The West’s manufacturing strategy is changing.

We want to make more things, but are having to learn it all over again.

The process knowledge built up during our manufacturing heyday has been lost, or at least severely depleted.

Those lost opportunities could have made tangible changes in people’s lives and the physical economy around us.

However, there are benefits to not having all that legacy, and looking at manufacturing from a fresh perspective. Leveraging tools such as Physical AI to progress in leaps and bounds.

To do what America did in a hundred years, or what China subsequently did in a few decades, can be compressed again by those willing to try.

But we need to analyze the incentives to bring certain manufacturing processes back from overseas.

In particular, complex assembly processes such as phone or laptop manufacturing which were traditionally very labor-intensive, due to their sheer variety and intricacy of tasks.

This complexity, and the need to ramp production at breakneck speed, meant that investment in dedicated capital equipment never made sense traditionally, for such esoteric processes.

This led to the hunt for lower-cost labor, where China cornered the market, and then moved up the value chain, to produce more high-value components. It then leveraged its own automation prowess to automate those manual complex assembly tasks.

Now countries such as India and Vietnam are progressing on this journey. But China does not want to let go, and there is now an appetite in the West to bring these assembly operations back.

However, to be cost-competitive the previous article ‘Assembly on Demand’ argued that assembly processes needed a standardized architecture, so that they could be assembled like lego blocks.

Equipment that can be quickly reconfigured to handle both the variety of operations and keeping pace with product iterations.

We are now going to research these assumptions in more detail to see whether there is a role for flexible assembly automation, or is the march to lower-cost regions inevitable? Until those possible regions run out.

The Manufacturing Stack

To understand assembly, we first need to know where it sits in the manufacturing ecosystem.

I previously introduced the concept of part manufacture (injection molding, vertical milling, 3D-Printing, etc.), vs assembling multiple components together.

However, there are many subtle layers in the manufacturing stack of which these broad concepts form part.

To get the final product in the customer’s hands, multiple types of raw material have to be processed into components or substances that can be used in an assembly.

In so doing, the manufacturing stack can be broken down into nine stages, from S1 to S9, shown below, and is where the concept of vertical integration comes in.

When I was at College, I was taught that Original Equipment Manufacturers (OEM) should focus on their core competences.

This core competence, often focused on the S7 to S8 stages, with earlier stages handed to trusted suppliers who had their own specialities and expertise.

This followed the Toyota Production System model where supplier relationships were close and often formalized through the keiretsu system.

In Japan and Korea, and to a lesser extent Germany, this meant companies throughout the S1-S9 manufacturing stack were co-dependent on each other, with cross-shareholdings, co-developed parts, and long-term sourcing commitments.

This permitted OEMs to concentrate on S7-S8, while Tier-1 and Tier-2 suppliers specialized in S4-S6 but with close knowledge transfer in order to optimize each stage of manufacture.

In much of the West, the importance of these close relationships was lost, with competitive bidding and more transactional relationships becoming the norm.

This allowed suppliers with close relationships to Toyota, and alike, to access companies such as Ford and GM, beginning the hollowing out at lower tiers of the supply chain in the West, with the loss of design and process knowledge.

Indeed, Ford once brought a large share of the S1-S9 manufacturing stack together at its River Rouge complex, complete with its own steel mill and glass factory.

Over time, it moved in the opposite direction, spinning off its automotive components and systems business, Visteon, and increasingly sourcing S5 and S6 modules from external suppliers.

Until Tesla came along and realized the importance of the close coupling from S1 to S9, pushing the boundaries of performance by enacting exact control at every stage of production.

This has become more important in the age of the software-defined vehicle. When vehicle behavior depends on tightly integrated software, electronics and sensing.

In this model, vertical integration is not about owning everything. It is about owning the layers of the stack where outsourcing would slow learning, limit final integration or affect product performance.

Apple, despite its legendary reliance on contract manufacturing, understood that outsourcing production does not mean outsourcing manufacturing knowledge.

Its own product designers, manufacturing engineers, subject matter experts and commercial teams are integrated into all stages of manufacture.

The result, a tightly controlled manufacturing stack in which suppliers provide scale, equipment and execution, while Apple retains control over process design, architecture of capital equipment, tooling and schedules.

In this model, final assembly and test become the pinnacle of where all this process knowledge comes together.

Without the knowledge of the final requirements. It becomes more difficult to integrate components from lower in the manufacturing stack.

Consequently, hand assembly becomes a way to navigate around these integration issues.

Caught in the Variety Trap

Presenting hand-assembly as a purely engineering problem is only one aspect and we need to push on this a little to provide a better perspective.

Analyzing the variety trap, shown below, you can see at each stage of production, the variety of product options broadly increases.

From just a few raw materials, such as aluminum, copper, or steel, to hundreds of thousands of components which can then be assembled into tens of millions of different products.

What is notable, is that the size of the overall market at each stage of manufacture is significant, accounting for $1-2 trillion annually. But that revenue is split over a few commodities for raw materials, as opposed to millions of SKUs for Assembly and Test (S5-S8).

Precision Components (S4), such as semiconductors or battery cells, are where product concentration and complexity is optimal to be able to specialize and hone processes incorporating advanced capital equipment.

The result is that the best companies have risen to the top. For example, TSMC for semiconductors, CATL for batteries, Rolls-Royce for Turbine Blades and Samsung which have cornered the market for phone displays.

Due to the scarcity of capable companies for Precision Components, they are able to command higher margins over standard 3D-printed or machined components.

This can be illustrated in detail by analyzing smartphone cost, where 54% of the cost is accounted for by a few key components, such as the display panels or camera optics, dominating the 100+ components that make up final smartphone assembly.

The same is true for a vehicle lithium-ion battery pack where 70% of its cost relates to the manufacture of the cells that go into that pack.

This margin and cost thread continues when you break down Assembly & Test into its constituent S5, S6, S7 and S8 buckets.

S5 functional sub-assemblies are still specialized enough to use automation. For example, Surface Mount Technology for PCBAs.

However, this automation is less specialized but has a higher degree of bounded flexibility, which means that it can be utilized for multiple product lines, year-over-year.

This allows a reduced capital cost to be amortized, while demand can be filled with as few or as many products as capacity allows, maximizing utilization, while also reducing labor.

Module (subassemblies, such as phone housings, or a robot actuator) and Final Assembly have typically not found a way to assemble vast combinations of components effectively.

Hand assembly, while becoming the default, also means that integration issues do not typically surface, because humans are adaptable, they typically find a way around assembly challenges.

Therefore, process knowledge does not get applied across products, preventing the development of flexible automated solutions whose cost can be amortized year-over-year.

The Integrator Dilemma

S5 processes such as SMT or Wire Processing, have revealed a path forward.

A reason why these processes were able to automate, is that certain input parameters were standardized (component packaging, geometry, etc.), while still providing a vast variety of product configuration, programmed directly from design data.

It is why, when you give your PCBA design to a Fab, they can tell instantly if you have violated a Design for Assembly (DFA) rule that will prevent your design from being assembled.

For S6 and S7, in addition to undertaking the cross-learning of DFA rules across multiple products to accurately bound what can and cannot be manufactured. There is also the integrator dilemma that assemblers have to face.

As OEMs began to automate certain hand-assembly operations. The challenge was given entirely to Integrators who developed bespoke solutions, because they had limited power to push back on DFA issues.

Therefore, the integrator’s business model became to develop custom solutions year-over-year for its customers.

These Integrators developed the mechanical design, software recipes, and process parameters, which was condensed down to a simple load or unload process for the operator.

Therefore, because of this lack of process knowledge by the OEM, and limited ability for Integrators to address DFA issues meant a reusable architecture never emerged.

That is why future Contract Manufacturers need to work with their customers to co-develop assemblies, as well as take on more of an integrator’s role to begin to architect an automated solution that can work across product lines and iterations.

This is one reason why Tesla has been able to make the production innovations it has, because it deliberately kept manufacturing engineering in-house, closely coupled with product design.

Keeping operations in-house also provides a critical speed advantage, as it eliminates the delays inherent in communicating across different companies.

What Does a Flexible Machine Look Like?

At the start of the article, I proposed that an automated assembly machine should be constructed like lego-blocks, almost suggesting an infinite number of bespoke architectures.

It was anchored in the Integrator model, that assembly operations have lived in for so long.

However, as we have seen in S5 of the assembly stack, it is possible to have a collection of machine types that is flexible enough to produce multiple products.

Consequently, there is a second concept, a single architecture that is capable of many operations, such as a six-axis robot with flexible tool changing.

At low volumes, one cell performs every assembly step. As volumes increase, these cells are scaled, each performing fewer steps and through data collection can optimize process parameters. Improving both throughput and yield.

In reality, a solution is likely to sit within these two bounds.

By incorporating vision and force feedback, robots are now able to learn and hone multiple operations, increasing the bounds of flexibility a single machine can achieve.

In addition, essential elements of automation: servos, motors, conveyors, part handling, and tooling are becoming more plug-and-play improving the ease of integration and reconfiguration.

In summary, we are at the stage where the bounds of flexibility are widening for the first time in decades, providing the possibility for a single architecture to be productive across multiple SKUs.

Beyond Assembly

As the bounds of flexible automation widen, manufacturers also begin to rebuild the process knowledge that was lost when these operations were outsourced.

Whether the relationship is between a customer and contract manufacturer, or between different groups within a vertically integrated company, this knowledge can then flow back into product design.

DFA rules no longer need to be discovered after the equipment has been built. They can be applied at the CAD terminal, improving process efficiency before the product reaches the line.

There are also higher-margin activities around final assembly, particularly Test.

By moving more validation into, or directly adjacent to, assembly, defective components can be identified earlier.

Process parameters can be optimized, while stronger evidence of product performance can be generated. This the customer more confidence, which many will pay more for.

Final assembly is also the stage where hundreds of parts from multiple suppliers converge.

This makes S7 the natural point to manage sequencing, traceability, cost and quality across the supply chain.

All these activities generate data.

That data improves yield, identifies supplier issues and provides the training data needed to teach Physical AI.

And notably, as manufacturers begin to teach their own machines, they start to rebuild the process knowledge previously handed to integrators.

This creates a virtuous loop.

More knowledge improves the product and process. Improved performance raises margins. Those margins then provide the investment required to scale manufacturing once more.

Making Reshoring Successful

So, back to the question we started with.

Is the march towards lower-cost regions inevitable for complex assembly operations?

Not necessarily.

Reshoring will not follow a single path through the S1–S9 manufacturing stack. In some industries it will begin with precision components, as we are seeing with semiconductor factories. In others, it may begin with subassemblies or final assembly.

Nor does the entire stack need to return at once.

What matters is that the stages which do return are closely connected to product design, allowing process knowledge to accumulate and support the next stage of investment.

Higher Western labor costs make automation more attractive. Shorter product cycles, supply-chain risk, quality and time-to-market strengthen the case further.

The opportunity is not to recreate the manufacturing system the West lost in exactly the same form.

It is to use flexible automation, Design for Assembly and Physical AI to rebuild that capability faster.

That is how the West can make more things again: not by competing with lower-cost regions on labor, but by compressing the learning curve and turning process knowledge back into an advantage.

Share

Subscribe now

Sources

The following references were consulted during the writing of this article:

• Sturgeon & Kawakami, Global Value Chains in the Electronics Industry, World Bank

• PwC, Mine 2026: Annual Review of the Top 40 Mining Companies

• Semiconductor Industry Association / WSTS, Global Semiconductor Sales, 2024

• Manufacturing Market Insider, Top 50 EMS Companies, 2024

• Berylls by AlixPartners, Top 100 Automotive Supplier Study, 2024

• Bain & Company, Automotive Profitability: How OEM and Supplier Margins Are Faring

• Precedence Research, Steel Market Report, 2024

• SKF, Amphenol and Murata, FY2024 Annual Reports

• Hon Hai Precision Industry (Foxconn), FY2024 Financial Results

• TechInsights, iPhone Teardown and Cost Analysis series

• Fomalhaut Techno Solutions, iPhone component cost analysis (via Nikkei)

• Kraemer, Linden & Dedrick, Capturing Value in Global Networks: Apple’s iPad and iPhone (2011)

• CNBC, Here’s How Much a Made-in-the-USA iPhone Would Cost (2025)

• BloombergNEF, Annual Lithium-Ion Battery Price Survey

• Argonne National Laboratory, BatPaC: Battery Pack Cost Modeling

• Wuxi Lead Intelligent Equipment, battery production line documentation

• Grand View Research, Outsourced Semiconductor Assembly & Test (OSAT) Market Report

• US Bureau of Labor Statistics, Employer Costs for Employee Compensation (2025)

• Reshoring Institute, Global Labor Rate Comparisons

• AppleInsider, How Apple Learned Automation Can’t Match Human Skill (2020)

Thanks for Reading

If you enjoyed this article, check out:

• Assembly on Demand

• What’s Next in Manufacturing?

• When Your Factory Starts to Think

Share

Subscribe now

After writing five articles on ‘Making Your Product Manufacturable’ you may be surprised to hear me say that those articles covered only one half of the equation. In fact, the articles were laser focused on the parts that must be assembled together to manufacture your desired phone, laptop or car.

However, there is also the ‘Machine that Makes the Machine’. The 5-axis Computer Numerical Control (CNC) milling device which must be programmed to subtract metal from a billet of aluminum to leave a supremely machined phone enclosure. Or the countless industrial robots welding together a car body from hundreds of stamped parts.

These two automated methods are common in the fact that they use precision fixtures to hold the part(s) robustly, while being machined or welded. However, they differ in how quickly they can be swapped over to produce an alternative product.

A 5-Axis CNC, can be reprogrammed offline, and with a new fixture dropped onto the machining bed could be machining an alternative part in a matter of hours.

Analyzing data from in-station and offline metrology tools, offsets in the machining program can be incorporated to ensure Critical to Quality (CTQ) features are qualified and meet specification within a day or two.

Compare this to a vehicle body shop. Reconfiguration for a new product, is a lengthy affair, more akin to conducting open-heart surgery on the production line over multiple weekends.

Installing new robots, retro-fitting fixtures, altering control logic, until months later the updated product can progress fully through the series of stations and pass into the Paintshop.

Often this open-heart surgery is not sufficient, and an entirely new production line needs to be commissioned.

Automotive Bodyshop (photo copyright @ ABB)

Why is Flexibility Sacrificed in Assembly Operations?

An important distinction to make with CNC processes, compared to our Body Shop example. Is that machining is a subtractive manufacturing process, while welding a vehicle body together is an assembly process.

Many subtractive and additive manufacturing processes have undergone a revolution over the past quarter century, driven by advances in CNC and digital fabrication.

This digitalization has enabled tool paths to be generated directly from CAD models, allowing parts to be milled, laser cut, or 3D printed with minimal human intervention, significantly narrowing the gap between design and production.

It is why companies such as Hadrian, Vulcan Forms and Xometry have gained strong market traction, offering flexible part production to both industry and hobbyists alike.

Why hasn’t this happened for Assembly operations?

Why can’t you upload your CAD model of a complete phone and have the manufacturing process planned and executed for you? Why hasn’t the same level of digitalization occurred?

The core difference is that while you can manufacture multiple geometries from a single general purpose platform, such as a 3D printer. Assembly requires configuring many subsystems, robots, conveyors, fixtures, in a precise sequence. It is not one machine, but many coordinated ones.

Mixed in with this coordination are many different process types, from welding, gluing, fastening, and even some forming processes such as bending or clinching.

This makes hand assembly the default starting point, just as it was for early smartphones and as it is now for humanoid robots.

As process understanding grows, these tasks are slowly broken down into highly detailed sub-tasks and automated. But each step that is automated creates a dedicated process, limiting reusability.

What’s missing is a layer of abstraction. Building blocks that can be configured digitally, to assemble anything. Embedding intelligence into those blocks could unlock the same level of code-driven agility exhibited by today’s best CNC tools.

Industry 4.0 promised this, but in practice it layered rules-based software on top of rigid, single-purpose machines. The result, very little improvement, in speed or flexibility.

Hardware needs to be designed from the outset to be configurable, discoverable, and reprogrammable, so changeovers feel like software updates.

Hand-Assembly of Optimus (Teslarati https://www.teslarati.com/tesla-optimus-pilot-production-line/)

Preparing for Automation

Using humans is the most flexible assembly process, because of our general intelligence and dexterity.

We can be trained to perform a range of tasks using a multitude of tools. This makes it simple to configure us into a long production line of processes as distinct as we are.

When competition is limited, unit price high, and product iterations fast, there is little incentive to move away from this model.

However, as competition increases, pressure is put on cost, quality and throughput, driving firms to make capital investments in automation to lower operating costs in order to improve or protect margins.

However, there are limits.

You must design your product, so that it can be assembled via automated means, otherwise you will end up supplementing these processes with hand rework and quality checks, slowing down production and increasing cost.

You need to consider how you are going to ‘Make your Product Manufacturable’.

This means reducing part count, either through elimination or combining parts into a single structure, such as Tesla’s Gigacasting. Doing so shortens assembly sequences and removes opportunities for failure.

Selecting the correct assembly datum features and applying them consistently throughout the process simplifies your tolerance chain making it more likely that the final product meets design specifications.

Further, assembling two parts together in a single orientation, rather than using a complex motion such as toeing in, simplifies fixture and end-of-arm tooling. This allows for repeatable location, and because motion is simpler it can be executed at faster speeds.

Automated handling is also aided by simplifying component geometry at clamping points, prioritizing shapes that are easy to grip and locate with end-effectors.

None of this means that design-for-automation rules should come at the expense of creating an exceptional product. As Sam Altman puts it, “Until you build a great product, almost nothing else matters.”

Many processes that were historically difficult or uneconomical to automate are now achievable using robots with more manipulation axes, combined with vision and force feedback.

The advance toward humanoid robots in production environments increases the scope of what can be automated. But this may not be the most efficient, there may be faster, cheaper and more robust ways of automation.

It is why, when you look at any production line that was previously manual, the machine that replaces a manual process looks very different.

The thought process behind automation is important, one which rewards simplicity.

Therefore, any automation effort should begin by decomposing the task into its simplest actions and reasoning from first principles about how each one is executed.

That process inevitably involves questioning the constraints imposed by the design, while preserving the intent and quality of the product.

Example of replacing human with humanoid robot in the loop (left), rather than rethinking the process to maximize speed and efficiency (right)

The Machine that Makes the Machine

Now that we have a design that can be assembled via automated means, we can consider those automation building blocks.

Taking a CNC milling process, it can be decomposed into a software recipe (G-code), work holding (fixture), tool (cutter), tool manipulation (gantry), and part loading (human or robot).

Many of these blocks are explicitly designed to be variable. Tools, fixtures, and programs can be swapped for standard or custom alternatives. A drill replaced with an end mill or a soft jaw vice exchanged for a dedicated casting fixture, for example.

However, there are elements, such as the machine bed or gantry, which remain fixed. Their drives and motors simply respond to new instructions via standardized control interfaces.

Assembly lines rarely enjoy this separation.

When changing from one joining process to another, or reconfiguring for a new product, not only do tools need to change. Fixed elements like robots, dedicated fixtures, and material storage often must be repositioned, replaced, or incorporated.

If flexibility is not designed in from the outset, changeovers become slow, invasive, and expensive.

To avoid this, assembly systems must be designed around a Single-Minute Exchange of Die (SMED) philosophy. Where the line behaves like a construction kit, reconfigurable in hours rather than weeks or months.

Making this practical requires quick interchangeability across the entire machine architecture. This interchangeability spans several tightly coupled dimensions, forming several tiers of modularity, as shown by the infographic below:

The Building Blocks of an Autonomous Factory maximizing Speed, Flexibility and Reusability.

Every element of the mechanical design and controls are digitized, allowing manufacturing engineers to take an assembly or test specification and convert it instantaneously into an assembly station design, ready to be inserted into the production line.

Autonomous Line Design from Product Specification to Fully Architected Assembly Station.

Getting to an Autonomous Factory

In the end, hardware alone is not the answer. Automation only delivers real leverage when machines, software, design tools, and supply chains are considered as a single system.

Building great products at scale will increasingly depend on software-defined machinery, generative design tools, and architectures that can flex across existing supply chains, from casting and machining to SMT and final assembly.

Autonomous manufacturing systems must be designed from first principles, not layered on top of rigid automation.

As sensing, vision, and robotic dexterity continue to improve, and critically, as their cost and cycle time fall, the boundary of what is practical will keep moving, favoring systems designed for adaptability rather than brute-force automation.

📚Thanks for Reading

If you enjoyed this article, check out the ‘Making Your Product Manufacturable Series’:

• When Your Factory Starts to Think

• Making Your Product Manufacturable: Part V: Getting to Ramp

• Making Your Product Manufacturable: Part IV: How To Believe In Your Measurements

• Making Your Product Manufacturable: Part III: Designing for Variation

Share

Subscribe now

We’re at the beginning of the AI era.

It has its sights on everything from content creation, customer support, household chores, and increasingly, manufacturing.

This pending manufacturing revolution is often compared to the historical transformation in agriculture as a way to make sense of what may come next as AI evolves.

Five hundred years ago, farming dominated employment with roughly 60-65% of the English workforce working the land. Today, in most industrialized countries, that number is closer to 1 to 3%.

And yet we produce far more food than ever before.

This has been a key for economic prosperity, enabling an increasing share of the population to move up the value chain into manufacturing and services raising the GDP per capita year-over-year.

It is proposed that AI will push manufacturing toward near-total autonomy, reducing the share of people employed in the sector from 10%, in the West, to agricultural levels.

If true, this would represent a structural shift on the scale of the First Industrial Revolution itself.

So together, let’s look back and analyze what happened to see what insights we can draw for what may come next in manufacturing.

Seed Drill Invented by Jethro Tull (1701)

The First Industrial Revolution

The First Industrial Revolution began in 1760 in the UK, spanning eighty years. In this period the population transitioned from an agrarian, craft-based economy to one which became increasingly mechanized and concentrated with factory-based manufacturing.

However, the seeds of this transformation were planted long ago and by the 1700s people had been spinning yarn, weaving and doing intricate metalwork from home for 200 years. The inclination to make had germinated and was in full bloom (de Vries, 2008).

At this point, the percentage of the workforce employed in manufacturing (35%) had almost reached parity with agriculture (40%).

The propensity to ‘make’ extended to agriculture, and by 1700 England was benefiting from early innovations in crop rotation, field enclosure and early mechanisation, triggering a period of sustained population growth as nutritional quality improved, quadrupling to 25 million by 1870.

These innovations matured throughout the eighteenth and nineteenth centuries, allowing productivity to double, requiring a reduced share of the labor force to work the land, halving to 20% by 1870.

In absolute terms, numbers did increase from roughly one to three million people, and when combined with productivity gains, allowed output to increase almost fourfold to sustain the growing population, meaning a massive surplus of people moved into manufacturing and services.

UK sector output (Broadberry et al, 2015)

Manufacturing Intensification

Well, looking at the graph above showing UK output by sector, agricultural growth now looks rather modest compared to the eighteen-fold growth in manufacturing output.

However, the share of labor employed in manufacturing barely shifted, increasing from 35 to 40%. But this hides the boom that occurred due to population growth, with manufacturing employment increasing roughly six-fold to five to six million people.

This still represented a tripling in productivity, enabled by the concentration of labor, capital, and power in urban factories, as people increasingly migrated from rural areas in search of industrial work.

But what was the motivation behind building those first factories? To answer this, you have to look at the most dominant driver for manufactured goods during this period, textiles.

As we found, in 1700 spinning yarn and weaving co-existed alongside farming. A structural shift that had not occurred in similar countries such as France.

The output from spinning yarn, and subsequently weaving that yarn into finished goods, was well matched up until the point the flying shuttle was invented by John Kay in 1733, increasing throughput of looms.

At this point, it took three spinners to keep up with the demand from the downstream weaver. This prompted innovations such as the Spinning Jenny in 1760, which allowed a person to spin multiple threads at once, amplifying their individual output.

However, at this point both spinning and weaving still required skilled handwork and could be done in dispersed cottage settings.

The Spinning Jenny invented by James Hargreaves in 1764

However, the increased output lowered prices, stimulating demand, driving pressure to intensify manufacturing processes further.

This pressure to innovate birthed Arkwright’s Water Frame in 1769, which replaced human power with water (and later steam) mechanizing spinning itself. The result was stronger, more consistent yarn and a sharp reduction in reliance on the skilled spinner.

But water frames were large, capital-intensive, and fixed in place. Consequently, they were concentrated in mills near power sources, marking the decisive shift from dispersed artisanal work to factory-based manufacturing.

The Spinning Mule followed in 1779 and became the dominant spinning technology of the First Industrial Revolution, shifting the production bottleneck downstream to weaving.

And so came the Power Loom in 1785, which again required continuous power and precise mechanized coordination completing the consolidation of textile production into factories.

If you look at the structural shift in manufacturing, it was enabled firstly by innovation in capital equipment, creating mechanized production, increasing throughout and improving quality.

Secondly, it harnessed power, with the steam engine replacing human effort, massively intensifying production.

In summary, mechanization automated skill, while power generation automated effort.

This in itself, drove manufacturing innovations, which included the coke-fired blast furnace, precision boring and improved iron refining enabling the production of reliable machinery.

However, grid scale power did not yet exist, power generation and its use in industrial equipment was co-located by the end of the First Industrial Revolution.

Double-acting Steam Engine invented by James Watt in 1769

Societal Shift

As factories scaled, they placed pressure on downstream distribution. Goods had to be transported, stored, financed, insured, and sold.

Canals, and later railways, dramatically lowered transport costs while creating entirely new industries. Finance, insurance, retail, and administration expanded rapidly to coordinate increasingly complex production networks.

This expansion was underpinned by legal and financial innovations that made industrial scale possible. Joint-stock companies allowed capital to be pooled from multiple investors, spreading risk and funding ventures too large for any individual.

Patent law protected inventors, giving entrepreneurs like Arkwright and Watt the confidence to invest years perfecting their machines. Contract enforcement and property rights provided the stability that long-term industrial investment required.

Services did not lead the First Industrial Revolution, but they were transformed by it. They absorbed labor at scale, particularly as manufacturing became more capital-intensive.

The predictable income from manufacturing and services, drew ever more people into the cities, looking for secure wages. This agglomeration pushed productivity higher through shared infrastructure, specialist skills and faster knowledge transfer within and between sectors.

By 1870, Britain had been transformed from a rural society into an urban industrial economy. GDP per capita had risen steadily, manufacturing output scaled enormously, and services became a dominant employer.

What’s Different Today?

Prior to any mechanical or power revolution, the foundations for the First Industrial Revolution were already laid. The market for textiles and metalwork were well established, the challenge was how to scale it.

The culture in the UK was primed to meet this challenge. It was a nation of makers and tinkerers and so people like Kay (Flying Shuttle - 1733), Hargreaves (Spinning Jenny - 1764), Arkwright (Water Frame - 1769), Watt (Steam Engine -1769), Crompton (Spinning Mule - 1779), and Cartwright (Power Loom - 1785) had both the reason and inclination to develop machines that mechanized production, or power it.

Through scale, the First Industrial Revolution helped to create the consumer economy you are familiar with today. Mechanization lowered costs and improved consistency, turning manufactured goods from luxuries into necessities, opening the market to more people.

Today, the starting conditions for the AI revolution are different. For advanced economies, population growth has slowed and many consumer markets are mature.

At the same time, the services sector, where AI automates information work and has historically absorbed workers from manufacturing, is likely to be disrupted faster. Manufacturing, by contrast, requires tight coordination between information and physical effort, making it harder to automate and slower to transform.

The productivity gains are obvious across both sectors. However the harder question is where is the new demand to power the growth that absorbs displaced labor?

In addition, the gulf between service and manufacturing productivity growth may be amplified by the innovation ecosystem that exists today.

During the eighteenth and nineteenth centuries, the people who learned how to harness new sources of power, invent machines, and mechanize production were close to the scaling challenge itself.

Design, prototyping, production, and refinement increasingly concentrated around the factory environment, shortening learning loops and accelerating improvement.

Today, those learning loops are far more dispersed. AI research, product design, capital equipment development, and manufacturing are geographically separated. Intellectual property, process development, machine building, and production no longer co-locate by default.

Hangzhou and Silicon Valley are AI powerhouses. Silicon Valley leads in GPU design through companies like NVIDIA and AMD, while Taiwan manufactures those chips via TSMC, using machines developed in the Netherlands by ASML.

If you look at the consumer products whose manufacturing methods are primed to be transformed by AI, Silicon Valley and China hold the lead in their Design and Engineering. While China dominates Capital Equipment Manufacture and Production, with Germany, Japan and Korea notable players.

While there is excellence within sectors, there are very few geographic locations that are vertically integrated to harness the fuel that is AI from concept to final product, creating a true AI manufacturing learning engine.

Integrated AI within Manufacturing

Growing with AI

If learning density is the scarce resource in modern manufacturing, how can this be recreated?

Just like in the First Industrial Revolution, it requires building out the next generation of intelligent manufacturing capability, co-locating in areas where there is existing AI and robotics density.

Even if manufacturing systems are imperfect at first, as the learning flywheel accelerates and assembly lines become more automated, labor cost will steadily lose its importance as a differentiator, lowering the barrier further to co-locate AI and manufacturing capabilities.

For the flywheel to accelerate, intelligence and instrumentation must be embedded directly into the design process and production line, allowing the data generated to compound into improvements.

These improvements will allow the manufacturing ecosystem to act through greater context. That context comes from learning from production data, what worked and what didn’t in machine design, process experience, and failure modes.

From there, learning can flow upstream into product engineering, improving automation capability and the refinement of AI systems themselves. But this requires close coordination and sustained practice.

As intelligence improves, it is likely that product design and manufacturing will become more tightly fused. Products can be co-developed, with their manufacturing processes in mind, optimizing geometry, tolerances, materials, and assembly paths simultaneously.

Further, AI driven design will be able to gain insights from the market and update or generate new products, shortening the lifecycles between iterations.

This will increase the drive for flexible automation. Systems that can plan, re-sequence, respond dynamically to demand fluctuation and be able to adapt to changing product designs without re-engineering the whole assembly line.

As AI integrates design and manufacturing, the gap between theory and application will become barely perceptible. The companies with the biggest advantage will have the smallest execution gap.

To capture that advantage, companies will increasingly need to vertically integrate the manufacturing AI stack, bringing intelligence, machines, data, and engineers into the same learning loop.

This is how learning density is rebuilt, and why the next phase of manufacturing will be defined less by who has the best models, and more by who can turn experience into compounding capability.

What Does This All Mean?

The First Industrial Revolution did not succeed because Britain invented better machines. It succeeded because learning occurred where production happened. Mechanisation lowered costs, demand expanded, labor was absorbed, and a flywheel took hold.

Today, AI offers a similar opportunity, but under very different conditions. Markets are mature, labor dynamics are shifting, and manufacturing knowledge is globally dispersed.

Productivity gains alone will not guarantee growth. Without new products, new demand, and places where learning compounds, AI risks becoming a tool that raises efficiency without rebuilding industrial strength.

The countries best positioned over the next twenty years will not be those with the most AI startups or the cheapest labor. They will be the ones that deliberately rebuild manufacturing as a learning asset, co-locating intelligence, machines, engineers, and production.

For the US, that means rebuilding manufacturing depth to match its leadership in AI and product design. China continues to benefit from dense production-led learning, while pushing upstream into design and systems innovation.

For the UK, the lesson is sharper. From being the first nation to industrialize, it now sits outside the world’s top manufacturing economies. AI alone will not reverse that. Without production density, the UK risks becoming a country that designs intelligence while the activities that make it economically transformative happen elsewhere.

The challenge ahead is not inventing the future, it is building the places where that future can be learned, refined, and scaled.

📚Thanks for Reading

If you enjoyed this article, check out the ‘Making Your Product Manufacturable Series’:

• Making Your Product Manufacturable: Part V: Getting to Ramp

• Making Your Product Manufacturable: Part IV: How To Believe In Your Measurements

• Making Your Product Manufacturable: Part III: Designing for Variation

• Making Your Product Manufacturable: Part II: Getting into Tolerance Stacks

• Making Your Product Manufacturable: Part I: Why It’s More Than Just a Tolerance Stack

Share

Subscribe now

If you work in manufacturing, you have more than likely had your brushes with quality systems and processes. Whether it is someone e-waving a template for a control plan to be filled in or having to route through multiple control charts to find what is out of specification, and then more importantly why?

In many companies, there are rafts of excel, word documents, databases, and propriety software, all disconnected, hard to retrieve and oftentimes has no active feedback to the design or manufacturing process.

Quality software falls into two camps, a document control system to store and version control procedures, control plans, buy-off checklists, etc. The other stores, displays and aggregates your production data.

If you need a deeper analysis, then begins the process of exporting, re-formatting and importing data into a statistical or business intelligence package. This is followed by a series of button clicks, mixed in with reviewing help manuals and YouTube videos, until you stumble on the right combination to spit out the desired result.

How can this lengthy endeavor be shortened so you as an engineer can get the information you want, using the language you would use when explaining an issue to a colleague?

Equally, how are all these repositories connected, so your manufacturing system knows how to respond, and course correct, to maintain healthy production with minimal intervention?

The controlled chaos of a quality system

🧱 A Disconnected Present

As we saw in the ‘Making Your Product Manufacturable’ series, there are many elements to a quality system, with one key aspect being the identification and management of your product’s Critical to Quality (CTQ) features. This includes, identifying CTQs, establishing acceptable tolerance limits and evaluating if they are feasible.

CTQs and their associated tolerances, should then be annotated within your engineering drawings, and a Process Failure Modes Effects Analysis (PFMEA) conducted to rank the risks in relation to each CTQ. Subsequently, preventative actions and detection strategies should be identified to mitigate these risks.

For each CTQ failure mode, a control plan details the exact method employed to prevent the failure from occurring. As well as the measurement and sampling method used to monitor the CTQ. If the measurement detects a failure, then a reaction plan details the actions to fix the process and correct the failure.

Your measurement method needs to be qualified to make sure that it is measuring what it is intending with the required accuracy, repeatability and reproducibility so that the risk of escape of a failed CTQ being passed downstream is limited.

Finally, any breaks in this process, need to be backed up by a robust investigation, putting in containment actions to, for example, rework the failed CTQ while a root cause analysis takes place to find the reason for the process failure.

Once root cause is found, a permanent corrective action must be implemented and validated. This corrective action must be robust enough to prevent the process control from failing in the future.

If we consider, this one aspect of a quality system, it falls into the description presented in the introduction, numerous excel sheets stored on shared areas, CAD drawings converted to dumb pdfs, only able to be interrogated by an actual engineer, and root cause investigations managed by JIRA with little learning across separate investigations.

How can we connect these systems?

🔗 An Integrated Future

How can the dots be connected in the situation described? AI is an answer to this problem, but this answer is too generic. ChatGPT or Claude doesn’t have access to your Excel templates or live data, so any response to your questions won’t be suitable for the specific situation you’re trying to address.

To do this, we need to consider how to store, embed and retrieve our disconnected data sources ready for a tuned AI model to access. That’s the premise of a Retrieval-Augmented Generation (RAG) architecture for quality. Let’s consider how all this information sources could talk to each other in real time in RAG framework.

A RAG framework for closed loop SPC control

As you can see from the above diagram, the actual generative AI model is one element in a wider system. For it to be effective, it needs a user input, an agent to provide the right context and planning function, as well as access to both a knowledge base and live data to execute the user query correctly.

Going into more detail, let’s consider an example, of an in-process measurement system feeding measurement results into an SPC AI agent, that is required to instruct the manufacturing system on the right course of action, so the production cell is entirely autonomous:

🗂️ Knowledge Base: Contains all the company specific knowledge, in particular the knowledge of how each CTQ characteristic should be monitored, and what corrective actions are valid for when a deviation occurs.

🧩 Embedding Function: To allow the information contained within the knowledge base to be utilized by the generative AI model, it needs to be encoded so that it can be searched by meaning. This is done by breaking company documents down into small chunks, which are then vectorized, providing a numerical representation of the chunks ‘meaning’.

🧠 Knowledge Base Vector Table: Once information has been vectorized, it is then stored in a vector table, making it searchable so relevant guidance can easily be retrieved. For example, if you’re in-process measurement system detects a drift in the deviation value of a CTQ. This could yield an automatic alert to the process engineering department to check the integrity of a particular fixture.

🔍 Retriever: For your SPC AI agent to pull the correct guidance from your Knowledge Base Vector Table. You need something for this vector table to be compared against, so that the top matching chunks of information can be fed into the AI model.

This is where your retriever comes in, as it will vectorize the system query, producing the top vector similarities to match with your vector table. Your agent then has all the information it needs to identify the correct course of action, relating to the alert to give to the process engineering department.

🤖 AI Agent: Your AI Agent is the key orchestration layer that defines how the SPC control behaves. Its action is governed by system prompts and context engineering, which establish its role, scope, and planning logic. This framing shapes how it interprets incoming queries and how those are passed to the Generative Model, ensuring the model knows what question it is answering.

Your agent coordinates targeted knowledge retrieval from the vector database and triggers live data queries when needed. The result, clear feedback to your manufacturing process, for example, automatically guard-banding a test specification after detecting an adverse result further downstream.

⚙️ Generative Model: This is the core of your SPC AI agent, serving as the reasoning engine, combining contextual knowledge from the retriever with live data from the production database.

This combination allows it to make a statistical interpretation of your production data and prepare a structured response to pass back to the manufacturing cell. This could be a natural language reaction plan for the process engineer to follow, or updated robot code to modify a welding sequence.

🧮 Text-to-SQL / Execution: This is your key component to be able to retrieve data from your manufacturing execution system, or quality database. It will convert an instruction, such as ‘provide the camera calibration data from the last 50 modules produced’ into a SQL query to return the required values from your database for evaluation.

📈 Live Data: As the title suggests this is all your in-process measurement results, sample metadata, and process configurations fed directly from sensors or PLCs in your production line. It is structured so that it can be queried, providing real time visibility into the stability of every data point within your production line.

💬 System Query → Response Loop: Within a closed loop control system, your AI agent receives the live inputs, interprets them and will execute a reasoned action. This action could be the automatic adjustment of a CNC offset to maintain control of a CTQ parameter, or issuing an instruction to the maintenance department to change the drill bit so the CTQ diameter does not drift below its lower specification limit.

Your AI Human Machine Interface (HMI)

🚀 This is just the Start

How does your quality system compare to the two described? For many, it will be somewhere in between, maybe you have everything fed through JIRA, so there is one knowledge store, or you have sensors within your production line giving it a level of autonomy. But maybe now you see the potential for more, how quality can truly integrate with your design and manufacturing organizations, rather than sitting off to the side keeping a watching brief.

The journey to a truly intelligent quality system doesn’t start with buying new software, it starts with connecting what you already know. In the coming weeks, we’ll dig deeper into each of these layers to realize a quality system of the future, I look forward to learning with you on this process!

📚Thanks for Reading

If you enjoyed this article, check out the ‘Making Your Product Manufacturable Series’:

• Making Your Product Manufacturable: Part V: Getting to Ramp

• Making Your Product Manufacturable: Part IV: How To Believe In Your Measurements

• Making Your Product Manufacturable: Part III: Designing for Variation

• Making Your Product Manufacturable: Part II: Getting into Tolerance Stacks

• Making Your Product Manufacturable: Part I: Why It’s More Than Just a Tolerance Stack

Share

Subscribe now

Also, please feel free to connect with me on LinkedIn.

This is it, the last step in ‘Making Your Product Manufacturable!’

To distil the overall aim of this series, it has been to lay out a methodical approach to develop a manufacturing strategy to deliver all Critical to Quality (CTQ) features that are essential for your product to function and delight your customer.

This included how you would assess the manufacturing feasibility of these CTQs in the virtual world. Now you must validate this refined strategy in the physical world using qualified measurement systems to be able to launch and ramp your product.

However, not all issues will appear all at once. For every incremental increase in volume, a new set of critical to function, process or product issues will appear, previously masked by rework, long cycle times or utilizing labor that now needs to be stretched thinner.

Therefore, beyond initial validation, measurement data must continue to be collected to enable informed decision making, as you ramp from single digits to hundreds or even thousands of units per day.

So, let’s explore how you can execute as best as possible to minimize the pain of production hell that you are about to experience.

🏗️The Build Process

Before going into the nitty gritty of getting your production running, it is worth going over the different build phases you will progress through as you launch and ramp production.

Terminology differs across industries, but you can broadly break the phases down into Prototype, Production Trials, Continuous Production, and Mass Production.

Prototype phases use non-production parts and tooling to build a low volume of vehicles with the purpose to validate the design. In parallel, the more complete production dies and automation are designed and fabricated for use in the later build phases.

Units produced in these early prototype phases are used by design departments to evaluate if their audio, lighting, display, etc., meets developed specs. In addition, units are allocated for system level reliability testing, homologation, crash test, etc.

It is important that there are gaps in these phases, so that you have time to perform these tests so you can feed any design updates into later prototype builds or roll them into production tooling.

At the end of the prototype phase, production tooling should be bought off via a Factory Acceptance Test (FAT), where various performance, quality and safety checks are performed before being packed and shipped to the production site.

Once onsite, equipment will be installed in its final location and bought off in a Site Acceptance Test (SAT). This is where production trials are first performed, getting parts and sub-assemblies through the production process so each stage can be commissioned.

Example Machine Performance Metrics for a Factory or Site Acceptance Test.

Then begins the process of integrating all these machines together, in the grueling march to ramp as initial units are built, evaluated and improved. Learning is incorporated, making the transition from station to station more seamless as volume gradually increases.

There are several critical activities that will allow you to move through these production build phases methodically, starting with tooling certification.

📏Tooling Certification

Tooling, be it molds for plastics, press dies for sheet metal panels, or fixtures for CNC parts or welded assemblies, will not align with your CAD geometry, because as you know, nothing is perfect in the physical world and there will be deviations away from CAD intent.

Metrology tools, such as the CMMs you encountered in the previous article, or portable 3D scanners are used to measure tooling, and record deviations of key locating pins, die surfaces, and clamping locations.

Adjustments are then made, which can range from a simple shim move on a weld fixture or benchwork to make fine adjustments to a mold surface, all the way to re-cutting a die, which is a significant undertaking.

Weld Fixture Locating Pin with Shim Adjustment (Left), Portable 3d Scanner used for Part and Fixture Scanning (Right).

Once adjustments are made, and the fixture or tool is within 10% of the part or assembly tolerance. A final baseline measurement is taken and filed as part of your Production Part Approval Process (PPAP).

PPAP includes all the documentation needed to buy-off your part or assembly, which not only includes quality documentation, but health and safety and machine performance evaluations, such as the SAT metrics shown in the earlier table.

In addition, for assembly fixtures, you should update the shim log to record the number of shims in the adjustment blocks for each pin, and clamp location.

These baseline scans and logs become important when root causing issues later in your builds, as the first thing you want to investigate, is whether anything has been altered from baseline. Most people who have worked in a body shop will know of the ‘phantom shimmer’ who makes undocumented adjustments for a myriad of unknown reasons 🙃.

Further, if you have parts or sub-assemblies produced using different fixtures to increase throughput. Baseline scans are vital to ensure the differences between the produced parts are within an acceptable limit, so that their differences don’t become a bottleneck downstream.

Tooling buy-off is not complete once fixtures are certified, it must produce parts accurately and repeatably, which may not happen with the current certified condition 😡.

🎯 Datum Pre-Qualification

When you send the first part from a die or assembly fixture, you will receive a report with a bunch of numbers, and chances are there will be a lot of red deviations, meaning features are out of specification.

So, the question becomes how to begin the process of methodically adjusting CNC offsets, or shimming fixture locating features to bring the product into specification? And the answer is datum features first.

One half of a Motor Casing with Datum Scheme detailed, (a) Input Shaft Bore, (b) Intermediate Shaft Bore, and (c) Output Shaft Bore.

In my very poorly drawn motor casing example. If you encountered the situation on the left where your datum features are within specification. However, your input, intermediate, and output bore positions were not.

If you adjusted your CNC offsets to bring the bores back into specification and then assembled the casing with its opposite mating casing (the mirror image of the one pictured) using precision dowels to align each part’s ‘B’ and ‘C’ datum features.

The opposing bores would also come into alignment, ensuring the assembled gear shafts are straight, parallel, and correctly positioned to mesh.

However, if you encountered the situation on the right and your ‘C’ datum was out of position relative to your ‘B’ datum, your resulting coordinate system would be skewed.

Even if you corrected the bore positions using that skewed frame of reference, the gears still wouldn’t mesh straight and parallel when the casings are bolted together. That’s because the two casings no longer share the same coordinate frame relationship, so their bores no longer line up in space.

Therefore, when correcting for out-of-specification features, you should take the logic of correcting datum features first, so all CTQ features are then brought into specification using a stable frame of reference which extend across multiple parts.

Taking this logic through to larger assemblies which are manufactured using a fixture to align the parts. Once welded or bolted together, and you have corrected the assembly datum coordinate frame. Your next focus should be the spatial relationship of the redundant datum features.

Redundant datum features are those part or sub-assembly datum features that are used to locate the part in the fixture using locating pins/clamps, but do not form the overall assembly datum scheme, as shown in the figure below.

Assembly process using a weld fixture, assembly datum features (red), redundant assembly datum features (black).

If you think back to the datum transfer process in Part I of the series, pictured below. Correcting redundant datums, at each stage of the assembly process, brings the part or sub-assembly coordinate frame into alignment with that of the overall assembly.

In effect, as you systematically do this correction throughout the process, you are ensuring a stable coordinate frame is maintained as you build the product, so that measurements at each level of the assembly tree tie together and don’t become disconnected.

Now you may think, well if I certified all the locating features in my fixture to be 10% of the tolerance, then the redundant datum features which locate onto these locating pins should automatically be in specification and have the correct spatial relationship.

Well, it is where I once again say ‘Well No’ 😭.

Due to part stresses, welding heat distortion, clamping sequences, etc., the associated sub-assembly won’t come out of the fixture as ideally planned for by your fixture certification.

That is why, you should methodically assemble parts, using what is termed a ‘slow build’ where you measure as you go. Understanding these micro-assembly effects allow you to compensate them out of your build, so that assemblies come out of the fixture with the associated parts having the correct spatial relationship.

Using the Datum Transfer Methodology to Correct Assembly then Redundant Datum Features.

🔁 You’ve Assembled it Once, Assemble it Again and Again

Now that you have built one assembly through your process, there are a few more questions you need to answer before you can buy-off and begin to ramp. Firstly, where does that assembly sit within the distribution of assemblies you are going to produce?

Further, how wide is the process distribution compared to the tolerance band you determined was feasible as you were constructing your tolerance stacks during the design process?

The only way you are going to answer these questions, is by building and measuring more assemblies, ideally around 30 parts, to calculate statistically meaningful Cp/Cpk figures for each CTQ. It is important that assemblies are manufactured at production rates to assess if accuracy is maintained at the planned cycle time for the assembly cell.

Like your tolerance stack studies, the results of the 30 measurements will yield a distribution of results, whose width and location can be compared to your required specification limits, using Cp/Cpk calculations, like the examples below:

Example Process Distributions

So, what do you correct first?

In situations where your Cp value is low, meaning your process variation is significantly wider than your tolerance band. Ideally, you should focus on correcting the variation, as in these situations any estimate of mean-offset is likely to be noisy.

Trouble is, tightening variation is often a trickier endeavor than adjusting the mean. That is why, in production situations, it may be necessary to perform a mean-shift to move your distribution so that a greater percentage of assemblies are positioned within acceptable tolerance limits, knowing you will likely have to make a further mean-shift later.

As you saw with your tolerance stack, there is no single source of variation, there are often several contributors. Therefore, you can review your tolerance stack results, comparing it to measurements of upstream contributing processes to identify excessive sources of variation.

If there is excessive variation from these upstream parts or sub-assemblies then these can be targeted for improvement. If these check out, then you can focus on your current assembly process.

Firstly, before going in-depth with Design of Experiments (DoE) studies, you should observe the process and look for problematic issues, that could be causing this additional variation.

Many times, I have observed a mis-taught robot position or sagging parts in an assembly fixture because the tooling design does not support it correctly. These process observations should immediately be corrected for, before going into more in-depth studies.

If upstream processes are fixed and you have corrected observed issues in your current assembly process. Then you need to understand all the different process parameters, e.g. clamp sequences, clamp pressures, weld order, robot loading, CNC machine speeds and feeds, die pressures, etc.

As mentioned, conducting DoEs, is one method that can be used to identify where the problematic sources of variation are arising. Taking weld sequence for example, welds are programmed to be performed in a set order. Firstly, to lock in the geometry of the assembly, and secondly, to create the joint strength.

However, this order may not be locking in the geometry reliably. Therefore, you can trial different weld sequences, and by holding other variables constant, you can assess the impact of weld sequence in a series of short studies, measuring the result.

If a promising weld sequence is found, you can fix the updated process and redo your capability study, which will show a fixed or at least improved process variability.

In summary, you should use your tolerance stack to pinpoint process inputs that affect your CTQ, then use process observations and DoEs to identify and fix those unknown issues that your tolerance stack won’t pick up.

Methodically going through this process, will help stabilize production, so that you can then reliably correct your mean-shift to hit your Cp/Cpk targets for your CTQs as you move your assembly through the process.

Tackling each of these process issues, will allow you to produce your product more seamlessly so you can ramp production to target volumes.

👀 Monitor Continuously

Once you have established your baseline quality it is important that you have in place a monitoring and control plan, so that any process drifts are quickly identified. Once identified the control plan should be able to point the engineer to the area of the fault, to address it quickly.

Often complex products will have many parts and several sub-assemblies that are measured on a continual basis. Larger sub-assemblies can have anywhere from a few hundred measurement points, to over a thousand.

The question becomes how do you monitor this amount of data effectively? The answer is again to use your datum and redundant datum feature structure, which will inform the process engineer whether each part in a sub-assembly has the correct relative spatial relationship.

Further, a simple traffic light system can be used to assess whether your sub-datum features are in specification or not, giving engineers confidence that key parts will fit as the assembly moves downstream.

In summary, you need a way to group the measurement points so you can monitor key sub-systems at high level, and if any drift, you can then dive into the detail to see what is going wrong.

Example interactive measurement report (left), simple status overview of sub-datum schemes for key vehicle sub-systems (right).

If you have in-process inspection, in-station alarms provide key personnel with instant feedback on whether a process is drifting, rather than waiting for the assembly to be measured by the metrology lab.

For smaller, less expensive parts, in-process measurements can automatically yield parts from the process, ensuring only good parts are passed downstream. For larger assemblies such as vehicle bodies, you need a method to identify potential issues before they occur, as scrapping even one assembly will cost multiple thousands of dollars.

One of the most effective, is to alert if a mean-shift has occurred in the process. This is done by comparing the mean-value of the current ten assemblies with the previous one hundred. If a mean shift is detected, process engineers can investigate the issue prior to it drifting out of specification.

As well as effective measurement and data reporting, having a control plan, and failure analysis / corrective action (FACA) plan is essential to ensure the process is quickly brought back under control if process drifts occur.

A control plan doesn’t just list CTQs and how they’re measured. It should also call out the critical process variables and input parts to check when a deviation occurs. This allows engineers to zero in on the likely problem area quickly, speeding up troubleshooting and getting the line back under control faster.

If the actions specified in the control plan cannot adequately correct the process or quality issue, a FACA provides a structured methodology to contain the issue via rework or extra process checks.

Root cause investigation can then take place, to systematically investigate why the issue occurred. Again, your tolerance stack and DoE studies can be used to find and fix the issue and verify resolution.

As your data collection, in-process measurement, and failure analysis become more structured, you start building a powerful knowledge base about how your process behaves. When this information is digitalized, it can be stored, vectorized, and made available to AI models. These models can spot subtle anomalies, predict drifts before they hit spec limits, and even recommend corrective actions.

When closed-loop control can’t self-correct production, the context in this data combined with AI-assisted prompting can guide root-cause investigations and help engineers converge on solutions faster.

🗂️ Production Part Approval Process (PPAP)

Now you are finally ready to complete your PPAP documentation, bringing together the baseline scans from your tooling certification, updated process logs (shim logs, CNC offsets, etc.) that brought your CTQs into specification along with final capability measurements.

Finally, control plans, FACA process, measurement system studies, health and safety buy-off and all machine performance metrics (cycle time, uptime, etc.) should all be bundled in with your PPAP.

Having this knowledge accessible will mean you enter production, knowing that any drift will be caught by your qualified metrology plan, and your control plans will tell you what adjustment to make to maintain process control.

Most importantly, it means you won’t have to re-live a mini production hell each time an issue arises.

🎬 Closing the Series

And with that, we wrap up the ‘Making Your Product Manufacturable’ series. We started by defining CTQs, then moved through datum schemes, tolerance stacks, variation strategies, measurement systems, and finally how to bring it all together at ramp.

Ramping does not have to be about living on the edge but leaning into the uncertainty with a disciplined approach to qualifying and maturing your manufacturing process, underpinned by data and robust design assumptions.

I hope you enjoyed the series. There is likely a myriad of subjects I have missed, so if you would like to learn more on a certain subject on manufacturing, do not hesitate to reach out!

📚Thanks for Reading

If you enjoyed this article and wanted to catch up on the other articles in the series, please click on below links:

• Making Your Product Manufacturable: Part IV: How To Believe In Your Measurements

• Making Your Product Manufacturable: Part III: Designing for Variation

• Making Your Product Manufacturable: Part II: Getting into Tolerance Stacks

• Making Your Product Manufacturable: Part I: Why It’s More Than Just a Tolerance Stack

Also, please feel free to connect with me on LinkedIn.

Share

Subscribe now

Welcome to the penultimate part of this series on making your product manufacturable. Before we get to the final step, we need to tackle one of the most overlooked questions:

“How do you know you can trust your measurements?”

As you have seen there are a lot of upfront design and manufacturing considerations when developing your product and its associated parts and assemblies.

Now we are getting closer to producing real parts, welding or bolting them together, and doing final testing to make sure the product conforms to design intent and is ready to be released into the happy hands of the customer.

To do this, you need to give engineers reliable measurement data to allow them to dial-in processes as they commission and ramp production.

Together, we will explore what you need to do to qualify your measurement systems, so that engineers can count on the results to make those necessary adjustments to bring your Critical to Quality (CTQ) parameters into specification.

Typical Automotive Inspection Systems: Left, Lab-Based Coordinate Measurement Machine (Hexagon Manufacturing Intelligence) and Right, In-Process Laser Radar Inspection (Nikon Industrial Metrology).

📏 Measurement Systems

To know if you have a good part or assembly, there must be confidence in the measurement result throughout the value chain.

You may think, once the engineering drawing is completed, it can be handed over to your supplier or quality department, and the Geometric Dimensioning & Tolerancing (GD&T) will ensure everyone has the same interpretation of the inspection requirement.

Well, not necessarily 😭.

Outgoing Quality Control (OQC), Incoming Quality Control (OQC) and In-Process Quality Control (IPQC) all must be set-up, qualified and correlated. If they aren’t aligned, results will be questioned, and uncertainty quickly snowballs into delays.

To align the results of all these measurement systems there are several pre-requisites that need to be in place to provide a foundation for good measurement.

🤔 Measurement Strategy

If you ask a machinist, designer, metrology technician or manufacturing engineer how a particular feature should be measured, you will likely get several responses.

This ambiguity occurs, firstly, because each may have their own concept of what information is important, meaning engineering drawings are viewed from multiple perspectives.

For a designer, it will be will the part function reliably once it is assembled into the final product. In contrast, a machinist wants actionable information to improve their machining process, and this may require a different or increased level of detail.

Secondly, physical parts are not the perfect surfaces or shapes you modelled in CAD. You need to decide how to construct features from a series of individual measurement points on the part and link them back to individual geometrical elements of the CAD.

Finally, there may need to be filtering or fitting of the measurement points to provide the correct functional evaluation, such as assessing whether a bolt will pass through a hole. Requiring points to be filtered to construct the smallest diameter, because believe it or not holes are not perfectly round.

The figure below provides an illustration of some of these considerations, where extra levels of detail are required in your measurement strategy beyond just GD&T.

Different Measurement Strategies

That’s a lot of questions from an unambiguous drawing, and some might say ‘if you refer to ISO standard 1101, it will tell what filtering you need to apply.’ However, it is rare I have seen these standards consulted and provide a definitive answer, only a guide.

You still need to use the knowledge available to make the right decision for the evaluation of the part or assembly. This is where developing internal standards are important or even having an internal AI agent that can guide you to select the right measurement parameters.

In summary, you and your company know the functional intent of the measurement, and it is your job to develop internal strategies to measure and output meaningful results reliably.

➕ Measurement Points

There are two types of measurement points, those individual points we just discussed used to construct the measurement feature. Then there is the measurement feature itself, for example a bore, trim edge, or slotted feature.

Construction points are dictated by the measurement strategy. While the features to be measured, should be documented so it is clear to all stakeholders what must be measured at each manufacturing stage.

Documentation of measurement features is necessary because the GD&T may not provide enough fidelity to monitor and control key production parameters.

For example, using the illustration in the bottom right of the diagram above, the profile may need to be evaluated as a series of points along the trim edge, to be able to evaluate exactly where the profile is going out of specification, and make the necessary changes.

Like datum features, measurement points should evolve as you move through the assembly process. At the part level, all aspects should be inspected to rigorously assess the form, fit and function of that part before it can be moved downstream.

Once the part is in a sub-assembly, many of the part-level points can drop off the inspection plan because they no longer add value. At this stage, you can’t re-punch a hole or change the profile of a parts surface.

What is critical is knowing the part’s location and orientation within the sub-assembly through measuring the part datum features relative to the assembly coordinate frame. Fixture adjustments can then be made to bring components into alignment.

In reality, you will also need to measure critical sub-datum features and their associated CTQs because they may span across several individual parts. Consequently, it is only possible to assess conformance to some requirements at assembly level.

Measurement Feature Rationalization Through Production

🗜️ Measurement Fixtures

How you hold the fixture during measurement can have a dramatic effect on the measurement result, particularly for flexible parts.

At a minimum for rigid CNC parts like phone housings, or motor cases, they should be held so they don’t move during measurement, and then the datum features measured directly to establish the part coordinate system.

Ideally for flexible parts, they should be held in the orientation they are manufactured and be constrained by the primary datum locations defined on the GD&T. This ensures measurement results are representative of the manufacturing process, because measurement fixtures should be a mirror image of their manufacturing counterparts.

However, if the part is being constrained by the datum features defined on the drawing, they will be inaccessible during measurement. Consequently, bushes should be added to the fixture which are accessible, to establish your part coordinate frame.

Datum features are then calibrated relative to these bushes and should be certified to be within 10% of the drawing tolerance, to ensure minimal inaccuracy occurs because of the fixture. Again, this should occur in both the measurement and manufacturing fixture.

Undertaking these methodical steps will ensure that measurement protocols are consistent throughout the manufacturing process.

Measurement Fixture Design

✅ Measurement Qualification

Now you have established the pre-requisites for good measurement, in theory your lab-based Coordinate Measurement Machine (CMM) should align with the results of your in-process vision inspection. However, this needs to be validated via a qualification process.

Just like a part or assembly, a measurement system has both mean-shift (systematic error) and variation around the mean (random error). The first is assessed via a Calibration, while the second is assessed via a Gage Repeatability and Reproducibility (GR&R) study.

🎛️ Calibration

Just like when you make a shim move on a jig or change the offset in a CNC program to make better parts. You need to follow the same method with measurements to provide better results.

To correct measurement systematic error, your measurement instrument, say a CMM, must be calibrated against a higher-accuracy instrument or artifact. This allows you to offset its results to bring them closer to the true value.

Typically, lab-based instruments such as CMMs are calibrated on an annual cycle. In-process measurement equipment is then correlated back to these lab-based references, which is discussed later.

As a rule of thumb, measurement accuracy should be no worse than 10% of your tightest part tolerance.

To do this, a reliable chain of calibrated artifacts needs to be maintained. This means that the granite block used to calibrate your CMM must be measured by a higher accuracy device, such as a laser interferometer, in a nationally accredited lab.

That lab, in turn, is calibrated to a national reference standard (e.g. NIST in the U.S. or NPL in the UK), which maintains even tighter accuracy and environmental controls.

Overview of Calibration and Gage R&R

📉 Gage R&R

Once systematic error is corrected, you must assess how much variation the measurement system itself adds relative to the process. This ensures that when you see differences in data, you can be confident it is due to part-to-part differences and not from a noisy measurement system.

The trouble is, as you can see from the diagram above, measurement system variation is wrapped up with overall part variation. In production, every time you measure there is a small amount of measurement error as well as part error, and unlike calibration, you don’t know the part’s true value to separate the two.

To resolve this, you must measure multiple parts, several times. By studying repeated measurements, lets you separate true part variation from noise introduced by fixtures, instruments, operators, or the environment.

A method known as Gage Reproducibility & Repeatability (GR&R) is used to breakdown these components of variation. It is done by having three operators each measuring the same part for three repetitions, for a total of ten parts.

It is important that for each measurement repetition the same process of measurement in its entirety is performed. This means you use the same CMM, load the part on the same fixture, trigger the same measurement program and then unload the part, for the measurement process to be considered complete.

From this process, you can then breakdown the different components of variation, which are repeatability, reproducibility and part-to-part (or process) variation.

Now I am going to attempt to explain how to decompose these different components of variation with as little math as possible . . . let’s see how I fare!

Starting with repeatability, which can be considered the inherent noise of the measurement system, meaning under identical conditions what is the variation in measurement result you are likely to expect.

Considering one operator, when they measure each part for three repetitions. Each part will have a mean error, which is the average result of the three measurement repetitions. Then there is a spread around this mean, which is the range of the individual measurements.

Repeatability for that operator, is calculated by removing all the mean part errors, so all the 30 repeated measurements are aligned, shown in the figure below. The standard deviation () can then be calculated which represents the by operator repeatability.

Scaling this up across the three operators, the same normalization of mean errors occurs across the 90 measurements (10 parts * 3 operators * 3 repetitions) to provide the final measurement system repeatability, see no math 🙃.

By Operator Repeatability

OK, now for reproducibility, which is the variation between operators, caused by maybe one operator loading the measurement fixture slightly differently, or mishandling a part.

To visualize reproducibility, we are going to start by taking one part, which each operator will measure for three repeat cycles. Now, because of these operator differences, the mean error for that part will be different for each operator, and the by part reproducibility is therefore the standard deviation () or variation of these means.

Once again scaling up, but this time over ten parts, each part’s true error is considered as the average of nine individual measurements taken by the three operators.

Removing the individual part errors, by normalizing or aligning them at a common center, will result in having 30 individual operator errors clustered around this common center. Overall reproducibility is then the calculated standard deviation () of these individual operator errors.

By Part Reproducibility

Bringing all this together, you will have three standard deviation values, part-to-part variation, calculated as the standard deviation () of the ten individual mean part errors, as well as repeatability and reproducibility standard deviations.

To calculate Gage R&R, you divide your measurement system variation (repeatability + reproducibility) by total variation (process + repeatability + reproducibility), giving a percentage representing the proportion of the total variation, that your measurement system accounts for.

More commonly you may want to know what percentage your measurement system variation is of your part tolerance. For this, you multiply your measurement system variation by six to represent 99.7% of the population and then divide by your tolerance.

Ideally you want your measurement system variation to be no more than 10% of your overall variation or tolerance. Although values of 20% or 30% can be considered acceptable.

Any more than 30%, then at the margins of your tolerance band you will start to question whether a part truly is in or out of specification. At this point, you have two options, improve your measurement system or guardband the specification.

Overall Gage R&R with Calculations

Improving the measurement system, can be something simple such as operator training, or improving the fixture clamping so it doesn’t move during measurement. All the way to investing in a new CMM.

Now you are not very likely to be investing in a new CMM in the short term. So, if the simpler fixes don’t work, then you will need to look at guardbanding which means tightening the specification limits to account for the excessive measurement system variation.

Therefore, if your measurement system variation is 10% greater than an acceptable GR&R, then you need to tighten your specifications by this amount.

This does not make the metrology department popular with manufacturing, who will usually ask for the opposite, to expand the specification by 10% so their part tolerance isn’t intruded upon by the frailties of the inspection 😱.

Tightening specifications protect the customer, but it does mean throwing away or reworking a lot of potentially good parts.

🔗 Correlation

Even if your measurement systems are calibrated and precise, you still need them to agree. Without correlation across suppliers, labs, and in-process checks, you risk escapes, overkill, and endless arguments about which numbers to trust.

You don’t want a parts failing your in-process inspection, passing when you measure a sample using your lab-based inspection system, termed overkill. Worse, if your supplier’s OQC is passing product only for them to fail at your IQC, termed escape.

For quantitative measurements, a process of correlation is used to check if measurement systems are aligned.

Again, trying to explain correlation visually using the diagram below. If the measurement values of different parts were plotted on a graph, with the y-axis recording the values reported by your supplier system, and the x-axis recording the values of your internal inspection.

If the results of each of your part measurements perfectly agree, you could plot a line running from zero at a gradient of 45 degrees, and all the measurement values would fall exactly on this line.

Elements of Correlation

But as you have learned, nothing is perfect and there are three types of error that can occur.

• Bias: A constant error between the two inspection systems, which causes the location of where the 45-degree line intercepts the y-axis to shift up or down.

• Span: A relationship error causing agreement between your measurement systems to increasingly diverge, altering the slope away from the one-to-one 45-degree relationship. This means disagreement is larger when part error is near the specification limits than if it was closer to nominal.

• Random Error: Error arising because of the natural measurement system variation of the two inspection systems.

Now you know that these errors occur, you need a way to quantify them and put acceptability limits on them. The first is R², which assesses how well your measurement comparison points fit on a straight line.

An R² equal to one means there is perfect fit. The closer to zero this metric gets, the poorer the fit and the more random the scatter becoming impossible to fit a line to the data. As a rule, you want your R² value to be no less than 0.9.

Your R² metric does not tell you if there is a constant bias, or if the gradient is more or less than 45-degrees. Therefore, you need a catch all to account for whether bias or span error becomes too great.

Usually mean-offset, or max-offset is used. Where mean offset, is the average difference between your measurement comparison points, while max-offset is, you guessed it, the maximum error between your measurement comparison points.

Depending on how critical the inspection is, determines whether mean or max-offset is used, and whether the acceptability limit is set at an error of 10% or 20% of tolerance.

You may be asking, I have my method and acceptability limits, but how many parts do I have to measure to calculate these values and for them to be reliable. The answer is it depends.

Ideally, you want the sample of parts you inspect for your correlation study to cover your entire process variation, so the more you measure the more of the process limits you will account for.

However, if you are producing cars, you may struggle to get more than ten vehicle bodies for comparison, while for smaller components, you may be able to run 1000 parts overnight and easily be able to get 30 or more parts for comparison, increasing your confidence in your calculated R² and offset.

Finally, you may encounter the scenario of a very tight part-to-part variation accounting for small proportion of your tolerance band.

In these cases, it will be hard to attain an R² of 0.9 or greater, because measurement results are too clustered to be able to construct an accurate line from which to calculate R². Therefore, correlation may be assessed by just max-offset value.

🛠️ Now You’re Ready to Produce Parts

Qualifying your measurement systems isn’t the glamorous part of product development, but it’s essential if you want confidence that what you’re shipping is truly in spec. Without it, you’re just guessing, and guesses don’t scale to production.

Once your gauges, fixtures, and correlation studies are in place, you finally have a foundation for reliable feedback across the value chain.

In the final part of this series, we’ll pull it all together, looking at part and assembly quality maturation, qualifying production processes, and how to use your tolerance stack as a real problem-solving tool.

📚Thanks for Reading

If you enjoyed this article and wanted to catch up on the other articles in the series, please click on below links:

• Making Your Product Manufacturable: Part III: Designing for Variation

• Making Your Product Manufacturable: Part II: Getting into Tolerance Stacks

• Making Your Product Manufacturable: Part I: Why It’s More Than Just a Tolerance Stack

Also, please feel free to connect with me on LinkedIn.

Share

Subscribe now

You’ve made it this far — and by now you’re starting to see what it really takes to make your product manufacturable.

From defining those important Critical to Quality (CTQ) features, mapping your process flow, developing datum schemes and conducting tolerance stacks to test the feasibility of each CTQ.

By the end you should have a Red, Amber, Green status of all your dimensional CTQs. For those not attaining Cp > 1.33/1.67 you know to re-examine the what-if process of assembly and part manufacture again or specify 100% inspection and rework procedures.

Now you are almost ready to begin manufacturing your product, whereby your Manufacturing Design Department develop and execute the Design for Manufacture (DFM). Following the sequence laid out in the tolerance stack, using datum features to guide their tooling design.

Design and Purchasing onboard suppliers to manufacture key parts while your Quality Department develop inspection methods to detect drift in manufacturing, as well as the resolution steps necessary to root cause and correct the associated deviation.

However, it is worth taking time out to explore a topic that does not fit neatly within the A-to-Z process being laid out. These are the design strategies that you can employ to make your product manufacturable.

🧠 What Are These Design Strategies?

If you were to design purely to optimize variation, you would have fewer more complex parts, less assembly operations, and precision fits. However, there are other considerations beyond variation, namely how you will tune production to compensate for mean shifts.

In addition, there are ways to de-sensitize your design to variation. However, they might leave your industrial design department thinking you are overstepping your jurisdiction by suggesting them.

Considering the less contentious ones first, these can be broadly separated into adjustability, design gaps, floats and clearances, and post-processing. These strategies can be used in isolation or in combination, to achieve the optimal manufacturing process.

🔧 Adjustability

Mean-shifts are deviations away from the perfect design nominal on average, rather than variation, which can be considered as the noise.

It is your job to move that average back to where it is desired in manufacturing, which may not necessarily be back to nominal, and for that you need adjustability.

Let’s take a battery frame as an example. It will have multiple CTQs, including fitment of battery modules, mounting to the vehicle body, and routing locations for coolant interfaces.

If a shipment of modules came in, whose mounting locations were off-center, you need a way to quickly compensate for this without affecting other CTQs.

By breaking down the assembly into components such as cross-members and side-rails. Then by using a combination of keyways and slip planes, you can independently adjust the location of these components in minutes.

Locations can be adjusted, by fine increments, through adding or removing shims in your tooling datum surfaces and locating pins, compensating for the mean shift in your battery module.

Further, by just adjusting the cross-member, the peripheral locations on the side-rail which line-up with the chassis mounting locations are unaffected. Therefore, by planning your adjustability effectively, means limiting the quantity of CTQs affected by any one adjustment.

Typically, the further you move downstream, the more you want to consider how you can incorporate methods of simple and quick compensation. It is here, where tolerance stacks are at their longest, and multiple mean-shifts culminate.

As well as this ability to adjust, you need a way to regularly measure the output from these processes, complemented by a steady stream of data from upstream, so that adjustments are made from a position of knowledge, not reaction.

At this point, some of you may be asking why giga castings then?

And you’re correct, on the surface these mega castings appear to be limiting adjustability. But there are other ways to manage mean-shifts and variation, and we will come back to these shortly.

📐 Design Gaps, Floats & Clearances

A key enabler for adjustability, are incorporating design gaps, floats and clearances, to provide space for both variation and adjustment to be absorbed, pushing it away from sensitive areas of the design.

Take a headlamp as an example. If it was precisely pinned to the vehicle body, the fit to hood and fender would be locked right where variation and mean-shifts are largest. The odds of a perfect cosmetic match are extremely low.

Instead, you can utilize oversized holes, slotted brackets, or floating mounts to allow for controlled movement of the headlamp during install.

You can then use a fixture, which is adjustable to correct for any drift, to position the headlamp to achieve the required gap and profile to hood and fender. The mounting points, which incorporate this float or clearance, are then torqued to set the lamp in position.

The result, a clean visual match despite upstream variation. Instead, error is concealed under the hood away from any interfacing areas.

As well as floats and clearances, small gaps in the CAD, which when it comes to reality aren’t there, can also be used to absorb variation from two complex sub-assemblies.

For example, when you are trying to join the side of your car to the floor, each sub-assembly may have 50 or more parts, that is an extremely long tolerance chain to manage.

Therefore, you want your tooling to set the relationship between these vast subassemblies, and for that you need a design gap to provide free space to accommodate variation because of poor part quality, weld draw, or adhesive thickness.

Your tooling will then close-up any gap that remains, allowing the assembly to be welded in the correct set position and orientation.

If a gap did not exist in your design, the long tolerance chain will cause parts to clash, introducing stress into the assembly. This stress causes the completed assembly to spring out of the tooling, losing the accuracy you painstakingly just set . . . sigh.

🛠 Assembly Post-Processing

Just as post-processing in computing refines a raw digital output into something polished and ready for use.

Manufacturing post-processing involves refining key interfaces and joining locations at the end of assembly steps, rather than performing multiple operations upstream on individual parts.

The advantage, well these dimensions can be formed, typically through machining or piercing operations in one go, using the assembly datum scheme, giving more control over their relationship and making them more relevant to the downstream operation.

Earlier, I mentioned giga-castings, because I knew you were thinking about it when we were discussing making all these adjustments on these individual parts.

Well, having never worked with giga-castings, I would expect it makes use of this post-processing methodology, by in effect combining several parts in a single casting, allowing all post-cast machining to be done in one or very few cells.

This enables precise control of key fitment locations, such as fascia or fender interfaces, that would otherwise be distributed across multiple components.

To illustrate further, the potential advantages of this method, I will extend our previous battery frame example. It might be that downstream you need to attach seats to the battery frame, as it forms the structural part of the vehicle floor.

You could weld the brackets that the seat will screw into onto the individual crossmembers and then weld the crossmembers together with the side rails to form the frame. However, this will mean the relationship between the four seat brackets are not directly set.

Further, if you moved the cross member to correct your previous issue with the battery module fitment. There is potential that you inadvertently cause a seat fitment issue because you have directly changed the relationship between the brackets.

Now, if you directly assembled all the seat towers into a fixture and assembled to the post-frame welding using a reference on the crossmember, the relationship is directly set.

Then, if you introduced a slip plane between the crossmembers and brackets, then any reasonable adjustment in the crossmember would not affect seat fitment, giving you independent control over your CTQs.

🤔 So Why Don’t We Do Everything This Way?

Well, the more complex the post processing, the more expensive it becomes.

For a large part such as a front fascia, fitment and datum locations will span over several components of the vehicle body.

Therefore, a metrology equipped cell is needed to measure the assembly datum points, to construct the fascia sub-datum coordinate frame used to guide the piercing tool that forms the fitment locations.

Further, the more complex the assembly, the more effort you need to make in stabilizing your assembly reference points that you will use in your metrology cell. If the relationship between the datum features is not stable, then the end result will be highly variable.

For this reason, it can be difficult to tune these post-processing cells during commissioning and ramp, where the upstream parts and assemblies are still being dialed in. This means that the ground underneath the sub-datum features is constantly moving.

A certain level of process maturity is required for these cells to be effective, and it is why start-ups and scale-ups can struggle with these cells, especially when they are launching and ramping at speed.

That is why it is best to bring these cells online later and not be fully reliant on them at the start of production.

Net Form and Pierce Post-Processing Method (Tünkers Maschinenbau GmbH)

💡 Suggestions That Make Design Go Hmmmm….

Now the methods to manage variation just discussed, may have engineering worrying about correlating crash simulation, or whether the clamp load of a washer will be sufficient with a certain clearance hole size.

But they are possibilities to be discussed and iterated over, until the right trade-offs are made for the build and product integrity.

Now the following methods of managing variation are purely cosmetic, concepts like slip planes transfer over, while others are new.

Being cosmetic options, whether they are utilized, are at the discretion of your Industrial Design Department because they affect the design language of the product.

However, if you have the chance to sit in a styling review and feeling brave, you can offer one or more of the following suggestions:

• Offset Surfaces: By deliberately offsetting one interfacing surface to another it desensitizes the user to any change in relationship between said surfaces, compared to if the interfaces matched, requiring less control in production.

• Slip Planes: Rather than locking in a part by having cosmetic interfaces on both its edges. If one edge was designed to be a slip plane, it allows the gap at the other edge to be adjusted to achieve ideal fit, without fear of affecting another interface.

• Minimizing Matching Features or Interfaces: Minimizing the number of interfacing parts that come together in a visually sensitive area reduces the number of individual controls needed to be managed in production.

• Disguising Interfaces: Concealing the interface through use of blackout at the interfacing area or by offsetting the surface so much that the interface looks hidden from the perspective it is viewed from, disguises the interface from the user.

All these principles are there to help you minimize the number of interfaces that are controlled during production, or a designed in such a way that that they don’t fight each other.

One more aspect that you need to consider for large products, is that gaps and offsets can look different when you view them from different perspectives. Therefore, you should design and measure them from the perspective of the customer as much as possible.

Angled surfaces are susceptible to this, as if you view it along the direction of the surface, the offset can look perfectly matched. However, if you then look at it straight-on, they can appear mismatched.

Be careful when you’re looking at interfaces close-up on the CAD screen, it can get you into trouble! Always think how the customer will be viewing it.

🧨 When Things Fall Apart

If you find yourself stuck with a critical matching interface, limited control in production, and a tolerance stack that still won’t solve despite several iterations.

Congratulations! You’re in rework territory.

This might mean introducing additional fitting operations, having a downstream machining step, or even looping the part back through an earlier manufacturing stage.

Whichever route you take, two things are essential. Firstly, an accurate method to assess if the product is out of specification and by how much, and second, a follow-up inspection to confirm your fix worked.

Inspection can be as simple as a quick visual check or a manual gauge measurement, or as advanced as laser or vision-based metrology systems.

The key is matching the inspection method to the resolution and accuracy of the interface you’re trying to correct, which brings us neatly to Part IV.

😮 An Unexpected Part IV

When planning these articles, it went from one article, to three, and now to four.

However, I hope the design and manufacturing strategies covered in this article have helped you see how you can make your process flexible enough to cope with all the disruption that is part and parcel of manufacturing.

Part IV is where the rubber really meets the road. We’ll dive into qualifying production and metrology equipment, and how your tolerance stacks and datum schemes can double as problem-solving tools when things inevitably go off track.

📚Thanks for Reading

If you enjoyed the article please consider sharing or subscribing. If you are not yet convinced please check out some of my other articles:

• Making Your Product Manufacturable: Part 2: Getting into Tolerance Stacks

• Making Your Product Manufacturable: Part I: Why It’s More Than Just a Tolerance Stack

• Quality Issues Rarely Start Where You Think They Do

• To build better batteries, knowing how to build phones helps.

• Abandoning Design Freeze.

Also, please feel free to connect with me on LinkedIn.

Share

Subscribe now

Since the last article, I expect you have all been looking at your favorite products in a new light. Casting a critical eye over key interfaces, wondering about their process of assembly, or the method used to fabricate a particular part.

As you’ve seen, a lot of thought goes into complex hardware devices, and we’re not even halfway through our journey of figuring out how to put all these parts together, so you don’t see that unsightly gap, or feel that annoying stepped transition between two parts.

Once you start spotting flaws, it is difficult to stop looking for them. It is like when I do my own home renovations, I know where I haven’t laid a piece of flooring quite right or smudged a paint line.

But it is a process of improvement, you learn each individual method, and the correct order in which to do things, to yield the best results.

It’s the same in manufacturing, every iteration teaches you not just the technique, but the best sequence that leads to a robust, repeatable outcome.

In this article, you’ll learn how engineers use tolerance stacks to bridge the gap between design intent and real-world build quality. But before that, there is one more layer to datum schemes I need to tell you about.

🔀 Not One Datum Scheme But Many

In Part I, you explored how, as you move through the assembly process, it becomes impractical to preserve all the original part-level datum features. Instead, only a selected subset is retained that are traceable back to upstream processes.

This can be considered as the global datum or coordinate frame for the product, one which every part is referenced in both CAD and the production process. But like any large river system, it’s not fed by a single stream.

Just as tributaries flow into the Amazon or Nile, there are sub-datum schemes that feed into this global structure.

Think of the production process of a computer chassis or vehicle body structure as the main river. A motherboard sub-assembly of a laptop or the trunk assembly of a car can be considered as a tributary that flows into the main river.

Each tributary has its own local assembly process and datum cascade until it converges with the main river. At this confluence, the two parts must align to achieve ideal fitment, which is done by matching their local sub-datum schemes.

Sub-datum, or local alignments, allow for Critical to Customer, Function or Process parameters of subsystems to be fully assessed. Evaluating them against the global coordinate frame will not provide an accurate assessment.

Taking the trunk to body structure fit in the above visual. The orange dots, which are part of the vehicle trunk local alignment, may appear out of spec when measured globally. But locally, when referenced against one another, they might have a perfect relationship.

Therefore, this local alignment more realistically describes the size and shape of the trunk opening, and when the trunk coordinate frame is aligned to it, the best possible fit for this interface is achieved.

How exactly does it do this?

Using another example, if you had a robot with a camera vision system and wanted to place the glass screen into a phone housing so that it is optimally centralized. The vision system would average the four x-reference points, shown in purple, to form the x-axis, as well as the four y-reference point, shown in blue, to form the y-axis.

The vision system would do this for both the housing and the glass and then align together their respective origins and axes to place the glass within the housing. This achieves the most even gap interface all around the periphery of the phone, as errors from top to bottom, and left to right have been averaged out.

This is exactly what happens for the trunk also, only in three dimensions and with a few more sensors and robots.

Now you know how to datum parts, tolerance relative to these datums, and do so in a consistent manner as you move through the streams and rivers of the assembly process. It is time to ask:

Will all these parts come together as intended?

To answer this question, at last, I can finally tell you about tolerance stacks 😮‍💨.

🔗 A Chain Of Events

As you move through the assembly process, more parts are introduced utilizing different fixtures, tools and people. Each additional element introduces a further component of variation to your product.

Unfortunately, you can’t just blame your supplier for this variation, it is literally part of the process.

If you pulled up engineering drawings from single part, through sub-assembly, to final assembly you will see the tolerances increase because of this reason.

To work out how this variation adds up to evaluate if a particular Critical to Customer, Function or Process specification is achievable, you need to calculate its tolerance stack.

To start, imagine you had all your engineering drawings laid out in front of you. First, pick out the drawings for the interfacing components of a critical gap, distance or alignment so they sit at opposite ends of a chain.

Then, add the links in the chain with the drawings of the parts that get added as you move through the assembly process, highlighting the relevant tolerances that influence the critical parameter you are trying to assess enroute.

Let’s use the Critical to Customer interface between a vehicle’s bright molding and sheet metal as an example. The target gap to be achieved is 1mm ±1mm to be visually acceptable.

As you can see, they have been connected using the defined assembly process.

Starting from the right, the rear quarter is connected to the bodyside via a fixture in the Body Shop. Both the vehicle bodyside and rear quarter are located onto this fixture using separate clamp surfaces and location pins, just like the example fixture you saw in Part I.

These clamp surfaces and location pins represent the datum schemes of the bodyside and rear quarter and are calibrated to be within a particular tolerance to each other.

Given slight variations, in placing the part onto the fixture, clamp forces, and free play due to clearance between the datum holes of the part and the pins of the fixture, the variation estimated from this process is ±0.25mm.

Once assembled, the vehicle structure completes the remaining Body Shop processes, passes through Paint Shop and arrives in General Assembly.

It is here, the bright molding is fitted to the bodyside using adhesive, which can be considered to act like the clamp surfaces of our locating fixture, while locating pins on the molding itself control orientation, with the rear pin being critical to control the gap to the rear quarter.

To complete the stack, the relevant tolerances on each part that influence the overall gap specification, along with the fixture tolerance previously defined are specified, illustrated in the example tolerance chain above.

You now have all the elements to calculate the expected variation. You can simply add them up, right?

Well, yes and no.

Indeed, adding each of the tolerances together is one method, known as worst-case. The second method is to calculate the Root Sum Squared (RSS).

Worst-case = 0.60 + 0.15 + 0.25 + 0.50 + 0.25 + 0.50 = ±1.75mm

RSS =√(0.60² + 0.15² + 0.25² + 0.50² + 0.25² + 0.50²) = ±1.00mm

The RSS method shows that expected variation exactly meets the Critical to Customer requirement. But if worst-case is chosen, you would literally have to go back to the drawing board and revise the assembly method.

Which one to choose? Well, the answer comes down to statistics.

Ideally, when a tolerance is specified, you‘d like to think that its limits nicely bookend the range of possible outcomes for the feature being specified.

Variation doesn’t behave quite so conveniently, and you still can’t blame your supplier for this.

Instead, it will likely be normally distributed, where values tend to cluster around the mean, and then spread out so that a very small tail will extend beyond the limits of acceptability.

In your calculation, assume the variation contained within each specification limit cover ±3σ standard deviations either side of the nominal value. This means approximately 0.3% of the produced parts will fall outside the specified range.

This 0.3% is critical to understand the difference between worst-case and RSS.

In the figure above, all individual components of variation are laid out for the bright molding gap tolerance chain, A through F.

When you add these tolerances together using a worst-case approach, you are assuming that it is possible that every element within your stack will be at its maximum limit simultaneously.

While it is possible, the chance is much smaller than 0.3%, in fact it is 0.3%^6, or 0.0007%. So you can say worst-case guarantees that each individual step stays within the 0.3% failure threshold, but the overall chance that all six fail together is tiny.

In contrast, the RSS calculation effectively mashes all these individual 0.3% failure rates together, combining them in a way that the overall bright molding gap has a 0.3% chance of failure.

Now you know the differences between the calculations, which one to choose?

It all comes down to your appetite for risk and cost. But if you want a rule of thumb, if your critical parameter is safety related, and the tolerance chain is small, say less then three components use worst-case, otherwise use RSS.

Don’t breathe a sigh of relief yet, there is still one more decision to make when it comes to numbers 🤯.

📊 Making Everything Capable

You may think because RSS tolerances are typically used, you shouldn’t care about that 0.3% defect rate?

Well, you do.

To reduce risk of defects, production processes need to demonstrate capability using the Cp/Cpk indices. These indices measure how well your calculated process limits fall within the required design specification.

Why two indices? Well, Cp measures the width of the expected process limits and compares them to the width of the design specification. While Cpk, considers both this width ratio and how well centered the process mean is.

Again, using the bright molding example, your RSS calculation perfectly matched the ±1.00mm design requirement. Assuming no mean shift, the process will look like the bell curve on the left.

However, the target is to achieve a Cp/Cpk > 1, typically 1.33 or 1.67, shown by the process on the right, reducing the risk of failure to less than 0.001%.

At this point, you may be thinking . . . ‘why did I spend all this time deliberating whether to use worst-case or RSS?’ Typically to choose RSS which has a higher failure rate, only to demand that this failure rate gets reduced 🤯.

I sympathize, but if worst-case is used, the expected failure rate depends on the length of the tolerance chain, meaning certain processes will be held to a stricter standard than others.

In addition, for complex assemblies, the tolerance chains can be extensive which mean the calculated stack result would border on the ridiculous, making it impossible to achieve the design requirement no matter how much refinement takes place.

Now that this is cleared up, what happens when your bright molding tolerance stack is good, but not quite good enough?

There are some choices open to you. Firstly, you could revisit the assembly process to identify where improvements in the tolerance chain might yield better results, and it might be the point where you can ask your supplier:

“Can you tighten the spec?”

Alternatively, visualization software can be used to assess whether an expanded specification, one that maintains Cp = 1.33 or even 1.67 capability, is visually acceptable to design.

A little insight, don’t expect the questions to your supplier or industrial designer to yield a positive response without doing a ton of homework to convince them.

If neither of those is viable, 100% inspection of the bright molding gap can be specified on the drawing, backed by an approved rework process to ensure conformance.

🔬 Getting More Sophisticated

Well done, you made it through your first tolerance stack calculation, and you did it with nothing more than a spreadsheet.

What you just did would be acceptable for most scenarios. However, there are times when you need more sophistication, particularly for complex assemblies.

For this there are specialized software that can run thousands of randomized tolerance calculations, known as Monte Carlo simulation.

In each simulation run, discrete values are randomly picked from each contributing source and stacked together, providing a more realistic picture of the tolerance distribution, particularly if it is non-normal.

Additionally, 3D assembly variation analysis tools (such as CETOL, VSA, or 3DCS) can model how parts fit together in space, capturing geometric effects like angular misalignment.

For large parts this angular misalignment can be more significant than the pure tolerance stack.

Finally, for even more advanced cases, Finite Element Analysis (FEA) can be integrated to simulate how structural loads, thermal expansion, or manufacturing stress influence dimensional outcomes.

🚀 Get Ready For Part III

If I haven’t exhausted you yet, there is one final part in this series! It is when expectation meets reality, and we will investigate practical ways to manage variation, and qualify the measurement tools needed to validate Critical to Customer, Function or Process specifications and assess their capability.

Only then will you truly know, is your product manufacturable?

📚Thanks for Reading

If you enjoyed the article please consider sharing or subscribing. If you are not yet convinced please check out some of my other articles:

• Making Your Product Manufacturable: Part I: Why It’s More Than Just a Tolerance Stack

• Quality Issues Rarely Start Where You Think They Do

• To build better batteries, knowing how to build phones helps.

• Abandoning Design Freeze.

• Where are our Hardware Unicorns?

Also, please feel free to connect with me on LinkedIn.

Share

Subscribe now

Have you ever stopped to wonder why those fantastic renderings you see of the latest concept car, mobile phone, or laptop, may not live up to reality once you see it in the flesh?

Maybe the color doesn’t pop as much, or the feedback from the on/off button feels tacky and light, or it could be the gaps between the edge of the glass and the metal housing of your computer display may appear uneven.

There are all sorts of reasons why a product may let us down in real life, and companies put a lot of effort into closing that expectation gap and even add features for you to discover that elevate your perception once the product is in your hands.

⚠️ Nothing is Perfect in Manufacturing

Now, coming back to the poor fit of the glass within your expensive computer display, why could this be?

If you think about it long enough, there are several reasons. The length and width of the glass, the straightness of the edges, the evenness of the black ink decoration around the edge of the glass, which is applied in part to disguise said gap, or how well the metal housing is located within the CNC fixture as its recess is milled out.

There are many reasons but one common theme, dimensions. If we cannot control the dimensions of individual parts and how they are assembled, then our manufacturing processes will be inefficient with endless amounts of rework just to get parts to fit together. The product won’t look well put together and eventually break.

If we could fabricate every part to be exactly the same size and shape of the CAD design and then locate them together with absolute perfection, we wouldn’t have a problem. But our world is not like that, even the most precise processes have variation.

Take ASML’s photolithography machines used in semiconductor manufacturing. These systems can etch features onto a silicon wafer with a pitch as small as 13.5 nanometers. Yet even they account for up to 2 nanometers of variation, in the precise placement of each layer.

Now you’re thinking, well nanometers are close enough, why can’t we do that for everything? Well, for a start photolithography is a very specialized process. Even if we could use it elsewhere, we cannot afford to distribute $400million machines throughout our production line.

Cost, weight, performance, and cycle time all play a role in determining whether a certain method is more suited to a specific manufacturing process. For instance, welding sheet metal panels together to make a car body is better in all four categories than CNC machining the same car body out of a billet of aluminum. Even though a CNC process is more accurate and less variable.

🏁 Start with the End in Mind

For complex products, managing variation begins in the design phase. When done well, it links product requirements to manufacturing capability and quality control.

When your industrial design or styling department unveils their latest foam or clay model, if you look at it with a critical eye, you will spot a whole host of features that need to be controlled with precision. If not, it will lead to customer dissatisfaction.

Typically, for dimensions this will get quantified as a gap plan, where the nominal or perfect dimensions are specified between critical interfaces typically in a brochure or drawing. Examples for a car include, the gap between the headlamp and hood, symmetry of gap either side of the headlamp, or the flush/offset between the front and rear doors.

As well as the nominal value, a tolerance is placed on each dimension, which is the allowance given to the manufacturing process for how much it can deviate from the nominal value in production, typically within three quarters of a millimeter for a car, tens of microns for a phone.

Now, there are ways you can desensitize these interfaces, so they are less visible to customer such as the black ink decoration border around the edge of the glass in our computer display example. For a car, offsetting the surface between the headlamp and hood, so the panels don’t have to be perfectly matched.

The level of desensitization is a design decision, and how much this can be influenced by product engineering or manufacturing depends upon the company. Try telling Jony Ive back in the day that a certain feature on the latest MacBook is unmanufacturable and you would get short shrift, without pouring a lot of blood, sweat and tears in proving him wrong.

This is only the first level of Critical to Quality (CTQ) parameters.

🌊 Cascade of Requirements

As the product moves into engineering and broken out into parts and sub-assemblies, further Critical to Customer, Function or Process, parameters are defined:

• Critical to Customer: width, length and profile of the edge of the on/off button to maintain a tight and even gap to the phone housing.

• Critical to Function: Position and diameter of the input, intermediate and output shaft bores in an electric drive unit housing to ensure ideal gear meshing.

• Critical to Process: Gap between the chassis and body so when they marry together in production they won’t clash.

At this point, it is vital that product engineering and manufacturing are in communication, because engineering won’t necessarily know capability of the production equipment, or how and in what order certain parts will be assembled.

Equally, manufacturing may not be aware of sensitive areas of the design that have to be controlled or protected in production. Oftentimes, on first pass there may be conflict between the interests of the product and that of the process.

For instance, a critical dimension between two castings within a battery pack may have to be maintained to comply with crash performance, and this must be considered in the Design for Manufacture (DFM).

And so, begins the iterative process of datum selection, part tolerancing, process planning and tolerance stack calculations.

Often a tolerance stack is used as a pass/fail gateway done in isolation, rather than a what-if process, where the inputs are questioned and trialed virtually to make the best overall decision for both product and process, bridging conflicts that may exist.

🔍 Getting into the Detail

The what-if process starts by analyzing where the datum features are positions relative to CTQ features. In theory, datum features are references on your part through which a coordinate frame is constructed, and all other features are referenced.

What this means is, they are the features you clamp, rest or locate the part against, to ensure that the part is locked in a repeatable orientation as it is machined, welded, etc. This minimizes variation by constraining all six-degrees of freedom (translations in and rotation about the x, y, z axes).

Let’s say you are laser welding two thin metal panels together. For the weld to be effective the gap and profile between the panels need to be tightly controlled. However, if the datum hole used to locate the part is far from the weld seam, any twist or bow in the parts will be amplified in the area that matters.

That’s why, in discussion between product and manufacturing engineering, the decision will be made to place the datum hole as close to the weld seam as possible minimizing both variation and geometric effects. The latter, a simple tolerance stack won’t catch.

In addition, experience counts, knowing when and where certain principles can be violated to help the process can prove beneficial.

For instance, because the panels being welded together may not be rigid, there will be localized sag. The engineer can use this to place more clamps along the weld edge, deforming the part to its ideal shape during welding, improving the result, even if it technically violates the six degrees of freedom rule.

In conjunction with datum selection, the order in which parts are assembled becomes critical. The goal here is to ask whether parts can be combined, such as moving from several panels to a machined casting.

Alternatively, can the process be re-ordered so that the constituent parts that influence a critical dimension are closer in sequence, shortening the tolerance chain in your calculation.

📊 From Requirements to Specifications that Work

Once your process sequence and associated datum features are selected, they can be annotated on your CAD model or drawing. Other features can have their tolerance defined relative to these datums using Geometric Dimensioning & Tolerancing (GD&T).

Typically, to control six degrees of freedom, a minimum of three datum features need to be defined, whereby the datum controlling the most degrees of freedom is denoted first using the letter ‘A’ followed by ‘B’ and ‘C’ for the secondary and tertiary datums.

Additional letters are then used to denote where local control is required of non-rigid parts, or for sub-datum schemes used for localized assemblies, which will be discussed further in Part II.

As you move through the process, it is impossible to maintain all the individual part datum features, as the downstream sub-assemblies would simply have too many. Instead, only a subset of datum features is maintained, those that are physically accessible and traceable back to upstream processes.

Adding in new datum features breaks the logic of the production sequence, and processes will no longer talk to each other.

What this means is, if you defined the coordinate system on an assembly using one set of datums, and then abandoning them for another set later. The results from the first coordinate frame, will have no relation to the second coordinate frame making it extremely difficult to debug your manufacturing process as you commission and ramp production.

🧮 Get Ready for Part II

Once you have done all these things, you are ready for the tolerance stack. But just like a Substack article, a tolerance stack should not be too long, and so you will have to wait until part II! In this article, we will move from tolerance stack through to the following activities that need to be done to make your product manufacturable.

📚Thanks for Reading

If you enjoyed the article please consider sharing or subscribing. If you are not yet convinced please check out some of my other articles:

• Quality Issues Rarely Start Where You Think They Do

• To build better batteries, knowing how to build phones helps.

• Abandoning Design Freeze.

• Where are our Hardware Unicorns?

Also, please feel free to connect with me on LinkedIn.

Subscribe now

Share

Manufacturing issues rarely originate where they first appear. Often, they stem from a long and overlooked chain of events.

It starts in Stamping. A flat sheet of metal is passed through a series of dies to press, pierce, and trim it into a door panel. But the draw die was set slightly out of balance, causing the shape of this batch of panels to differ slightly.

This subtle deviation has a cascade of downstream effects.

Now the panel has reached Body Shop. It sits differently in the hemming fixture, altering the form further, affecting the fit to the vehicle.

Before the vehicle moves to paint, the changed fit condition makes it difficult for the production associate to apply the jig that holds the door in position through the multiple dips and ovens of the paint process.

The door warps further as it is baked.

Now it’s in General Assembly. The door no longer sits flush with the fender. A fitter, under pressure to keep the line moving, improvises a fix by adjusting the fender, and in turn the hood.

At the next station, the hood doesn’t cinch properly. The latch is adjusted. Now the gap between the hood and headlight grows. The issue is flagged at final audit.

A quality concern is raised, for an issue that began days earlier in stamping.

🔍 Now begins root cause analysis

My example of how a quality issue manifests itself might be slightly exaggerated, but I am sure there are some of you out there that lives this on a regular occurrence.

For everyone else, bear with me as I extend the chain of events, to how the problem resolution process plays out.

Now, Quality and Process Engineers are likely to be trained in root cause analysis processes, fishbone diagrams, 8D, A3s and alike. But these days, the process usually begins by raising a JIRA ticket and assigning an owner.

Now that owner is under pressure to show some progress, with the first step being containment.

Correctly, they go to investigate the fit process. After some discussions with the fitters, they find a few tweaks that reduce the issue. It gets the begrudging approval of the quality auditor and vehicles can keep shipping. But the underlying condition still needs resolution.

Then the night shift comes in.

There’s no pass-down. The issue flares up again. At 2 a.m., the quality engineer gets a call, multiple cars have hit rework. They make their way in, and present the containment to the night shift, and production begins flowing again.

Stressful, right? And we haven’t even started to investigate root cause.

Fast forward a few days. The containment mostly holds, despite occasional escapes. The quality engineer finally gets what they think are the most likely parts measured, the hood and fender. Both are within spec.

While waiting for the hood and fender results to return from metrology, they search around and find some continuous data collection of gap and flush measurements between the body panels in Body Shop, prior to being shipped to paint.

Aha, they see a shift in the trend of the flush measurement between the front door and fender the previous week, but it has since returned to normal.

They speak to associates and hear anecdotal reports:

“Yeah, last week’s doors felt different.”

However, without traceability, there’s no way to isolate the door assemblies. Instead, they grab a random sample from the line for measurement and request inspection of the hemming fixture.

They update the JIRA ticket - waiting on measurements - enough progress to get them through this week’s quality review.

But then there are reports of cars in general assembly failing containment, first the occasional black car, then a steady stream of white and blue ones.

The quality engineer, now armed with partial data and a bit more confidence, proposes a test, remove the containment for ten vehicles, check fit, and review.

Eight out of ten cars pass, and it is agreed to remove the containment and send anything failing to rework.

A few more days pass. No major incidents, fewer and fewer cars are routed for rework, except a batch of yellow cars which mysteriously showed up all at once with the old issue. But that passes too.

Surprisingly, the issue seems resolved.

Eventually, the remaining measurements come back, Nothing conclusive. The only real insight remains that shift in the door fit chart from Body Shop.

The engineer uploads the chart to JIRA, along with a note:

"There was a process change in Body Shop that affected door fit. Recommend reaction plan if metrics go out of control again."

The ticket is closed, without the investigation ever progressing to Stamping.

And the quality engineer sleeps peacefully, until the next time the die isn’t balanced correctly.

⚙️ How can this chain of events be improved?

Industry 4.0 and AI hold promise, but being able to leverage these tools is an end result. Less is discussed about the process of getting there.

While this article focuses on data, there is a whole skills and attitude aspect as well. No quality process or data insight will offset the skill of logically and mechanically thinking through a problem and getting hands on trying to tangibly see how a process behaves.

But the reality is that while there are forward thinking manufacturers that have live dashboards showing manufacturing health, and human-in-the-loop feedback alerting engineers of process deviations that can be dealt with before it becomes a quality concern.

Over 90% of the manufacturing sector are Small and Medium Enterprises (SMEs) and the vision of Industry 4.0, is far out of reach.

What does exist is a scattered mess of proprietary software, spreadsheets, PDFs, and whiteboards. Measurement data lives in Excel. Maintenance logs are scribbled in notebooks. Quality escapes are buried in a poorly organized JIRA system.

Process engineers are stuck cutting and pasting across tabs, trying to link information that was never designed to be connected.

Like your product design, production system and Manufacturing Execution System (MES), your quality and process data system should be designed from the outset.

Where should measurements be collected? What process parameters are critical? Where should sensors be placed within the production line? These are all questions that should be asked, along with the specification of data formats, traceability, database architecture and quality reports.

Now with Large Language Models (LLM) there is a place to centralize informal communication, slack messages, maintenance logs, meeting transcriptions. All this can be analyzed along with quantitative data to remind the team of that shim move that was made two weeks ago, or the robot that had to be reprogrammed following a fault.

All of which would add context for the poor engineer in our example. The maintenance team in Stamping may have recorded an anomaly when they loaded the door die. This could then be tied to the batch of doors that caused the quality concern.

But there are barriers. Propriety software wrapped in layers of permissions, tiers, guardrails, and configuration steps. I remember being sent a quote for $5k just to be able to get a poorly formatted excel from a gear measurement machine.

Then there are SaaS tools tailored for data scientists, tools that overlook the needs of quality and process engineers, and end up pushing teams back to tried-and-trusted Excel.

These are problems which can be overcome with discipline. If the right data is captured early, engineers don’t have to start from scratch every time an issue shows up. Instead, they can look across connected datasets, spot patterns, and apply lessons learned.

🧩 Start Where Engineers Start

The goal isn’t just automation. It’s not AI for AI’s sake. It’s building systems that meet engineers where they are, on the floor, in the problem, figuring things out.

That means tools that surface the right data at the right time, preserve context across shifts, and make it easier to solve problems.

We need systems designed to support the messy, human-centered work of engineering problem solving, helping people connect the dots, remember what happened last time, and fix things properly.

Because the future isn’t just AI-driven. It’s engineer-enabled.

📚Thanks for Reading

If you enjoyed the article please consider sharing or subscribing. If you are not yet convinced please check out some of my other articles:

• To build better batteries, knowing how to build phones helps.

• Abandoning Design Freeze.

• Where are our Hardware Unicorns?

Also, please feel free to connect with me on LinkedIn.

Subscribe now

Share

Can you think of what high-value manufacturing sectors are still present in the UK, US or Europe?

You will probably guess at automotive and aerospace, and some of you might stretch to defense, satellite technology and pharmaceuticals. In addition, as investment from the ‘Chips Act’ comes online, semiconductor fab will soon join this list in the US.

Analyzing the production rates of these industries, they range from 75 jobs per hour for high volume automotive, all the way down to one job every few years for some defense projects, like ships.

Typical single production line capacity of different industries.

These volumes are fine for the products that they produce, but there is a problem, we now want to get good at making batteries whose throughput is measured in thousands of jobs per hour from a single production line.

Skills from automotive or clean energy tech that need to be adapted to incorporate batteries within their production lines or supply chain don’t easily transfer. Stacking of electrodes, electrode insertion, filling, sealing, degassing, placing cells into modules, laser welding hundreds of cells to a busbar, all require a different sort of automation.

What critical industry is China, Korea and Japan so strong that allowed them to transfer over to become world leaders in manufacturing batteries?

Well, if you have never built consumer electronics at scale, you’ll struggle to make batteries.

📉 What was Lost

Until recently, there has not been much desire to rejuvenate a domestic consumer electronics sector in the West. But in the late 1990s, the UK did harbor ambitions to develop homegrown manufacturing, like our ambitions today to bring gigafactories online.

Wales was the hub of this development with Sony and LG investing in new factories and producing televisions under their own brands, as well as products like iMac under contract for Apple.

However, a factory does not make an ecosystem, and many components had to be imported, also the lack of skilled manufacturing engineers meant that Welsh plants could not keep up with their sister plants in Asia implementing automation, solving quality issues and developing local supply chains.

Indeed, as Patrick McGee’s book ‘Apple in China’ highlights, the Welsh iMac assembly line reportedly became known as the ‘toaster line’ for its tendency to burst into flames.

And so, one by one, the factories disappeared, with LG withdrawing in 2003 followed by Sony a few years later, limiting South Wales options in transitioning to alternative industries with the closure of its coal mines.

Even burgeoning local companies such as Dyson, with a more vested interest to make British manufacturing work, found the UK’s supply base too fragmented and planning laws too onerous to bring new capacity online effectively.

Without an ear in government that truly understood the requirements of industry, Dyson moved manufacturing to Malaysia in 2002.

🚀 How Asia moved on

In the intervening 25 years since the UK and US could say that it was competitive with Asia in the manufacture of TVs and Laptops. The sector has moved on with the development of smartphones, wearables and tablets, which the West simply has minimal experience in manufacturing.

An increasing consumer base, development of new products and disruption of existing ones, has seen the value of the market increase five-fold since 2000, to $1 trillion.

TVs have transformed from 130 million Cathode Ray Tubes in 2000, to 250 million flatscreen devices today, while smartphones have gone from a standing start to annual sales of 1.5 billion units.

In that time, consumer tastes have become more demanding, as well as the entrants of new competitors, has meant companies needed to scale, automate and invest in quality all over a short period to remain competitive producing ever more complex devices.

Vast industrial complexes were built, containing 30 or more buildings each with ten floors that contained a mind-boggling amount of production capacity.

These facilities became specialized at going from sketch to prototype, and then from prototype to hundreds of thousands of devices per week, all within a 12-month period.

Typical consumer electronics production site.

📘 What was the Playbook?

It starts with designers sitting side-by-side with manufacturing engineers and contract manufacturers figuring out how to make one perfect unit. Every detail is scrutinized, surface finish, tolerances, cleanliness, even the repeatability of the inspection method.

As each process step is defined, the ‘machine that makes the machine’ is developed. Examples include, developing the CNC program to machine a laptop housing, along with the fixture and tooling to go with it, qualifying it and then rolling it out to the hundreds of other CNCs that are situated on the factory floor.

Alternatively, it may require developing a whole new automated line consisting of several stations linked together with linear XY stages or robots to transfer parts between stations. Each of these stations is specialized to do one assembly operation.

These stations may also integrate (a) devices to perform electrical or leak tests, (b) laser line sensors to perform in-process metrology of critical dimensions, or (c) leverage machine learning in combination with a vision sensor to reject defective parts from the line before further value is added to them.

What is common about both CNC and modular pick and place systems is how quickly they can be reconfigured. CNC programs can be updated, requalified and rolled out. Automated machines can be copied to increase scale, their stations reconfigured, and fixtures swapped out to manufacture product variations or updated designs.

Billions of dollars have been poured into building skill and capacity, with machines capable of producing parts with micron-level accuracy. They can be repurposed quickly to spread the efficiency of the cap-ex investment over multiple product generations.

Automation has reduced the quantity of people you see on the factory floor. Even Quality Control, one of the last places you still see rows upon rows of people inspecting the product before shipment, is being replaced by in-process vision inspection.

Modular automated consumer electronic assembly cell (DWFritz).

🔋 Applying the Playbook to Batteries

Production lines for batteries, particularly cell and module assembly, mirror those developed for consumer electronics. Stations are linked by robotic handlers, linear stages or conveyors. Each dedicated to a specific task such as welding, bonding, testing, cleaning, etc.

Processes are tightly controlled and validated with a level of detail not seen in other areas of automotive assembly. For example, after plasma cleaning of cell terminals, vision systems inspect for micron-sized dust particles which, if present, could affect contact between the cell and the busbar.

Testing is continuous throughout the supply chain and in-house production. Terminal cleanliness is inspected at both cell casing assembly and prior to laser welding within the module line. If particles are found, cells are re-routed to be cleaned again or removed from the production line.

The concept of modularity has also transferred across, with whole stations being able to be removed or added, fixtures swapped out and robots being able to be reprogrammed to use a new custom end-effector. When a design change is introduced, updates can be batched as part of an agile process and introduced to the line.

The tooling, the line design, the reconfiguration strategy is all part of the manufacturing innovation which needs to constantly evolve to keep up with the pace of technology change within the batteries themselves.

🚗 Going Forward

Asia is way ahead in terms of its ability to manufacture current generation lithium-ion batteries, be it China’s dominance in LFP or Korea’s and Japan’s edge in NMC. Therefore, it makes sense for automakers to partner with the giants be it CATL, LG-Chem, Panasonic or Samsung.

However, there is a whole raft of innovations coming down the pipeline, such as lithium-metal solid-state cells or nanoscale silicon anode cells, which promise greater energy density and faster charge speeds.

There are startups around the globe trying to get ahead in this race, to do so means that a whole host of new companies are going to have to adapt the battery manufacturing playbook to their own designs.

Simply trying to outsource to the main players without significant collaboration won’t work because of how linked design and manufacturing is. Too many start-ups die in “manufacturing purgatory” where scaling-up hits hurdles, yields fail to improve and as a result firms come to the end of their funding runway.

As Dr. Halle Cheeseman states:

“Even excellent technologies can’t survive the manufacturing purgatory stateside. By the time they launch, the market’s moved on.”

Whether we are trying to build our own gigafactories or scale new chemistries, consumer electronics manufacturing thinking needs to be applied to battery manufacturing.

We have a lot to learn, and it takes time to master high-speed automation, in-process inspection, rapid DFM cycles, modular production lines and tooling that can accommodate design iterations.

But you must start somewhere, and we are starting.

Thanks for Reading!

- Glen from WattFactory

The concept of design-freeze in product development was one which I was taught at an early stage of my career. It is the point at which Design Release Engineers down their virtual CAD pens and upload final designs to the company Product Lifecycle Management (PLM) system.

This freeze is considered essential to allow all designs to be released to part suppliers, tooling vendors and equipment integrators to fabricate all the elements of the manufacturing system and bring them together in a coordinated manner to produce a complete car, plane, rocket, etc.

🏗️ A Rigid Process

Prior to the advent of gigacastings, a car body structure typically consisted of 400–500 individual sheet metal panels, assembled by a comparable number of robots.

To coordinate the complex set of activities required to assemble these hundreds of interdependent components into a robust final product, legacy automakers have long relied on their own proprietary versions of a tightly controlled Stage-Gate process.

In this process, the product design advances through a series of predefined milestones, or stage-gates (shown as the red diamonds in Figure 1). These gates mark progress from initial market positioning, through concept development and detailed design, followed by testing & validation, prototype manufacturing, and finally into full-scale volume production.

Each milestone comes with a defined list of deliverables, rigorously assessed in the often-dreaded stage-gate review, where a go/no-go decision will be made on whether to progress the vehicle program through to the next development stage.

As the product design makes its way through each stage gate it steadily grows in maturity, until it reaches Concept Confirmation. At this point, the vehicle is considered fully defined digitally.

Figure 1: Magna Product Development V-Model: https://www.magna.com/stories/inside-automotive/concept-creation/new-car-development-process.

📐 The V-Model

The increasing level of detail is structured using the V-model, where the design cascades down the left-hand side of the diagram, shown in Figure 1, from customer level to the point where every individual component has its own full set of specifications at the bottom of the ‘V’.

Between Concept Confirmation and Functional Confirmation, the design is validated using a range of digital simulations and physical models. Virtual tools include Computer Aided Engineering (CAE) to assess crash performance. Virtual Engineering (VE) to simulate the manufacturing process and Dimensional Variation Analysis (DVA) to model assembly tolerances as they stack through the manufacturing process.

Physical tools are used to complement virtual validation, with the fabrication of models such as the ‘Function Cube’, shown in Figure 2. This is a CNC machined exterior and interior model representing the visible surfaces of the CAD engineered product.

This model serves to validate the digital geometry before tooling kick-off and provides an opportunity to assess the physical look of vehicle, which can be evaluated under different lighting conditions to assess door gaps, feature lines, component interfaces, etc.

Figure 2: Example of a CNC milled function cube model of a vehicle.

❄️ Design Freeze

Once virtual and physical validation is complete, final CAD changes can be made, and the magical design freeze is reached. Design Release Engineers breathe a sigh of relief and hand-over, often several weeks later than intended, to Manufacturing or Supply Base Engineers to get all the bits manufactured, installed and bought-off for prototype and volume production.

There are imperfections with this process, namely being late changes, slippage in timing and the processes failing to acknowledge the to and fro between design and validation. But broadly these deficiencies were masked because all automakers were following variations of the same stage-gate process.

But this began to unravel when EV start-ups in the US and China applied lessons from the software and consumer electronics worlds which pushed speed and iteration over perfection.

⚡ What did EV Firms do differently?

While the established Stage-Gate process did ensure a robust product was released to the market, with steadily improving quality and safety with each iteration, it took approximately four years to progress from project Kick-Off to Launch.

EV firms and new automakers in Silicon Valley and China, while being unincumbered by legacy processes, had to learn how to build a vehicle from scratch.

Due to their proximity to the software and consumer electronics industry, they leant on these industries to guide their product development efforts.

🔁 Agility & Iteration

I remember my first months on a new project at Tesla back in 2014, where I had the expectation of attending multiple project reviews, reporting back progress against a predefined list of activities.

Instead, I was exposed to JIRA for the first time, where an ill-defined list of activities was bucketed together to be completed in a compressed timeline which bordered between aggressive and ridiculous.

It was my job to flesh out my own activities and work with other Design, Quality and Manufacturing functions to ensure the manufacturing process was planned out so that the product could be assembled within a given tolerance range, while maintaining the integrity of the product design.

This required developing multiple options, often involving design variations. However, with improvements in simulation software and compute power, selected options could be crash simulated and iterated for improvement.

While this was very rough initially, the process permitted learning, required cross-function communication via working meetings with engineers, and leveraged advances in virtual testing allowing for more iterations without the need for formal change requests.

In time, organizational memory grew so that interactions became more instinctive and through learning there was a clearer awareness of required deliverables, while still leaving room to explore new ideas.

This agility differs from the stage-gate system, prioritizing fast coordinated iteration cycles above pushing for a single perfect design release allowing issues to be discovered faster and permitting changes to be managed locally. This is especially critical when the product and production system is being debugged during manufacturing qualification and ramp.

🧩 The Power of Modularity

The speed of iteration in the design realm, along with multiple design updates, meant that EV entrants had to learn to apply the principle of modularity within manufacturing.

Just like in software, where code is broken into discrete functions that can be reused, swapped, or updated independently, allowing developers to improve or add isolated components while preserving the broader architecture, this modular mindset can be applied to hardware as well.

The first evidence of modular hardware in manufacturing was first seen on a large scale in consumer electronics, and is cascading into automotive production particularly in newer areas such as battery module assembly.

A modular manufacturing line is characterized by discrete, reconfigurable stations linked by robot or linear transfer systems, which enable flexibility in the face of inevitable design changes.

Without modularity, late-stage design modifications can render entire stations obsolete. Utilizing swappable fixtures, adjustable tooling, and standardized end-effectors enables changes to be accommodated quickly.

This flexibility protects capital investment, maintains product launch timing, and accelerates iteration, turning rigid production lines into adaptable platforms for continuous improvement.

♻️ Modularity: A Virtuous Cycle

The benefits of modularity have flowed back into the digital space with standardized components being designed that can quickly plug into a customized manufacturing equipment design.

This has allowed equipment vendors in China in conjunction with their high quantity of skilled labor to produce a fully detailed production line design in as little as 30 days, in comparison to the three to four months of traditional equipment vendors.

This time compression continues, where both standardized and custom components can be released to a dense network of suppliers that work from digital files, rather than 2D drawings to fulfill orders in short lead times.

The overall time compression has seen some automakers achieve product launches in two years, rather than the standard four.

🧠 First Principles over Established Processes

In my early experience as an apprentice, I saw how engineers responsible for commissioning production lines focused primarily on project management and sign-off criteria. The technical work of debugging, tuning and process optimization was largely left to the equipment vendor. At Tesla, Apple, and Rivian, my experience of this dynamic shifted significantly.

A lack of established standards and/or due to the novelty of the equipment meant that commissioning demanded a deep understanding of the manufacturing equipment itself. Engineers had to apply first principles thinking to troubleshoot in real time and pivot quickly as designs evolved.

For hardware disruptors, such as EV startups, it is necessary to cultivate this first principles thinking by hiring individuals who are capable of deconstructing complex problems to their fundamental components. This approach not only fosters quicker iteration cycles but also provides a source of learning and innovation.

With this hard-won knowledge, engineering teams are better equipped to scrutinize every part, process, and specification and apply the same thinking in the design phase to drive out unnecessary complexity in second-generation products.

A good example is SpaceX’s evolution of the Raptor rocket engine, where learnings from the first generation led to a design evolution that became simpler with each iteration, shown in Figure 3.

Figure 3: Space-X Raptor Evolution Reducing Design & Manufacturing Complexity

🚀 Design is Never Done: Embracing the Iterative Edge

My experiences in automotive start-ups and consumer electronics companies have changed how I think about industrial product development. While traditional stage-gate processes brought consistency and accountability, they were designed for a slower era—one where designs could be locked down, changes minimized, and longer lead times tolerated.

But as technologies became more complex and timelines more compressed, rigid workflows begin to act as constraints rather than safeguards. Modularity, iterative development, and deep process understanding are essential ingredients for speed, resilience, and learning.

Ultimately, the goal should not be to reach a perfect design freeze, but to create a system that embraces controlled evolution. The most successful companies treat each product launch not as a fixed endpoint but as a launchpad for continual improvement, where every iteration sharpens the design, strengthens the process, and accelerates learning.

Whether developing rocket engines or electric vehicles, the ability to question assumptions, simplify intelligently, and adapt quickly will define the next era of hardware innovation.

The potential for AI in manufacturing is undeniable, evident by the explosion of SaaS startups entering the space. Over 350 companies have now progressed to at least Series A funding, each aiming to disrupt incumbents such as Dassault Systèmes and Siemens, as well as relative new kids Atlassian (hello JIRA 😊).

Recently, I started to question the proliferation of these platforms, starting with how much does all this cost?

It feels like manufacturers are being offered a plethora of discrete solutions ranging from AI-powered CAD and PLM tools to MES platforms, EHS compliance, and beyond. Often leaving the customer to work out how to integrate everything together.

It’s the equivalent of me subscribing to a Notion, Todoist, ChatGPT and twenty more productivity apps, each with a different promise, with a vague hope they’ll somehow hang together to lift every part of my life.

🛠️ Finding Manufacturing Customers is Hard

Selling to manufacturing is a very different game compared to design, finance, or tech firms. Operationally, manufacturers question every cent, and SaaS subscriptions need to at least ‘wash their face’ demonstrating a tangible benefit to their customer, ideally in terms of dollars saved per part.

Manufacturers are like frugal consumers, scrutinizing every line of their energy bill, in contrast to the tech sector, eager to blow their budget to get their hands on every shiny new release. This is especially true in contract manufacturing or among tier-one and tier-two suppliers, where margins are razor thin and volume is king.

Nowhere is this truer than in China, where competition is fierce and major customers like BYD and Huawei have clear visibility into material and labor costs, enabling them to negotiate aggressive terms.

Contract manufacturers, acutely aware of this cost vs. value equation, build their own software platforms tailored to their needs, skipping the license fees altogether, protecting their razor thin margins.

If SaaS providers want to win in manufacturing, they must think beyond selling isolated features that can be developed in-house. Instead, they must embed themselves deeply into workflows, capital projects, and evolving the way manufacturers design, build, and scale their operations.

So, what are the key considerations that manufacturing SaaS firms need to cognizant of when finding their value proposition?

📐 Integrate Early

The upstream phases of product development are where designers and engineers focus on embedding value into the product. Here, companies are more open to adopting specialized tools, if the value proposition is clear.

Key manufacturing decisions around tolerances, materials, and process flows are made in these upstream phases of product development.

If SaaS providers can help define the process parameters for a manufacturing digital twin, one that carries seamlessly into production with minimal on-cost, they can dramatically lower the barrier to adoption.

Another strategy is to work with equipment vendors, the companies that ‘make the machines that make the machine’, to embed digital tools as part of the production line's initial capex purchase, rather than trying to retrofit software after production begins.

🧠 Help Customers be AI-Ready

Many SaaS firms are leveraging the benefits of AI for tasks like defect detection, optimizing asset utilization, or generate process instructions. But even when these tools are integrated early via simulations, they often diverge from reality once production begins.

To provide manufacturers with the opportunity to course correct quickly, data infrastructure must scale in line with production. This will allow information from closed loop feedback and AI to be leveraged, updating simulation assumptions and improving process parameters.

Yet many firms experience a data 'cold start' problem where they haven’t defined sensor placements for in-process measurements, data standards, or storage systems needed to support real-time learning.

Without forethought during the design-for-manufacturing phase, critical systems like traceability, in-line metrology, and communication protocols are often left undefined or poorly aligned with actual decision-making needs.

To overcome this, SaaS providers must help firms treat data system architecture as a design deliverable, not an afterthought. Traceability, in-process metrics, and communication protocols must be defined early and refined through prototyping and early build phases.

📊 Build Full Workflows that Capture and Amplify Learning

Working with manufacturers early-on to define data collection, reporting metrics and in-process closed loop control requirements, will help SaaS firms to build complete workflows rather than discrete apps.

To optimize the entire production line, it is important to consider manufacturing problems are rarely isolated. In complex products, contributors to a defect can be spaced hours or even days apart in the process.

Specialized apps can silo this data, preventing the system from learning as a complete entity. Just like my productivity apps, my sleep schedule may be optimized without consideration of my work or exercise regime.

Being able to capture all the learning from the subtle adjustments that are made to solve issues and continually improve the product, will allow manufacturing knowledge to accumulate and strengthen over time.

The genesis of these micro-innovations often begins through offline analysis, Slack threads, and production line trials. But in so doing, create a shadow knowledge that is never fully integrated into the manufacturing data architecture.

Being able to capture this data to build large language models (LLMs) trained on real operational history, will help teams query past fixes and avoid solving the same problem twice.

If we think of the future of manufacturing, a few highly skilled individuals will supervise intelligent systems. Giving them tools, such as LLMs, to interact with those systems through natural language to troubleshoot fluctuations and refine operations will be a critical enabler of a smarter platform.

🤖 Build for a Robotic Future

Production lines today are already highly automated, with robots executing precise sequences of welding, bolting, and riveting, adjusting dynamically to sensor feedback.

The introduction of humanoid robots on the shop floor will open-up the number of operations that can be automated, such as assembly operations with limited access, or using flimsy parts.

Initially, they’ll behave like today’s robots, trained to do repeated task, using integrated vision sensors to adjust for process variability, to ensure the operation is completed right first time.

But how will they interact with humans? Being AI driven, they will use natural language rather than being programmed through a terminal.

Also, in time will they be isolated performing one operation?

The inevitable push will be for these robots to be connected to the factory floor, taking in data from upstream and downstream operations. Correlating what they are sensing in their own process to communicate issues and fixes up and down the production line.

Eventually, they may be flexible enough to move out of station, to investigate problems for themselves acquiring more data by performing multiple different operations to seek out the cause for anomalies.

I have presented my vision of the future production line. Manufacturing SaaS companies will have to gaze into the future to understand themselves what it will look like and adapt their software tools to integrate with this robotic future.

🌐 Conclusion: From Tools to Infrastructure

The future of manufacturing will be powered by intelligence embedded into the infrastructure of how we design, build, and iterate. For SaaS firms, this means moving beyond individual apps and dashboards and toward platforms that help manufacturers learn, adapt, and improve, continuously and autonomously.

This is creating a new type of manufacturing intelligence, one where knowledge compounds over time and naturally flows into production optimization, design iteration and the specification of future products.

Manufacturing data will continue to evolve towards a robotic future, where a few skilled individuals collaborate with AI-driven automation through natural language. Guiding and reconfiguring flexible production lines in real time as part of an intelligent, adaptive team.

Over a decade ago, walking into my first day at Tesla’s Fremont factory, I had little idea of its place in automotive history or of the chapters still to be written. I was only vaguely aware of how Telsa purchased the sprawling facility for the knockdown price of $42 million from Toyota complete with operating press lines, alone justifying the cost.

Even today, Fremont remains the second-largest manufacturing plant on the West Coast, surpassed only by Boeing’s immense assembly complex in Everett, Washington.

Birth & Decline of Bay Area Automotive Industry

Yet, Toyota wasn’t the factory’s original occupant. Built by General Motors in the 1950s, Fremont was part of a broader wave of automotive manufacturing that swept into the East Bay, with Ford opening its own plant just a few miles south.

However, by the early 1980s, both plants were failing. Productivity had plummeted, quality control was non-existent and absenteeism rampant. The only metric trending up and to the right was drug use amongst its workforce.

From here, the story of the two factories diverges. Today, the only remaining trace of Ford’s San Jose assembly plant is an easily overlooked historical display inside the Great Mall of the Bay Area, the shopping complex that now stands on the plant’s former site.

Former Ford San Jose Assembly Plant Display at Great Mall of the Bay Area

The Fremont factory, meanwhile, was also shuttered as it faced up to the rising competition from Japanese automakers which had set new standards for quality and reliability.

Toyota and Honda vehicles held a magnifying glass up to GM’s offerings, exposing their deficiencies, nowhere more so than the vehicles that came from Fremont’s production lines.

As the American ‘Big Three’ were being asked why their cars broke down, political pressure also mounted for Japanese manufacturers, with the threat of import taxes and growing ‘Buy American’ sentiment within government and sections of the public.

Toyota needed to find a solution, but a big question remained unanswered. Could their lean manufacturing philosophy, termed the Toyota Production System, or TPS, the source of their competitive advantage, work in America?

A New Dawn: Lean Manufacturing

One of the core methods of TPS was the art of running small experiments in the pursuit of continuous improvement, and this logic was applied with the objective of answering their big question.

GM on the other hand, needed to find a solution for Fremont and its wider quality and productivity issues, and so in 1984 the New United Motor Manufacturing, Inc (NUMMI) was born. An equal partnership to see if TPS could be applied outside of Japan with American workers.

By the end of the year the GM owned brand Chevrolet Nova rolled off the production line, and by 1986 the Corolla followed, the first Toyota produced in America.

In the subsequent decades, Fremont rose to the top of GM’s rankings for quality and efficiency, and Toyota had their answer and built factories of their own across the United States.

By the 1990s and early 2000s, TPS was being embraced by the West. The publication of “The Machine that Changed the World” bought widespread attention to its core principles, Just-In-Time, Kaizen and 5S. NUMMI was its prime case study of how these principles could work not just in Japan, but in the U.S. and Europe as well.

At university, these principles were held up as best practice. We learned how Toyota only designed and manufactured what it must in-house, relying instead on a close-knit, highly efficient supplier network for everything else, from fasteners to seating systems.

One standout memory was being told that some Japanese factories still used paper filing systems in certain areas because, incredibly, they were faster and more reliable than digital systems for specific tasks.

By the time Ford had taken over Land Rover in 2000, where I was an apprentice, companies were fully leaning into TPS principles. Overnight, we were painting walkways, implementing Statistical Process Control (SPC) and unimaginatively calling it the Land Rover Production System (LRPS).

But it became clear that elements of its adoption were a struggle. For example, 5S, the process to create an organized and efficient workplace, through five structured stages. We found that, we were forever sorting and cleaning, but we never really got to the deeper layers of standardization that made the system sustainable.

Cracks in Lean Manufacturing

Fast forward to my first days at Tesla. In the manufacturing team, there were many ex-NUMMI, versed in TPS and lean thinking. But there were elements that surprised me. One of them was the way data was handled.

In the heart of Silicon Valley, people were still taking measurements manually, writing them down on paper, and filing them away. Coming from a world where German automakers had long since digitized their quality systems, this felt strangely analog.

This reliance on manual processes was seen elsewhere. In the midst of a sea of robots, there were curled-up pieces of paper outlining work instructions that were never paid attention to.

These created islands of data, formatted to permit only the most basic level of analysis, owned by those ‘in the know’. Whilst many new recruits and those drafted into support the launch, were flying blind.

A knowledge debt began to appear in the company, leading to duplicated efforts, misaligned changes, and ultimately waste, a cardinal sin in the lean manufacturing world.

Tesla was still learning how to build cars. It also needed to sell them, fast, to enthusiastic early adopters who were willing to forgive poor door gaps for the thrill of a three-second 0–60mph and the 17-inch center touchscreen.

And so, Tesla leaned into Digital Tools and Agile Development methods, that Silicon Valley was famed for. Control systems, traceability, analytics and in house enterprise systems were favored over manual controls like kanban cards.

But that digital ambition clashed with the residual practices of TPS and sometimes, with the people who promoted it. One oft-repeated story in Tesla folklore marked the symbolic end of TPS at Fremont.

A Quality VP, trying to channel classic Toyota principles, stopped the line to investigate an issue, only to be told by Elon that he was no longer a Tesla employee in front of the production staff.

Did Lean Manufacturing Fail at Tesla?

Well, yes and no. Toyota’s core lean principles, eliminating waste, empowering workers, continuous improvement are embodied in Silicon Valley principles of ‘go fast and break things’ encouraging fast iteration by all, regardless of the chain of command.

However, digitalization has challenged some of the ways lean manufacturing has been implemented in the past. Over-reliance on manual processes and local experts creates silos of knowledge, which in today’s data-rich, complex systems does not enable enterprise-wide efficiencies.

I remember, years later, trying to publish a dashboard, only to be asked by a former Japanese OEM engineer if I was planning to share it with the whole company. They were genuinely surprised, it was their belief that their group alone should see the data, because “they were the experts.”

Tesla’s rise and that of other EV startups challenged other areas of TPS beyond the shopfloor. Breaking the supplier-led, modular production approach, was a significant one. Instead of relying on a tightly coordinated supply base, Tesla pursued vertical integration designing and building its own seats, electronic systems and more.

It did this, because just like the factory floor, the vehicle tech stack could no longer be considered as individual islands managed locally by a team of experts. For an electric software-defined vehicle you need tight integration across multiple domains.

Trying to revolutionize the ways cars are designed, built and operated through a network of established suppliers, as Tesla found, is difficult.

As more suppliers found working with Tesla’s methods challenging and its innovations as risky or threating. Tesla bought more and more design and manufacturing in-house.

Agile, Challenging Western Automotive Practices

As well as TPS, the western stage-gate model to manage new product introduction came under threat. Using this model, each stage of product development is rigorously defined and reviewed at a formal stage-gate review.

A cornerstone of this model was the concept of an all-encompassing design freeze, a milestone where the company declared the design complete, releasing drawings so that tooling and production equipment could be fabricated.

But when you’re building something fundamentally new, placing so much confidence in a single design freeze can create problems. It delays learning, discourages iteration, and locks in assumptions too early.

Tesla didn’t have the luxury of doing that. It had to keep learning while launching. Therefore, the rigidity of stage-gate was set aside in favor of agile software principles, organizing product improvements around rapid sprints and iterative releases.

Ironically, this move echoed one of the core tenets of TPS, delaying design freeze until enough was learned. In this way, Tesla’s path wasn’t a rejection of lean principles but a reinvention of them, updated for a digitally integrated, fast-moving product landscape.

Agile, Reshoring & the Manufacturing Reset

I am currently a few chapters into “Your Life is Manufactured” by Prof. Tim Minshall and I agree with him that there are real advantages in being able to make the things you invented. In my experience, designing how a product is manufactured and producing it, is part of the learning process.

This loop, again like the TPS principle of “Plan-Do-Check-Act”, creates a virtuous cycle where micro-innovations are incorporated back into the design improving both manufacturability and customer satisfaction.

However, if we are to reshore, as this article shows, the world of manufacturing has moved on, and China has taken agile processes to new levels and can move from design to manufacture at a breathtaking pace.

Regaining the ability to make things, once it is lost, is not only extremely difficult, as Tim Minshall points out. But others take what was there previously and runs with it, to the point it is almost unrecognizable to what we remembered and catching up with the knowledge deficit is doubly difficult.

Over 5,000 prototypes, that’s how many iterations James Dyson went through in five years before creating a bagless vacuum cleaner ready for market.

Overwhelming confidence and a supportive wife helped him reach that point. Yet, he still faced resistance from incumbents refusing to license his technology for fear of disrupting the recurring sales of vacuum-bags. Meanwhile, other companies tried to copy his design.

In the end, Dyson brought manufacturing in-house and took his product to market with £650,000 in bank financing, equivalent to roughly $3 million today. By today’s standards, that might be comparable to a Series A round for a consumer-facing company.

By contrast, firms like Nothing (earbuds and cell phone technology) and Humane (AI pin) have both raised over $200 million in debt and equity to fund innovation and growth.

Has the start-up market changed?

Are vast funding rounds the way? Or do dogged designers have a chance bootstapping their hardware products to market? Eventually achieving billion-dollar valuations while retaining full ownership?

Intrigued, I looked into UK unicorns to learn about their founders, business models, and funding. An article by Beauhurst served as my starting point, allowing me to investigate the founders, funding history, and market segments of 50 unicorns to compare them with Dyson.

First, more about James Dyson

James was in his thirties when he started to miniaturize the cyclone technology he witnessed in a factory environment years earlier for use in household vacuums, and it wasn’t until he was over 40 that Dyson Ltd was formed.

Before focusing on vacuums, he spent his twenties at art college and later designed and sold unique sea vehicles at Rotork, before launching his first venture.

That first venture reimagined the wheelbarrow by replacing the wheel with a ball. Although the “barrowball” gained popularity, Dyson was ultimately forced out by his investors, and the company failed.

You might assume such a background would appeal to venture capitalists who promote a fail-fast mindset. But in the UK at the time, corporate experience was still paramount, and his bitter experience led him to strike out on his own.

Today’s Funding Landscape

The UK now boasts a thriving venture capital ecosystem, and vast sums are raised with 20% of our unicorns receiving over £100 million in funding prior to achieving unicorn status.

There is evidently an appetite for investment, but does this money go to founder with a bright idea and a bucket full of determination?

Well, let’s see.

Looking at the age profile of unicorn founders in Figure 1, it appears many were younger than Dyson, suggesting youth is not a deal-breaker for investors. Yet a closer look at founders under 25 at companies like BrewDog, GymShark or Paddle, reveals they were already successful entities before raising venture capital.

These companies had demonstrated product-market fit and shown resilience in scaling. Venture funding was primarily used to accelerate growth and go global.

Figure 1: Age profile of unicorn founders in the UK

So, would a 30-year-old James Dyson, armed with an idea, a clear customer pain point, and a sizable addressable market but no revenue, still struggle to secure venture funding today? Probably, unless he also had robust experience.

Revisiting Figure 1, the largest cohort of unicorn founders are in their thirties. Examples include Flo Health, Monzo, Revolut, and Wise—each founded by people who built domain expertise in their twenties, often at start-ups or in finance.

The founders of Motorway previously worked at Zoopla, those at Darktrace came from Autonomy after its acquisition, and Zego’s founders had been at Deliveroo.

Therefore, for founders in their twenties, the product must already prove itself. For those in their thirties, domain expertise and a solid track record are critical.

This justifies James’ decision to go his own way, at least initially, to prove product-market fit and potential to scale. After this point, investors today would be more willing and aggressive in their funding.

The UK’s Appetite for Hardware

But is the UK prepared to fund hardware on the scale it demands? The answer, illustrated in Figure 2, is largely no.

FinTech, SaaS, and the emerging AI sector dominate among today’s UK unicorns, with minimal presence in robotics, quantum computing, or other physical product domains. Graphcore is an exception in semiconductors, but even it is primarily software-focused.

We literally have Nothing, as well as a few MedTech and Biotech firms like CMR Surgical and Oxford Nanopore, which benefited from academic research funding and patent-based validation.

Figure 2: UK unicorns by business sector

Still, the investment culture leans heavily toward software, likely due to London’s central role, where 75% of UK investment capital is deployed. Although AI funding is a bright light, which has demonstrated a track record with successes like Deepmind.

However, if the UK wants to develop companies that can employ thousands of people, which Dyson does, we need the risk appetite to fund more Nothing’s and go into sectors such as robotics, or quantum computing.

But we currently have no track record and these fields require a ton of capital and risk, which our investment culture is not ready for. Part of the success of AI, despite the cost of compute, is that it fits into our software-centric model that can be proven with lower investments then scaled.

Conclusion

If you’re a modern-day James Dyson, whether 20 or 50, and aiming to launch a hardware business, you still need that unwavering determination to develop the product, market the brand, and establish product-market fit. Only then might venture dollars come your way.

However, there is potential for change.

If Nothing or other hardware unicorns continue to succeed, the UK may become more willing to back similarly challenging endeavors. Just as Castore rode the momentum from Gymshark in sports apparel, we may yet see a new Dyson emerge in the wake of Nothing’s success.

It’s interesting how words take on different meanings depending on context. Take 'agglomeration', in my world it refers to a defect in batteries, where electrode material clumps together and inhibits performance, a negative connotation.

Recently, I came across a more positive usage: agglomeration as the concentration of businesses and workers in connected geographic areas, driving innovation, productivity, job creation, and higher wages.

Does Agglomeration Exist?

Well yes, the San Francisco Bay area is the best example, where multiple small cities down and around the bay seamlessly merge into a network of largely affluent cities, from San Francisco to San Jose.

Top class universities in the form of Stanford and Berkeley attract the best and the brightest, some of whom gravitate to the start-up mecca of Palo Alto and with Venture Capital funding create firms, and in time they outgrow the city and set-up in neighboring areas such as Menlo Park (Meta), Mountain View (Google) or Fremont (Tesla) attracting ever more highly skilled and ambitious people to the area in a virtuous cycle.

How Did it Start?

The story is well known. After World War II, the U.S. government invested heavily in defense research and high-tech industries. NASA established a base in Mountain View, and companies like Lockheed Martin expanded their aerospace projects, pushing advances in radar and electronics. This, in turn, ushered in the semiconductor revolution of the 1950s, led by startups like Intel.

Inevitably, people began to figure out how silicon chips could be used by people in the workplace and at home, with firms such as Hewlett Packard and later Apple being part of the computer revolution.

The level of innovation drove investment into the area with the establishment of a thriving ecosystem of Venture Capital firms setting up shop on Sand Hill Road, next to the Stanford Campus which continued to provide a steady stream of ideas to fund.

As ideas became companies, they spread out across the bay, connected through I5, I280 and 101 freeways, and so, the computing era morphed into the internet revolution, and on into social media, smart phones, mobile apps, software defined vehicles and artificial intelligence.

This success was fueled by an influx of talent and investment—first from the government, later from the private sector. A well-connected mass transit system enabled businesses to expand into new locations like Redwood City and San Mateo. These areas, in turn, became hotbeds of innovation, particularly in South San Francisco, which emerged as a hub for high-tech biomedical firms.

Does Agglomeration Exist in the UK?

The UK, particularly England, has one of the highest population densities in Europe. In theory, we should be the agglomeration capital of Europe with great companies and a skilled workforce.

Intrigued, I decided to research and dug into the trove of data on the Office of National Statistics website, and with a python script and help from ChatGPT was able to produce the following maps, which show the UKs Gross Regional Product (GRP) per capita with an increasing level of fidelity:

Fig. 1: GRP per capita for UK by (a) region, (b) county and (c) district.

GRP per capita is the average value per head in a particular geographic region of the UK, you can think of it as their average salary. Zooming out on the left, the dark green spot representing Oxfordshire & Berkshire could be one of those areas.

Shifting one map to the right, we see Berkshire is the richer district, with poor old Oxfordshire, despite its University and F1 teams dragging the larger region down. However, moving right again, we see the city of Oxford is rich with the surrounding green belt being less affluent.

Indeed, while the south of England is evidently richer, it is difficult to spot a network of cyan high-income hotspots, which would show an agglomeration effect.

There is potential, the Cambridge-Oxford-Milton Keynes corridor has been identified as an emerging agglomeration hub. If infrastructure investments, such as the long-discussed Oxford-Cambridge rail link are realized, this region could develop into the UK's answer to Silicon Valley.

Which areas are networked together?

Our initial investigation shows isolated urban hubs of wealth, but no spill over into the surrounding counties, which can then morph together with other spill over areas, to create a silicon shire, or alike.

To see if there are any signs of this network effect, I looked at growth, rather than the absolute GRP figures. The maps below show the top 25% of areas by population densities on the left, and the GRP per capita growth over the past 25 years on the right:

Fig. 2: Districts of UK by (a) population density and (b) economic growth.

Ah! There is strand on cyan counties running west of London all the way to Bristol, along the M4 motorway. Indeed, north of London there also another promising sign going north up the M1, through Milton Keynes and on to the Coventry area.

What is so special about the M4 corridor?

It appears our motorways are a positive tick in the box for agglomeration. But what about the other factors, such as urban areas, universities, government and private investment, and a continual renewal of ideas.

Stretching across the M4 from London to Bristol and then up north along the M5, there are the urban areas of Reading, Swindon, Bath, Gloucester and Cheltenham, well connected but typically 20-30 miles apart, where in the Bay Area cities are smaller but much more bunched together.

On to Universities, of course London has top class institutions but on the west, there is both University of Bristol and University of Bath which are also highly ranked. On either side then, there should be a flourishing of ideas and top talent ready to spread out across the region either creating their own businesses, or ready to work in world class firms and institutions.

Indeed, as you move from west to east you have some great companies, Airbus, Rolls Royce Aerospace, GKN and Dyson which are co-located close to government agencies of Ministry of Defence and GCHQ, which are customers of some of these private firms, employing highly skilled individuals themselves.

Going further west the story continues, the story of tech and government agencies continues, with BAE Systems, Vodafone, Thermo Fisher and Eli Lilly, along with the Atomic Weapons Establishment. But there is also a significant spill-over from London with wealthy areas having great transport links into city and high value services supporting these families.

In summary, the M4 corridor is a story of transport links, universities, advanced technologies and government agencies operating in high-tech fields (aerospace, biotech, consumer electronics, defense, pharma, telecommunications) creating a positive economic eco-system around them.

What does this agglomeration effect tell us for the rest of the UK?

Britain has become an increasingly service-based economy over the past 25 years, but high value services concentrate in cities, such as Edinburgh, London or Manchester. They do not serve the wider regions.

For agglomeration, we need to invest in companies that need space, in other words advanced tech and manufacturing, with urban hubs in between which can be vibrant, such as Bristol, to attract the younger skilled individual, or more settled such as Wokingham, which encourage the more established to stay.

The importance of renewal

Reviewing the companies set up along the M4 corridor, what is noticeable is the lack of newer larger companies. If we want agglomeration in sectors such as new energy, we need our larger companies to evolve and be joined by emerging giants.

If you look at the largest companies in the UK vs USA, the UK is made from established firms, while USA has a significant number of companies under 25 years old, which we need to create.

This renewal factor is crucially important. The M4 corridor has adapted and moved into fields such as pharma or consumer electronics. This differs the Newcastle-Sunderland-Middlesborough agglomeration, where industries such as ship building and coal mining have been decimated, resulting in anemic growth over the past 25 years.

This could have been different, if the ship-building sector continued to adapt and invest in new manufacturing technologies or engineering tools, it could still be competitive with the ship building giant of South Korea. Indeed, with government foresight, if coal mining shifted focus to new energy, there could have been new engineering and manufacturing firms establishing themselves in this field.

What is the lesson?

If we don’t have innovation, investment and renewal in advanced technologies that have room to scale, the UK will increasingly spend its tax income supporting a decaying system. It is like not replacing a broken window and spending more and more of your income on heating.

So, while the M4 corridor can be seen as a success of the last 25 years, it must continue to adapt and be joined by new hungry agglomeration regions with the freedom to expand along a growing network of transport links, with urban hubs that can attract the brightest minds.

If we get comfortable, we will and are getting quickly overtaken by other more hungry countries, and any tech innovation coming from our universities and top minds will have no avenue to grow once they reach a certain point, unless it is overseas. And then, in economic terms, agglomeration may transform its meaning to have a negative connotation.

Moving back to the UK, I knew that job opportunities in manufacturing would be scarce, in comparison to where I was previously. I was fortunate that my ten years in Silicon Valley, the region experienced a renaissance in hardware manufacturing. Spurred on by Tesla, behemoths like Meta and Google began creating physical products, while start-ups such as Lucid initially manufactured batteries before venturing into designing and building a luxury EV.

The post-dotcom bubble within Silicon Valley was built on Web 2.0, search, social, mobile, and then evolved into the ‘internet of things,’ enveloping the physical realm. During this transition, software companies, accustomed to the idea of bugs, weren’t surprised to find that they also arise in manufacturing. However, their approach to solving these bugs was different to the traditional incumbents.

Avoiding the Bug

In the traditional heartlands of automotive manufacturing bugs are meant to be avoided. Over time, these regions learned what will and will not work, developing processes to critique designs and ‘desensitize’ them so they could be produced using tried-and-true methods. As a result, both design-led and manufacturing-led innovation slowed. This was recognized by outsiders as early as the 2000s, when Bill Gates commented:

“If GM had kept up with the technology like the computer industry has, we would all be driving $25 cars that got 1,000 miles to the gallon.”

In addition, a supply base emerged that designed and manufactured more of the advanced components for vehicle producers. This led to a loss of expertise within OEMs and to increased standardization, ensuring that different systems could integrate during final assembly. Consequently, cars were developed in parts, rather than as a cohesive whole.

Embracing the Bug

In contrast, Silicon Valley took the stance that sometimes it is better to fix the problem than to avoid it. To build revolutionary, high-quality products, it needed to learn not only how to manufacture as effectively as the incumbents but also how to push the boundaries of what was thought possible. This required retaining manufacturing knowledge in-house, at all levels of the production process, from minerals to motor.

Even companies that outsource to contract manufacturers recognize the difference between the labor that physically produces the product, and the expertise required to design the manufacturing system itself. Hence, if you visit a contract manufacturer, you will often find engineers from major tech giants working side-by-side with the contract manufacturer and designers back at their home base. Through iteration, they come to understand the product’s functional specifications and figure out how to produce ten, then a thousand, and ultimately hundreds of thousands of units.

Manufacturing-led innovation goes hand-in-hand with design-led innovation, be it through revolutionizing the production process or making incremental improvements that significantly boost productivity and quality. This mindset permeates the entire supply chain, driving the integration of new technologies and novel manufacturing methods so that the system as a whole is both revolutionized and optimized.

A Painful Yet Necessary Process

Returning to the UK, there is still a core manufacturing base, but it has been significantly depleted. Many products are designed here but manufactured elsewhere, resulting in limited understanding of the intricacies of production processes. There are positive exceptions. Formula One, for instance, where innovation is integral to staying ahead. Here, design and manufacturing remain closely bonded. Components designs can be iterated and built quickly, while constant data and failure analysis means manufacturing processes can be improved in parallel.

Realizing a physical design is difficult. Production rarely goes seamlessly at first, which can be likened to ‘death by a thousand cuts.’ Over many years, we have outsourced this pain to the point where we now impose requirements on our vendors that, without our knowledge, may be nearly impossible to achieve. Nevertheless, our vendors have persevered and now we envy their manufacturing capabilities.

It doesn’t stop there, because our competitors excel at manufacturing, they can translate their designs and disruptive ideas into reality more quickly, with less bureaucracy and faster feedback loops. This raises the question, what is now the route for our own design and R&D innovations?

Get on Board

The UK has excellent early-stage R&D: our universities are strong, and we have many small companies that display real entrepreneurial flair. But when it comes to scaling, we rely on others’ capabilities the journey is a long and slow one, which we get out compete. Worse, for innovations further down the supply chain, for example a new process to refine a particular material, there is a very limited ecosystem for this innovation to flow into.

To scale, we must not shy away from bugs. Instead, we should embrace the lessons they bring, seizing the opportunity to create new designs rather than crafting processes that weed out risk. In short, we must endure more pain and take pride in having made it to the other side of that struggle.

I often think about the process of designing a car. As an engineer, I’m comfortable diagnosing problems and implementing fixes, but I’m not equipped for the initial, more artistic phases of vehicle design.

However, perfectly formed lines aren’t what design is all about. Concepts are linked to personas, market requirements, and technology trends. It must fit within a certain envelope, achieve a target drag coefficient, and have a specific wheelbase.

The vehicle design exists within certain constraints, living at the intersection of engineering feasibility and creative flair.

Using these constraints, I could use my drafting skills to lay out the wheels, box off the area for the trunk, and mark out the overall length and height of the vehicle. However, as soon as I draw one artistically inspired line, my technical drawing would look like it had been graffitied by a two-year-old.

I simply haven’t done the reps of drawing vehicle concepts over and again, learning about perspective, shading, etc., to be able to put my imagination on paper.

Yet authors often use vivid, precise language to describe an object’s texture, contrasting colors, and how a character reacts to these sensory details—bringing everything to life in the reader’s mind. So, I wondered, why not feed a detailed prompt into ChatGPT to see how it interprets my vision?

I thought about what I desire for my next car and came up with the following prompt:

“Design an understated electric sedan for a 45-year-old professional male who values minimalist design, with some distinctive features, such as narrow angular headlights or sharp feature lines on the door. It should be able to corner well with an aggressive rake angle and sit low to the ground and have low-profile 22-inch alloy wheels. The ideal body color would be white, and the overall design should communicate refinement and performance. Produce an image that is photorealistic, studio lighting, ultra-detailed, high resolution.”

Now did ChatGPT come up with a mock-Tesla because that was what I was describing or because out of all the electric sedans the model was trained on, the majority were Teslas? There is an element of creativity missing.

So, could an AI be as creative as Frank Gehry playing with cardboard to develop his iconic building concepts? Not on its own, and not without an ongoing interaction with the designer, through sketches, sharing ideas and iterations.

But it is a useful tool, one which could rapidly increase the speed of product development, holding requirements in mind as the design morphs into the end product.

I remember once working with a scooter company being amazed at how quickly they iterated and launched new vehicles, more akin to the development timelines of a cell phone than a car. Designers rapidly responding to changing tastes, altering the shape and color of the scooter and having it on the market within six months.

In a traditional automotive company, product development would begin with concepts, and some of those concepts would be chosen to be made into scaled clays and a further subset into full-size versions.

It is not until the clay has been reviewed, observed under different lighting, lines being subtly altered in response to wind tunnel results, and had initial feasibility studies done for crash and manufacturability that it is frozen, and engineering can be in a period of twelve months or longer.

This is a huge barrier to entry for a start-up.

But a designer working one-to-one with a client can drastically reduce that barrier. Conversing to understand their needs and wants whilst simultaneously interacting with the virtual design, feeding in additional prompts, requirement, or directly altering the feature line of a door.

At a broader scale, the same principles could apply to an agency building a brand in conjunction with the product, gathering customer data and then A/B testing the generated designs. It opens new ways in which a product can be developed, with faster feedback loops and lower cost.

Then what’s next? There is still a huge amount of capital investment in engineering the design, validating it, manufacturing it.

But maybe not?

There is no reason your design can’t be generated into a three-dimensional model and broken down into parts, components selected and built. Leaving you to focus on adding value to the customer, building the brand and exciting the market.

How design tools develop in conjunction with emerging AI trends will blur the line between the sequential process of design and manufacturing, leaving many of us wondering how we can respond to these changes?