Field Notes from the Future

Agentic commerce: Why value and trust may become the new scorecard of business

Deloitte OCTO — Wed, 02 Sep 2026 14:03:17 GMT

Key insights

Agentic commerce is emerging quickly: AI agents are increasingly helping consumers research, compare, and purchase products online.
“More value for the price” is becoming a critical competitive signal: Consumers are prioritizing brands that deliver the strongest value exchange, not simply the lowest price.
Predictive customer insights may outperform traditional loyalty metrics: Signals that track shifts in customer perception relative to competitors can forecast growth months ahead.
Trust is becoming a foundational requirement for AI adoption: Reliability, transparency, capability, and humanity all influence whether people trust AI-driven decisions.
Businesses may need to organize around outcomes: Companies that focus on delivering measurable customer outcomes, not just selling products, may be best positioned for the era of AI-mediated commerce.
What is agentic commerce and why is it important?
First web revolutionized commerce, then mobile did the same. The next transformation may be driven by AI agents. These tools are increasingly becoming part of the buying process, helping people research options, compare prices, evaluate product reviews, and sometimes even complete purchases.
This emerging model, often referred to as agentic commerce, could reshape how brands compete online. Instead of consumers manually searching and evaluating products, AI systems increasingly act as intermediaries by analyzing data and recommending the best option based on the customer’s goals. In this environment, the brands that win may not be the loudest or most visible. Instead, they may be the ones that consistently demonstrate value and trust.
Why agentic commerce is on the rise
Economic uncertainty and shifting expectations are pushing consumers to rethink how they spend. But the shift is about more than finding the lowest price. Consumers are increasingly evaluating value. Premium brands can still deliver strong value if the experience, quality, or outcomes justify the price. Conversely, even discounted products may struggle if they fail to meet expectations.
These dynamics are already changing consumer behavior. For example, higher-income shoppers who historically avoided discount retailers are increasingly visiting value-oriented stores when they perceive a stronger value equation.1 The implication for companies is significant: customer loyalty is becoming more fluid and consumers are increasingly willing to switch brands when they find better value.
The ways in which companies used to connect with customers and develop relationships with repeat buyers are becoming less relevant in the current data-driven world.
Why agentic commerce demands new metrics
For decades, companies relied on customer satisfaction scores, loyalty indicators, or net promoter score to gauge brand health. But these indicators often measure past experiences rather than future momentum. What matters more may be how customer perception is shifting relative to competitors.
When organizations analyze large-scale customer feedback over time, they can often identify moments when perceptions begin to diverge. Such signals may predict changes in revenue growth or market performance months later. In this sense, predictive customer insight can act as a forward-looking navigation system, helping leaders see what may be coming rather than simply measuring what already happened.
The rise of AI agents amplifies the importance of value signals. AI systems evaluating products do not respond to marketing messages the same way humans do. Instead, they rely heavily on data signals that indicate which product delivers the best outcome. This could fundamentally reshape marketing and brand competition. If AI agents prioritize measurable customer outcomes and verified experience data, brands that consistently deliver real value may rise faster, while those relying primarily on advertising or brand perception could struggle.
Trust becomes the foundation of AI-driven commerce
For agentic commerce to scale, however, trust will become even more important than it currently is. Consumers and enterprises are still cautious about AI systems making decisions on their behalf. Trust depends on several elements working together, including reliability, transparency, capability, humanity, and purpose. Brands that can demonstrate that they’ve centered their products and services around these ideas are likely to perform better with consumers.
Another emerging factor is the quality of the data behind AI systems. As AI models increasingly generate synthetic data, human-verified insights may become critical for grounding decisions and improving trust. Without trustworthy data, even powerful AI systems may struggle to gain widespread adoption.
Agentic commerce shifts organizations toward outcome-driven goals
Taken together, these trends suggest companies may need to rethink how they operate. Historically, many organizations structured themselves around products, sales channels, or geographic markets. But as AI mediates more decisions, those structures may be less relevant. The companies that succeed may be those that organize around a simpler question of whether they created meaningful outcomes for their customers. Because when AI agents evaluate options, compare experiences, and recommend products, they will likely prioritize brands that consistently deliver strong outcomes. That could redefine the scorecard of business.
This post is an adaption of Deloitte’s new FWD: Technology podcast with special guest Rob Pace, Founder and CEO of Hundred X. Hosted by Bill Briggs (Chief Technology Officer), Brett Davis (Chief Innovation Officer), and Simona Spelman (Chief Human Officer), the show brings clarity to a world of complexity, helping you move from buzzwords to business impact. Each episode features stories, insights, and hard-earned lessons from builders who know that innovation only works when people know what matters…and why.
This article was originally published on April 15, 2026.
Click here to hear the full interview and catch up on previous episodes.
Help spread the word — forward this to your network.
Share
Join our community to get emerging tech insights every week.
Subscribe now
^{This article contains general information only and Deloitte is not, by means of this article, rendering accounting, business, financial, investment, legal, tax, or other professional advice or services. This article is not a substitute for such professional advice or services, nor should it be used as a basis for any decision or action that may affect your business. Before making any decision or taking any action that may affect your business, you should consult a qualified professional advisor. Deloitte shall not be responsible for any loss sustained by any person who relies on this article.}
^{As used in this document, “Deloitte” means Deloitte Consulting LLP, a subsidiary of Deloitte LLP. Please see www.deloitte.com/us/about for a detailed description of our legal structure. Certain services may not be available to attest clients under the rules and regulations of public accounting.}
^{Copyright © 2026 Deloitte Development LLC. All rights reserved.}

Rob Pace, interview with Deloitte, March 5, 2026.

Real-World Wins: How back-office AI is delivering results

Deloitte OCTO — Wed, 26 Aug 2026 15:37:15 GMT

A couple of months ago, we wrote about the promise of back-office AI: greater speed, sharper insights, and improved engagement. But what does success look like when AI moves from theory into practice?

Let’s explore how leading organizations are unlocking real value by integrating AI into their core business operations—the transactional processes, reconciliations, and administrative duties that keep enterprises running smoothly.

Subscribe now

Real-world use cases

Across industries, organizations are deploying AI to tackle longstanding back-office challenges, like late payment detection and supply chain vulnerabilities, with measurable results. The following examples show how companies and public institutions are using AI to reduce costs, improve forecasting accuracy, and build operational resilience.

Smarter collections in Telecom
A US-based telecom company used machine learning to detect late payments. By equipping its finance department with intelligent collection strategies, it achieved potential annual savings of US$20 million through more targeted invoice management.1
Proactive maintenance in utilities
A North American utilities provider adopted AI-powered predictive maintenance, identifying 70% of component failures up to five days in advance.2 This enabled proactive planning, reduced downtime, and optimized resource allocation.
Precision forecasting in beverage manufacturing
An alcoholic beverage company improved forecasting accuracy by 95%, leveraging machine learning to integrate diverse data, from market indices to weather patterns, for better-informed supply chain decisions.3
Resilient drug supply chains in public services
A US government entity enhanced its drug supply chain monitoring with AI, automating workflows and proactively identifying vulnerabilities. This proved critical for managing shortages and building resilience after the COVID-19 pandemic.
Efficiency gains in higher education
A public university automated refunds, deposits, and fee follow-ups using AI, saving approximately 3,800 full-time equivalent hours annually and improving the experience for more than 45,000 students.4
Transformation in healthcare
A major US healthcare system achieved up to 70% faster employee onboarding, a 10–20% improvement in collections for self-pay accounts, and a 12–15% reduction in back-office expenses, all supported by advanced automation and AI-driven data extraction5

Strategies that support success

Success with back-office AI requires more than just technology. Organizations that see outsized results apply a business-first approach built on the pillars below:

AI readiness: Assess maturity across data availability, volume and diversity, quality and integrity, governance, and ethics.
Use case scoping: Align AI solutions with strategic challenges for measurable impact.
Data infrastructure evaluation: Build systems that support scalable and secure AI.
Data quality and integrity: Ensure insights are trustworthy and reliable.
Data governance and ethics: Establish policies to guide responsible AI use.

This framework provides a pathway for sustainable transformation—helping organizations move from experimentation to meaningful impact.

Lessons for all

Back-office AI is not limited to industry giants. The technology is increasingly accessible, and the starting point is straightforward:

· Select a process

· Test a use case

· Measure outcomes

· Scale success

Organizations that listen to their teams and foster a culture of experimentation can achieve greater results. AI is a tool to free human talent, so people can focus on creativity, problem-solving, and relationship building.

Beginning your AI journey

The future of back-office AI is about preparation more than adopting cutting-edge technology. Whether you are just starting small or scaling your efforts, success begins with the right foundation.

Key capabilities to access

Data Availability: Do you have access to the right data?
Volume and Diversity: Is your dataset comprehensive and varied enough for robust AI models?
Quality and Integrity: Can you trust your data for reliable AI-driven outcomes?
Governance and Ethics: Are policies in place to ensure responsible use?

Essential strategies

Conduct an AI maturity assessment: Identify strengths and gaps.
Prioritize use cases: Focus on challenges where AI delivers real business value.
Strengthen data foundations: Invest in infrastructure, quality control, and governance.
Invest in skills and culture: Build teams ready to learn, innovate, and adapt.

Leave a comment

The bottom line

Implementing back-office AI isn’t about chasing trends—it’s about building for long-term, scalable impact. Organizations that thoughtfully assess, prepare, and act are setting the standard for tomorrow.

Ready to begin? Start by evaluating your organization’s readiness, scoping high-impact use cases, and building your data and talent capabilities. The next era of business is agentic, autonomous, and insight-driven, so ensure your organization is prepared to excel.

Let’s build the future together!

https://www.deloitte.com/content/dam/assets-zone3/us/en/docs/services/consulting/2025/back-office-ai-pov.pdf

— Nabhanil Mondal | Associate Vice President | AI and Data

This article was originally published on February 11, 2026.

^{This article contains general information only and Deloitte is not, by means of this article, rendering accounting, business, financial, investment, legal, tax, or other professional advice or services. This article is not a substitute for such professional advice or services, nor should it be used as a basis for any decision or action that may affect your business. Before making any decision or taking any action that may affect your business, you should consult a qualified professional advisor. Deloitte shall not be responsible for any loss sustained by any person who relies on this article.}

^{As used in this document, “Deloitte” means Deloitte Consulting LLP, a subsidiary of Deloitte LLP. Please see www.deloitte.com/us/about for a detailed description of our legal structure. Certain services may not be available to attest clients under the rules and regulations of public accounting.}

The new AI edge is discernment, not enthusiasm

Wed, 19 Aug 2026 14:01:12 GMT

Somewhere in your company right now, someone is deciding whether an AI model should answer a question that used to have a simple yes-or-no answer: can this person see payroll data?

It sounds like an odd place to draw a line. AI can summarize a contract, draft a first-pass email, or catch a pattern in support tickets that no one ever wrote a rule for. Surely it can handle “who’s allowed to see what.” But access control isn’t a judgment call. It’s already defined, by an org chart, an HR system, and a policy someone wrote down years ago. The correct answer exists and is knowable with certainty. Handing that question to a probabilistic model doesn’t make the answer better. It replaces a fact with a guess: an 85% confidence “probably not” standing in for a rule that was already fully specified.

That’s the trap a lot of companies are walking into. For the past couple of years, the working assumption inside most enterprises has been that more AI is automatically better. Bolt it onto the workflow, add it to next year’s roadmap, count the use cases in the board deck. Today, there’s increasing focus within organizations to make sure workers are mindful of the cost of the tokens they’re using. The days of tokenmaxxing are definitively dead, with most enterprises realizing the emphasis on widespread AI use bought them nothing more than sky-high bills. Now the focus is shifting to using AI for workloads where it actually moves the needle. There are two major benefits to this approach. One is, obviously, lower costs. But the other, which may prove even more meaningful, is about using the right tool for each job.

The enthusiasm for all things AI is increasingly meeting the hard math of ROI. S&P Global Market Intelligence’s research on enterprise AI adoption found generative AI has outpaced the AI approaches that came before it: 60% of companies investing in AI have implemented generative AI, ahead of longer-standing approaches like rules-based systems (54%) and pattern recognition (51%). But the same research found the rush has a cost. The share of companies abandoning most of their AI projects before production climbed from 17% to 42% in a single year, with the average organization scrapping nearly half its proofs of concept. Even so, close to half of companies say no single business objective has seen a real, measurable lift from generative AI. Adoption went up, but so did the pile of work that never shipped.1

None of this is an argument against AI. It’s an argument against skipping a step. Most enterprise leaders have gotten good at asking “how do we add AI here?” and out of practice asking the question that should come first: does this problem actually need it?

A lot of what gets labeled an AI problem is really a process problem, a data-quality problem, an integration problem, or a plain old automation problem wearing an AI costume. When the right answer to a task is fixed, governed by policy, regulation, or a clearly written rule, a deterministic system gets there faster, cheaper, and more consistently than a model approximating an answer from patterns in its training data. A deterministic system also doesn’t arrive with the extra weight AI carries by default, such as the licensing and token costs at scale, the need to watch for drift and hallucination in a way no one ever had to watch a fixed rule, or the validation and documentation that model-risk and compliance teams now expect before anything touches production. That overhead buys nothing when the underlying task never had any ambiguity to resolve in the first place.

Think of it as putting a fit gate in front of every candidate use case. Teams should consider an explicit, upfront checkpoint, cleared before a line of code gets written, that routes the problem to one of three places. Some problems genuinely belong with AI, such as synthesizing messy, unstructured information, or catching fraud patterns that don’t reduce to a rulebook. Some belong with conventional automation, like anything with a defined right answer, a policy behind it, or a compliance obligation attached, like payroll access, tax withholding, or invoice matching. And some don’t belong with either tool. In these cases, it’s better to redesign the process, or retire it, rather than automate a bad workflow with a more expensive kind of automation.

That third option is the one most companies skip, and it’s often the one with the best return.

Going into 2027, plenty of companies will still be counting AI use cases and calling it a strategy. The ones that pull ahead will be the ones that built the discipline to treat “no, not here” as a legitimate, well-reasoned outcome of a real evaluation, not a failure of ambition. Enthusiasm was the right instinct for getting started. Discernment is what it takes to actually win.

The payroll question, and likely a number of other workflows, still deserves a boolean answer.

Enjoying this post? Subscribe now so you never miss the next one.

Subscribe now

Have you seen similar trends in your industry? Share your thoughts in the comments below.

Leave a comment

Help spread the word — forward this to your network.

--Caroline Brown | Group Creative Director | Deloitte Consulting LLP.

S&P Global Market Intelligence, "Generative AI Shows Rapid Growth But Yields Mixed Results," https://www.spglobal.com/market-intelligence/en/news-insights/research/2025/10/generative-ai-shows-rapid-growth-but-yields-mixed-results

Context is king: Optimizing infrastructure for inference

Wed, 12 Aug 2026 14:04:03 GMT

Here’s a funny thing that happened in AI compute over the last few months: global supplies of dynamic random-access memory (DRAM) hit an unexpected bottleneck, and prices spiked. For decades memory chips were relatively cheap commodity components. The current shortage caught almost everyone off guard.

That’s especially notable given where all the attention in AI hardware has been. Over the last couple years, GPUs were where the action was, as hyperscalers and frontier AI labs grabbed as many as they could get their hands on in order to train bigger and bigger AI models. GPUs were the bottleneck everyone worried about. So it’s telling that memory chips are the ones in short supply now. It may be the first hint that we’re entering a new era in how people and enterprises use AI, one that places a premium on AI tools’ ability to handle context during inference.

The previous era was all about scale. Bigger models, trained on more data, and more parameters, delivered better performance. That will likely continue, but the current memory shortage is a signal that inference, or how an AI tool actually performs once it’s out in the world, is becoming just as important. It’s not just about a model’s size anymore. It’s about that model’s ability to handle context.

Users who expect larger context windows (a chat tool’s working memory of the conversation), agentic systems that hit multiple models over an extended time period, and governance policies that require audit trails all force AI tools to hold extreme amounts of data in memory. In the past, users might have gone to an AI chat tool for a quick answer to a simple question. Now they’re increasingly using them for deeper business problems. They’re feeding AI tools with larger amounts of business data and asking the tools to produce more complex outputs after many more turns in a conversation.

I’d file all this under “context.” It may be one of the single biggest factors in how useful an AI tool is in an enterprise setting. Experts increasingly say that a model’s contextual awareness will be one of the most important factors in an AI tool’s ability to work within business processes and deliver return on investment.1

Tech companies have taken notice, and many more are looking at how to help AI tools do more with less, rather than simply cramming data centers with more DRAM. Google, for example, recently announced a memory compression technique that it says will speed up inference workloads significantly without requiring additional hardware.

We’re also starting to see software companies compete directly on context, though it’s a different flavor of context than the kind driving the memory crunch. Where a model’s context window is short-term working memory for a single conversation, this next wave of tools is about giving AI agents a longer memory of the business itself. AWS, for one, just released a suite of tools built around what it calls a context layer, a knowledge graph that maps how a company’s data connects and gets smarter over time based on how its AI agents actually use it. Snowflake and several other vendors have released similar tools.2 It’s a sign that as enterprises move the use of agents move from pilot projects into real workflows, whether the agents handle context well, in the moment and over time, will decide whether they’re actually useful.

The memory shortage may have been hard to see coming, but the growing importance of context isn’t. It’s the difference between a tool that gives a quick, shallow answer to a simple question and one that can actually operate on its own inside your business. That’s the difference I’d be watching for as you decide where your next AI dollar goes. Context is king.

Join our community to get emerging tech insights every week. 

Subscribe now

Have you seen similar trends in your industry? Share your thoughts in the comments below. 

Leave a comment

If this post resonated, forward it to colleagues who would also find the insights helpful!  

--Kelly Raskovich | Senior Manager | Deloitte Consulting LLP.

https://siliconangle.com/2026/06/18/context-becoming-missing-layer-enterprise-ai-thecuberesearch/

https://venturebeat.com/data/aws-enters-the-context-layer-race-with-a-graph-that-learns-from-agents-not-manual-curation

Open-weight AI models are all about what you want to manage vs. what you don’t

Wed, 05 Aug 2026 14:03:46 GMT

Earlier this year, nearly one third of the AI workloads that American companies ran through one of the largest AI developer marketplaces were handled by open-weight models. Two months earlier, that share was about 11%.1 Those numbers describe something that was once a niche, hobbyist corner of the AI world but now has grown into a legitimate choice for enterprises.

“Open-weight” sounds like a technical distinction, and it is one, but it has quietly become a strategic one too. A closed model lives behind an API that someone else controls. They can raise the price, tighten the rules, or cut off access, and there is not much a customer can do about it. An open-weight model is a file. Once a company has it, that company can run it on its own servers, fine-tune it on its own data, inspect what it actually does, and keep using it exactly as is even if whoever trained it disappears tomorrow.

That difference used to come with an obvious cost. Open models were meaningfully behind the best proprietary ones in performance. That gap has closed fast. Comparisons that once separated open and closed models by a full technology generation are now often described as a matter of months.2 On some tasks, coding especially, one recent comparison put a leading open model within a percentage point of a top proprietary one on a closely watched benchmark.3 Open-weight models accounted for less than four percent of total AI spend.4

Here is the actual trade-off. Once a model’s weights are released, they cannot be recalled. Anyone with enough computing power can retrain away whatever safety behavior the original developer built in, and there is no equivalent of a vendor quietly patching a vulnerability after the fact. That is a real, structural risk a closed model does not carry in the same way.

Open-weight models may also be vulnerable to a specific kind of attack. A cybersecurity researcher recently demonstrated that it’s possible to manipulate these models’ behavior by feeding them curated data sets that cause the models to take specific actions that the original user never intended.5 In last year’s Tech Trends report, we wrote how this kind of data poisoning attack may start becoming more common. Unsecured open-weight models risk elevating this kind of attack.

For most companies, this is landing less as a philosophical choice and more as a portfolio one. There are likely enough AI workloads to go around, and using proprietary AI models for certain jobs and open-weight models for others may make sense for enterprises going forward. Proprietary tools bring security and stability while open-weight models enable customization and experimentation. The type of workload will determine which model is the right tool for the job. The choice is all about what you want to manage. For teams that are still in the early days of their AI journeys, offloading some of the governance and technical support to a vendor may make sense. For more mature AI teams that have more in-house tech talent who can build and manage implementations, open-weight models may help them drive new capabilities.

None of that makes the choice simpler. It does mean the choice is now a real, ongoing operating decision rather than a one-time technology pick, and it deserves the same continuous scrutiny a business would give any other piece of critical infrastructure.

Join our community to get emerging tech insights every week.

Subscribe now

Have you seen similar trends in your industry? Share your thoughts in the comments below.

Leave a comment

If this post resonated, forward it to colleagues who would also find the insights helpful!

--Ed Burns | Production Manager | Deloitte Consulting LLP.

Zentera, “The Weights Are Coming Home: Why Enterprises Are Moving AI On Premises, and Even On Device”: https://www.zentera.net/blog/why-enterprises-are-adopting-open-weight-ai

WinBuzzer, “DeepSeek V4 Ships 1M Context, Open-Weights”: https://winbuzzer.com/2026/04/27/deepseek-v4-open-weights-launch-xcxwbn/

MLQ News, “Chinese AI Models Surpass 30% of US Developer Traffic on OpenRouter as Cost Gap Over OpenAI Widens”: https://mlq.ai/news/chinese-ai-models-surpass-30-of-us-developer-traffic-on-openrouter-as-cost-gap-over-openai-widens/

Zentera

The Register, “Researcher poisons open-weight AI model for under $100”: https://www.theregister.com/ai-and-ml/2026/07/16/researcher-poisons-open-weight-ai-model-for-under-100/5273880

When measuring the wrong thing is worse than not measuring at all

Wed, 29 Jul 2026 14:00:30 GMT

In a turn of events that surprises everyone I meet, I have a degree in optimization. Two, actually, and arguably three, if an MBA counts. That, combined with the facts that I live in San Francisco and am a management consultant, might lead you to believe that I am someone who tends to optimize my life—planning my days to the minute, tracking my fitness, packing my weekends, ordering ahead. On the contrary: In a job interview a few years ago, I was asked to what degree I make decisions based on intuition. I replied 100% (shockingly, I got the offer).

I tell you this because I am about to make a controversial statement. In an era where the suffix du jour is “maxxing” and every week a new company pops up with a new bodily function to measure—I’ve personally been in conversations about skincare trackers and dog wearables within the past six months—I propose we have over-indexed on metrics.

Let me preface this argument by acknowledging that I believe in the importance of measurement, particularly for the large, public organizations I serve, who must be responsible financial stewards, transparently reward their people, and be able to trace decisions and impacts from the micro to the macro and back again. It is absolutely critical to define success measurements for any program or product so that it can be continuously improved upon and its existence can be justified. If I were compensated on vibes alone, I would be frustrated if I found out my peers were being rewarded differently and that difference could not be explained. And if my doctor didn’t have a way to monitor my physical condition, my hypochondria would probably spiral out of control.

But the problem with over-indexing on metrics is that we optimize for what we measure. Certain things are much easier to measure than others, which means we tend to over-fit our behavior toward those things. I made a joke earlier about “maxxing,” but just in the past couple months “tokenmaxxing” was a perfect demonstration of this argument. Because AI adoption is very difficult to measure, many organizations chose to measure token usage as a proxy for adoption. The problem is that this incentivizes in many ways the exact opposite behavior than was desired. If an employee is incentivized to use as many tokens per day as possible, they will probably use the most expensive models, use AI to do things that they should really do without it, and perhaps even use enterprise accounts to do personal errands that have nothing to do with work. This often results in outrageous costs, inefficiencies and decreased productivity, and learned behaviors that may be difficult to undo.

Frequency of usage is a marginally better measure—a leader should want to see their employees leveraging AI frequently, especially early in the AI journey, because it’s important for people to experiment and build fluency—but what if that frequent use isn’t effective use? Even productivity, if leaders are able to align on its definition (which is a big if), can be challenging to measure if you’re not also gauging the quality of outputs. Over-indexing on individual efficiency may cause leaders under pressure to show efficiency gains to miss the forest for the trees. The focus on these metrics that are relatively easy to measure means that the wrong behaviors might be hardwired into the organization. A few of my favorite leaders at Deloitte recently released a report entitled “From AI adoption to adaptation,” which finds that enabling real value from AI requires true behavior change, such as strengthening judgment, comfort with experimentation, and divergent thinking. When organizations measure usage or activity alone, those behaviors are less likely to be learned, and meaningful business outcomes are less likely to be realized.

I believe the pitfalls of overoptimizing for misaligned metrics become abundantly clear as organizations try to incentivize AI adoption and justify tremendous investment, but this is not only an AI problem. In the words of Derek Thompson, host of Plain English, “I think [metrics] force us to play the games we can measure rather than the games we value.” For example, I think the best way to measure a sports team’s likely longevity would be to look at the passion of the fan base, but passion is incredibly difficult to measure. So instead we measure who is spending the most money. These fans may or may not be passionate, but we can measure their investment, so we might over-rotate on serving them. This could leave fans who have more passion but less financial means in the dust, eroding long-term trust and excitement in the franchise, and ultimately leading to its demise.

I first became passionate about this topic when I picked up Matthew B. Crawford’s pop-philosophy book Shop Class as Soulcraft. In it, he argues that work was stripped of meaning when labor was divided into “thinking” and “doing;” in other words, management and strategy vs. building and executing. This more or less coincided with the rise of Taylorism during the Industrial Revolution, in which complex systems and craftsmanship were modularized into discrete, sequential activities. This required that outputs of component activities be measured rather than the outcomes derived from the overall process. Realizing how that has translated into our modern economy has governed my approach to work to this day. I am relentlessly focused on driving true value and outcomes, and I don’t particularly care how those outcomes were achieved (as long, of course, as they were morally sound). If my team uses AI to create incredible deliverables quickly, that’s fantastic. If they didn’t use AI but created incredible deliverables quickly, that’s also fantastic. What is not fantastic is if they don’t use the tools at their disposal to create the best outcomes possible within the time that we have allocated. I trust them to use their judgment to determine what those tools are.

So, if over-indexing on activity metrics isn’t the answer, what is? I am, unfortunately, still trying to figure that out—as is the business community as a whole—but I propose the following “Metrics Principles” to help govern decisions around KPIs in the interim.

1. Work backwards from value. At the end of the day, the best metric is, “Did the thing I want to measure create business value? Did it increase the bottom line?” Much like a personal OKR, program or product KPIs should be designed to tie back explicitly to value creation. If the measure is more likely to incentivize behaviors that hinder value creation, it is probably not a great metric to measure.

2. Less is more. In my experience, driving toward a few defensible, value-oriented metrics has proven more effective than proposing dozens of KPIs for a given program or product. Ultimately, you are trying to figure out if the thing you are doing is creating value, so the combination of a few proxies should be enough to provide a thumbs up or thumbs down.

3. Think long term. This is, of course, easier said than done. But if a metric is incentivizing short-term outcomes or behaviors, it’s probably not really driving value. Number of clicks in a week, probably not so helpful; number of clicks in two years? That might drive some real insights.

4. Just because it’s hard to measure, don’t skip it. My biggest frustration with a metrics-obsessed culture is that often the most important things are the hardest to measure, so we don’t measure them. Emotional labor, collaboration, mentorship, commitment are hard to measure quantitatively, but qualitative measurement is better than no measurement.

My last word of wisdom might be the scariest phrase ever uttered by a consultant, but here it is: At a certain point, we must trust our intuition. Perhaps not 100%--even I, despite what I told that recruiter years ago, do in fact rely on data to make decisions—but particularly as we progress in our careers and our lives, our intuition is based in real experience. Gut feeling may not be THE metric to follow, but it is a good one. And it’s likely better than a metric that tracks the wrong things.

Enjoying this post? Subscribe now so you never miss the next one.

Subscribe now

What would you add to this list? Let’s build it together in the comments.

Leave a comment

Help spread the word — forward this to your network.

--Dany Rifkin | Manager | Deloitte Consulting LLP.

Quantum's awkward middle years are finally getting interesting

Wed, 22 Jul 2026 14:03:06 GMT

For most of the past decade, quantum computing has lived in a strange limbo. The physics was real, the demos were impressive, and—yet—practical use cases remained limited. That gap between promise and practice made quantum easy to file under “someday.”

After attending the Economist’s quantum conference recently, I came away with a different read. The conversation is shifting. People are no longer arguing about whether quantum computers will do useful work. They’re arguing about which useful work comes first, what it will cost, and how to be ready for it. That is a much healthier debate, and it’s the clearest sign yet that the field is maturing.

A date on the calendar

The moment that crystallized this shift came during the conference. Amazon Web Services and QuEra Computing announced Libra, a fault-tolerant quantum computer that will be available through Amazon’s Braket service in 2028.1 The hardware alone will be a significant achievement. But this is also notable because a hyperscaler and a service vendor are putting their reputations on the line with a specific year for a machine designed to support commercially useful work.

The specs give a sense of what “useful” means here. Libra is designed for more than 256 error-corrected logical qubits and roughly a million reliable logical operations.2 This contrasts with today’s devices that support handfuls of logical qubits and hundreds of logical operations. Without logical operations, quantum computers accumulate small errors that cause large calculations to dissolve into noise before they complete. These logical operations are based on engineering tricks that let a machine catch and correct its own mistakes mid-calculation, which is the difference between a lab curiosity and a tool you can build a workflow around.

QuEra’s own framing is telling. The company describes fault-tolerant quantum computing as moving from a scientific milestone to an engineering and deployment roadmap.3 That’s the whole story in a sentence. The hard question now moves from whether vendors can deliver useful quantum computers to how they’ll deliver them on schedule and at a price that makes sense.

If Libra arrives roughly on time, it has a strong claim to being the first genuinely useful quantum computer. Which means the interesting work now falls to the clever technologists who have to figure out how to do valuable things inside its real constraints of cost, capacity, and access.

The clock that’s already ticking

There’s a less exciting but more urgent implication buried in all of this, and it belongs to your security team.

The first useful quantum machines won’t be able to crack the encryption that protects today’s internet. But the direction of travel is clear enough that regulators have stopped waiting. A 2026 executive order in the US requires federal agencies to transition to post-quantum cryptographic algorithms by 2030 and federal contractors to make the switch by 2031. One of the biggest effects of the order is that it shifts quantum readiness from voluntary preparation to mandated deadlines.4 The UK’s National Cyber Security Centre has set the same 2035 target for completing migration.5

Those dates feel comfortably distant until you do the arithmetic. As we wrote about recently in our Tech Trends report, large organizations often need three to five years just to inventory where their cryptography lives and to migrate it, and complex estates can take longer. There’s also the harvest-now-decrypt-later problem, where adversaries can quietly collect encrypted data today and simply wait for the hardware to catch up. This not only raises the criticality of good cyber hygiene (e.g., one can’t decrypt data they don’t have), but drives the need to migrate to post-quantum cryptography which runs on ordinary computers. The sensible time to start scoping that migration was a while ago.

Value while you wait

Even as quantum technology matures, you don’t have to wait for 2028 to get something out of it.

Quantum-inspired techniques, which run on classical hardware but borrow the mathematical structure of quantum algorithms, are delivering value now and can help you build the internal muscle you’ll want later. Two areas stand out for enterprises. The first is sharpening machine learning outcomes in places like fraud detection. The second is reworking the heavy Monte Carlo simulations that sit underneath risk calculations such as value at risk and derivative pricing, which are exactly the workloads researchers expect quantum methods to accelerate.6 These techniques not only provide ROI and business value today, but teams can look under the hood and start understanding how quantum computers operate. Teams that start implementing these techniques for value and reskilling today are the ones who will move fastest when the hardware lands.

Optimization is another target use for quantum computers. The special purpose devices called quantum annealers support real optimization problems.7 Here, the trend is people learning to translate more complex and messy business challenges into optimization problems, from scheduling and logistics to working out crystalline structures in chemistry. That skill of problem reformulation may be the thing that separates the organizations that benefit from quantum from the ones that keep waiting.

The takeaway

While the fully realized potential of quantum computing may remain on the horizon, enterprises should be focusing on a few simple things. The path to value has a shape now. There’s a date on the hardware, a deadline on the security work, and useful things to do in the meantime. Quantum is no longer a someday technology you can safely ignore. It’s a roadmap, and roadmaps reward the people who start reading them early.

Join our community to get emerging tech insights every week.

Subscribe now

Have you seen similar trends in your industry? Share your thoughts in the comments below.

Leave a comment

If this post resonated, forward it to colleagues who would also find the insights helpful!

-- Scott Buchholz | Managing Director | Deloitte Consulting LLP.

AWS Quantum Technologies Blog, “AWS Deepens Strategic Collaboration with QuEra to Bring Fault-Tolerant Quantum Computing to Amazon Braket,” June 15, 2026.

QuEra Computing, "QuEra Announces 2028 Fault-Tolerant Quantum Computer and Expanded Multi-Year Strategic Collaboration with AWS," June 15, 2026.

PR Newswire, “QuEra Announces 2028 Fault-Tolerant Quantum Computer and Expanded Multi-Year Strategic Collaboration with AWS,” June 15, 2026.

White House, “Securing the Nation Against Advanced Cryptographic Attacks,” June 22, 2026.

UK National Cyber Security Centre, "Timelines for migration to post-quantum cryptography," March 20, 2025.

Quantum (journal), "Quantum Monte Carlo simulations for financial risk analytics: scenario generation for equity, rate, and credit risk factors," April 4, 2024.

D-Wave Quantum Inc., “D-Wave Announces Advancements in Annealing and Gate-Model Quantum Computing Technologies,” Jan 27, 2026.

How technology is reshaping music production

Wed, 15 Jul 2026 14:00:38 GMT

A few weeks ago a song came on while I was in the kitchen and it made me stop what I was doing. It was an old track, recognizable, but the production had been rebuilt from scratch. It had a recognizable melody but completely different atmosphere. I picked up my phone to check who’d done the cover. The track credited no producer or remixer. It just said “AI-generated.” I genuinely had no idea. Not for a single second.

That got me thinking about how far the use of technology in music production has come in the past few years. I grew up watching the behind-the-scenes videos of large-scale film score recordings, the kind where a full orchestra filled a studio to record a single song. I have a very specific picture in my head of what making music used to look like. It looked nothing like one person working alone with a laptop.

What a Recording Session Used to Be

Those old film recordings were a major logistical operation. It required a proper studio packed with musicians, strings on one side, brass on another, percussionists, a choir, a lead vocalist, all scheduled in layer by layer, session by session. There was a literal queue of artists waiting for their slot to contribute maybe thirty seconds of a song. The composer spent weeks, sometimes months, arranging each part. Someone’s whole job was just managing who came in and when. And after all of that, you still had to mix it, edit it, master it, and record it to tape.

A three-minute song could represent six months of work, dozens of people, and a budget that would make your eyes water.

The Revolution That Started Before AI

Here is what people forget: AI is not the first tech tool that turned music production upside down. Digital audio workstations already let one person do what previously required a full band, an engineer, and a studio booking. Drum machines replaced session drummers. Synthesizers replaced string sections. Autotune gave singers a safety net that made recording approachable to people who might have frozen up in front of a live microphone. Then sampling happened, the internet happened, and suddenly a teenager in their bedroom anywhere in the world could release music that sounded competitive with professional productions.

Every single shift came with someone saying the old way was being killed. Every single time, music became something more people could make. The line of musicians waiting for their recording session slot became a folder of electronic samples.

What AI Actually Adds

AI has compressed another gap, the part that still demanded years of technical skill. You no longer need deep music theory knowledge to compose something that sounds composed. You do not need a vocalist if you can describe the feel of a performance. You can write lyrics, generate a composition, record and tune a vocal, layer harmonies, and produce the whole thing alone. You can take an existing song and reimagine it in a different style, with a different artist’s voice, in a different genre entirely. You can even generate versions of songs that were never released, alternate takes, what-ifs, things that exist nowhere else.

For someone who grew up watching that orchestra queue in action, that is genuinely hard to hold in my head.

None of this means the use of AI tools in music production is without challenges. The same tools that free a solo producer also make it easy to clone a voice without consent or build on someone else’s creative work without crediting or paying them. The industry is mid-argument over how these issues should be handled and the sides are nowhere near settled. But the question of how something was made has always been separate from whether it deserves to exist. What is not optional is making sure the people whose work contributed to the training of these models get a fair seat at the table.

I went back and replayed that track after dinner, this time hunting for the seams. I found a couple. Small things. But I only noticed because I was looking. If I had just let it play, it would have been part of the evening like any other song. The feeling was real regardless.

Music has always absorbed whatever technology was available and kept going, stranger and wider and more accessible each time. AI is the newest instrument in that lineage. I do not know exactly how musical artists will use AI in their productions. But that world was already changing long before a machine learned to make music. What AI has done is make these changes undeniable.

Enjoying this post? Subscribe now so you never miss the next one.

Subscribe now

What would you add to this list? Let’s build it together in the comments.

Leave a comment

Help spread the word — forward this to your network.

--Makarand Kukade | Production Senior | Deloitte

Robots are ready. But where do you start?

Wed, 01 Jul 2026 14:03:48 GMT

North America’s largest robotics show made one thing clear: the machines work. The harder question is what to do with them.

Walking the floor at McCormick Place during Automate 2026 felt like a trade show transformed into a whirring factory that someone had wheeled into downtown Chicago. Robotic arms welded behind protective curtains. Bipedal humanoids drew crowds of spectators with smartphones to see if they would dance or drop. Autonomous mobile robots threaded within booths, carrying payloads from vehicle doors to inventory pallets to pharmacy carts.

The energy was infectious, as was the money behind it. In the weeks around the show, multiple large manufacturers announced multi-million-dollar funding deals.1 The capital is frothing because the underlying technology has crossed a threshold. As Deloitte’s Tech Trends 2026 report puts it, the transition from prototype to production has been happening for some time now, with robots already inspecting power grids, assisting in surgery, and working alongside people in warehouses.

But spend a day talking to the people building these machines, and a more sober picture comes into focus. The robots are capable. The industry is still working out how to deploy them responsibly, which form factors actually earn their keep, and what the real deployments of value are hiding.

Safety stopped being a sidenote

The most striking shift from previous years was tonal. In the sessions of the third annual Humanoid Robot Forum, which now overlaps directly with the flagship Automate show, safety dominated the conversation. This is different than often crowding out the capability demos that used to steal the spotlight. That is a healthy sign for an industry that wants to put human-like machines next to human bodies.

Underneath the product announcements, the slower and arguably more important work is happening in standards committees. Industrial safety standards move on multi-year cycles, and the people writing them face a genuine tension for setting rigorous design parameters for machines that learn and adapt without deterring innovation. This is not an abstract worry. Tech Trends 2026 notes that as robots leave controlled factory floors for public spaces, regulators will need new frameworks for certification, liability, and oversight, and that companies already face overlapping and sometimes contradictory requirements across jurisdictions.2

You could see the caution expressed in hardware model iterations, too. Select humanoid makers that had prioritized two legs have started adding wheeled options, trading some of the human-like mobility for the simple fact that a wheeled base is far less likely to topple over near a person. It was a small detail that captured the mood of the whole event. Stability, predictability, and trust are becoming the baseline requirements as deployed units scale from single digits to multitudes.

The debates are everywhere. The answers are not.

If safety was the dominant theme, consensus was lost on form factor capability. Stationary or mobile? Hands or grippers? Fingers or suction cups? Legs or wheels? Every one of these questions has passionate advocates, and almost none of them had a settled answer.

That ambiguity is not a failure. It is what an industry looks like when driven by innovation. But it creates a real problem for buyers, because the pace of new model releases is fast enough that any hardware bet you make today may look dated in a year. The more useful framing is to focus less on form and more on function. What task do you actually need done, and what is the simplest machine that can do it reliably?

That principle cuts against some of the hype. A general-purpose robot that promises to do a little of everything is an impressive concept but a difficult business case. A machine engineered around one valuable, repeatable task is far easier to justify.

Finding the focus

So where does the value actually live today? On the Automate floor, the convincing demonstrations were narrow and specific: ship welding, folding small boxes, pushing carts, manipulating containers. These are not glamorous tasks. They are exactly work that companies struggle to staff with humans, and that robots can be trained to perform at high volume. But automating tasks alone does not equate to displacing human influence on a one-for-one scale. Instead, human roles will need to be redesigned around workflows that interchange robot-capable tasks with machines.

The shipbuilding example is worth dwelling on, because it shows how a focused strategy plays out in practice. Shipyard welding is high-skill, hazardous, physically punishing work, and the trade is hard to staff, with attrition in some key roles running high.3 A robot that does nothing but weld in a constantly dangerous space is not a compromise. It is a precise match of capability to human risk and automation need.

There is a geographic dimension here as well. Shipbuilding is concentrated in Korea, Japan, and China, and the appetite for automation in that part of the world is enormous. South Korea has the highest robot density of any country on earth, with more than 1,200 industrial robots per 10,000 manufacturing employees, and Singapore ranks second.4 When skilled labor is scarce and the work is grueling, the deployment of automation becomes abundant. However, the understanding of what human labor does next is not as clear of a case.

This is the throughline running through the whole event. To eventually get robots handling a broader range of difficult human work, the hardware will keep maturing as does the vision for what people will do in exchange. The path to more interchangeable human and machine work does not run through a single do-everything robot arriving fully formed. It runs through thousands of narrow, well-defined tasks that can be trained, simulated, and delivered at scale today, each one teaching the machine models – and people – a little more.

The robots really are ready, but the models cannot be the only ones that should be learning. The smarter question, for any leader watching this space, is not whether to adopt them but where to start. And the answer, increasingly, is wherever the work is hard, dangerous, and specific enough that a machine can do it well many times over.

Join our community to get emerging tech insights every week.

Subscribe now

Have you seen similar trends in your industry? Share your thoughts in the comments below.

Leave a comment

If this post resonated, forward it to colleagues who would also find the insights helpful!

--Matt Shannon | Product Manager | Deloitte Consulting LLP.

Hyundai to acquire SoftBank’s remaining stake in Boston Dynamics for $325M

Tech Trends 2026 | Deloitte Insights

Hyundai-backed humanoid robots to transform welding in shipyards

Robot Density Surges in Europe, Asia, and Americas - International Federation of Robotics

Tech Trends 2026 | Deloitte Insights

From Code to Conversation: How We Stopped Building and Started Talking

Wed, 24 Jun 2026 14:02:40 GMT

Cast your mind back a few years. A developer building a feature would open an integrated development environment (IDE), write hundreds of lines of code, run it, watch it break, refactor it, optimize it, test edge cases, repeat. It was craft. It was slow. It required deep expertise. The barrier to creating anything was the ability to write it — character by character, line by line. Then something shifted.

Low-code and no-code platforms arrived and said “you don’t need to write it; you just need to configure it.” Suddenly a business analyst could build software features that once needed a developer. The drag-and-drop canvas became the new IDE. Creation moved from syntax to structure.

But that was only the first gear change. We are now in the middle of a second change driven by AI, and it is far more radical. Today, you don’t configure. You converse.

The Three Eras of Creation

The first era of software development was defined by writing code, which was the only mode for expressing intent digitally. The gap between idea and output was technical skill. If you couldn’t write it, you couldn’t build it. This gave way to the second era of development, in which users could express their intent through visual interfaces that allowed them to drag, drop, connect, and configure components without needing extensive coding ability.

The gap between idea and output shrank dramatically. Non-technical users became builders. We’re now entering the third era. Coding tools leveraging AI are allowing users to express their intent through natural language, whether typed or spoken. The gap between idea and output is now almost nothing. You describe what you want, and an AI tool builds it.

It Is Not Just About Code

This is the part many people miss. The conversation about AI and automation tends to focus on software development. But the shift is close to universal and is happening to many acts of creation.

Designing an email campaign? You used to open a design tool, drag blocks, choose fonts, pick colors, write copy, A/B test subject lines. Then came template builders — pick a layout, fill in the blanks. Now you tell an AI tool “create a re-engagement email for customers who haven’t purchased in 90 days, warm tone, include a 10% discount, our brand colors are teal and white.” Done.

Creating a presentation? You used to build it slide by slide. Then came templates. Now tools like built in tools in presentation software generate a full deck from a one-line brief that includes structure, content, visuals and all.

Analyzing data? This once required SQL, Python, or a BI developer. Then came drag-and-drop analytics. Now you ask in plain English “Show me which customer segments had the highest churn last quarter and why,” and an AI tool queries, analyses, and narrates the answer.

Even designing a physical product is changing. Generative design tools let engineers describe constraints like weight, material, and load and AI creates structural forms no human would have conceived.

Look across all of these examples and you see an identical arc. Work has moved from expertise-dependent to template-driven to intention-driven. The tool used to require you to speak its language. Now it speaks yours. This is not incremental improvement. It is a fundamental inversion of the relationship between human intent and machine output.

What This Means for Organizations

Enterprises are handing off more decisions to agentic AI. The organizations that will lead are not necessarily those with the largest technology budgets but rather those that recognize this shift earliest and redesign their operating models around it.

The question is no longer “Do we have developers who can build this?” It is “Do we have people who can describe what they need clearly enough for an agent to build it?”

That is a different skill. And it sits with a different person in the organization.

The New Scarce Resource

In Era 1, the scarce resource was technical skill. In Era 2, it was process knowledge. In Era 3, the scarce resource is clarity of intent — the ability to articulate what you want, for whom, under what conditions, with what constraints. In short, this is the ability to think well and communicate precisely.

Prompt engineering is often dismissed as a temporary workaround until AI gets smarter. That misses the point. The underlying skill — structured thinking expressed as clear language — will only become more valuable as agents become more capable. Better agents do more with a good brief. They also amplify the cost of a poor one.

A Word of Caution

None of this means expertise is obsolete. The best outcomes still come from people who understand the domain deeply enough to know whether the agent got it right. A developer who can review AI-generated code. A marketer who can judge whether the generated campaign will actually convert. A lawyer who can spot what the AI contract draft missed.

The human role is shifting from maker to director, and directors who don’t understand the craft tend to make poor films.

The Bigger Picture

We are living through a compression of the gap between imagination and execution. What once took a team, a quarter, and a significant budget can now take a conversation and an afternoon.

That is not a threat to human creativity. It is potentially one of the greatest amplifiers of it...

The question for every leader, every team, and every organization is the same: are you still optimizing Era 1 skills when the real leverage has moved to Era 3?

Join our community to get emerging tech insights every week.

Subscribe now

Have you seen similar trends in your industry? Share your thoughts in the comments below.

Leave a comment

If this post resonated, forward it to colleagues who would also find the insights helpful!

-- Srivani Yadati | Manager | Deloitte USI.

Navigating hospitals with some help from physical AI

Wed, 17 Jun 2026 14:03:22 GMT

Hospitals are large, complex environments where sensitive materials need to be moved quickly and securely. It’s the ultimate proving ground for physical. To get a look at how robots are navigating these challenges, Deloitte managing director Franz Gilbert sat down with Vivian Chu, co-founder and chief innovation officer at Diligent Robotics. What follows is as lightly edited transcript of their conversation. It covers how Diligent’s robot, Moxi, works in hospitals, the kinds of use cases it works in, the design considerations that went into building it, and the role that physical AI plays in its capabilities.

Franz Gilbert: My name is Franz Gilbert. I’m a managing director at Deloitte, and I have the sincere pleasure of being one of the leaders in our humanoid robotics practice. And I’m so excited, Vivian, that you’re here. And I don’t even want to try to do the introduction because what you’re doing is amazing.

If you wouldn’t mind for the audience, share a little bit so they kind of understand why I’m so excited you’re here.

Vivian Chu: So my name’s Vivian. I’m the co-founder and chief innovation officer at Diligent Robotics and I think what’s so exciting and why people want to hear about our story and journey is that at Diligent we’ve put robots in hospitals at scale and truly are making a difference.

So at Diligent, Moxi is our flagship robot that goes around doing deliveries like medication, lightweight equipment, labs, anything that would take staff away from their patients. And then by having Moxi make these deliveries, they spend more time with those patients.

What makes Moxi unique is that Moxi has an arm so the robot can actually interact with the world, ride elevators, and do things that no other robots can do right now.

FG: And if you wouldn’t mind, can you share the elevator story just to kind of paint a picture of how well you’ve thought about this?

VC: So right now, you need to have the ability to ride the elevator to do any deliveries. If you’re stuck to a single floor, you’re not going to be helpful to the staff. So what Moxi does is that Moxi will actually perceive the environment, look at where the buttons are and use its arm to push. Like, I want to go to floor 2 and the robot will push the button and then get on and ride the elevator with people.

And it’s so interesting because not only can Moxi do that, but we also have the AI underneath to detect people because sometimes we actually interact and ask people for help.

So one of the, one of the really cool things about having our robots deployed, it’s not just a novelty anymore. We understand the dynamics of how people get on the elevator just like you and I do. And we’ve discovered that when Moxi goes into an elevator and asks for help, people do that. And they love doing that for the robot.

FG: That’s the reason I’ve been so excited about it is that you really are in a real-world environment. You’re not in a work cell. You know, it’s more truly how we do work. So how does Moxi learn from the hospital staff and kind of what’s your vision for training them, you know, as it goes forward?

VC: So today we think about it as white glove service. The hospital staff, we talk to them, we understand their workflows, we understand their pain points. And then we make recommendations actually of how the robot should pick up the task or do a delivery that makes the most sense.

Our vision on the future really is to continue that white glove service where people should think about what they want the robot to do so they can focus on their work, and then we figure out the how. Right now we do a lot of work under the hood around templates and scripting so that it’s very straightforward and we can dynamically set up the environment very quickly and easily so the robots can do the task for the people.

FG: As you think about this, I mean you’re in a hospital environment, so how do you think about cyber and trust? What are some of the things that you’ve done in that regard?

VC: That’s a great question. That’s definitely a big aspect of being in the hospital environment and something we take very seriously. So some of the first things we do are the basic things like [compliance and privacy standard System and Organization Controls, or] SOC 2 compliance, like really thinking about how do we make sure that our base systems, all our company processes are all locked down.

And then we kind of go into layers. All the things that we picked on the robot, for example, the arm, we chose an arm that had a form factor that was designed to go on a wheelchair, which is actually the original use case of this arm. So it’s just naturally safe because they thought about pinch points. The arm torques are very low. We’re only pushing buttons right now. So you don’t need a lot of force. So naturally that’s already safe.

We layer things on like hardware safety and we have our sensors. We do things like safety bubbles. And then we build out all that to finally have safety behaviors of how it acts when it sees a person or wheelchair. What is that behavior it should do?

And so we think about it in all these layers to really be able to have a robot be in this mission critical environment where safety is incredibly important.

FG: Give some sense of scale to the audience. Talk about giving the time back, because you hear a lot about job replacement? But this is a different approach.

VC: Right. So this is actually really important. When we went to hospitals, we wanted to think about where do people want to spend their time? Where should they spend the time and where should robots fit into this picture? And so there’s actually research that shows that for every minute and extra touch point that a nurse spends with the patient, the patient outcomes improve. What we think about is how do you get them in the rooms more? How do you get them off of just running around the hospital?

And so for us, it’s around augmenting the staff as opposed to saying we’re going to go do nursing work. No, the nurses should go spend more time with their patients to actually make that big improvement.

FG: So aside from the human, let’s get a little bit into the technical side. Talk about some of the underlying technology components of the physical AI and what you think some of the biggest challenges are right now either for y’all or for the industry as a whole.

VC: So physical AI has many stages of growth at this at this juncture in time. Early in the robotics time span, it was a lot about how do you get all the little modules to work together. Now we are starting to approach more end-to-end systems where you can actually say here’s an example and use case that we couldn’t handle before, put that in your data flywheel and now the robot can handle these corner cases.

It’s a really interesting time and we’re tackling this problem in the same way that a lot of the cutting edge approaches are. But the caveat is that because we’ve been in the space so long, we’ve thought about sort of the safety modules, the guardrails, because one of the biggest challenges is actually operational. What happens when the robot does fail? What happens when you hit an example case that you just didn’t have in your database because you will always have that problem.

And we’ve built out the safety nets, the communication, the workflows, all of that with the team so that when the robot does encounter issue, we take that as more data to our data flywheel. We also have that safety net so that the patients get what they need because that’s the most important aspect of it.

FG: And it is it true in some environments that is that it actually has a badge? Can you share a little bit more about that?

VC: So the way the Moxi navigates hospital wide is it needs to interact with doors and elevators. And the doors are either ADA-compliant buttons or secure areas where you have to scan a badge in order to get to that area. The most challenging one that people never think about, there are many floors and elevator that are badge access only. So you have to scan and then push on a button, which is an incredibly challenging task because you have to scan within a time frame and if you don’t push the button, then you lose access.

So Moxi has a badge in the wrist that we use to be able to interact the environment to get to these secure areas.

FG: One of the things that we hear about a lot is the simulation-reality gap. Share a little bit about how you think about simulations and versus reality training. What does that look like?

VC: We do use a lot of real data, but we see simulation as having a really crucial aspect of setting up the environment when things are static and you’re trying to build out your base controllers. That’s a fantastic place to then use simulation, continue to stress test it. But when it comes to really training out your corner cases and models, that’s where we see things start to fail. And this is where for us, having our data flywheel and just continually collecting all the edge cases that you can’t even dream about really allows us to have robots out there and hit the reliability numbers that you need to hit in order to actually do real deliveries.

FG: When it comes to the design philosophy, you also did something different. A lot of folks have been thinking about humanoid robots. Arms, legs, head. You all came up with a very different form factor. What drove that thought process?

VC: So we in general take a very practical approach. So in 2017 is when we really started the company we went out and raised investor money. We had an arm on our robot and the questions were like why do you have an arm like? And now when we go out and fundraise people ask why do you only have one arm? Where are your legs?

Having one arm is incredibly important because that’s what our task needs. We need to be able to interact with the environment. But if you have two, it’s twice the expense. Same with legs.

We operate in ADA compliant environments. So for us rolling around just makes sense. It’s stable. If the robot loses power, which inevitably it will at times, you can just roll it off to the side as opposed to something on legs falling over or whatnot.

For us, it’s really around what form factor do you need for the work and how does it augment the ability for a person and a robot to work side by side in a teammate fashion?

FG: Let’s talk about edge cases and use cases. What are what are things that Moxi is going to be doing?

VC: One of our most common use cases is chemotherapy drug deliveries. So when you go into a hospital, patients come in and they essentially come in regularly to in order to get their chemo medication. And so for them, the faster they can go get that done and leave, that’s essentially a marker of success because they don’t have to spend as much time there. So before, because the medication is patient specific and has to be made and then delivered to that patient in that time window, you would have a medic, like a pharmacy technician, make that medicine and either they or a nurse are hand walk it to the patient. And often where the pharmacy is and where they have the patients have long corridors and hallways. It could be a 10 minute walk.

Now for a lot of our hospital clients that have this workflow, Moxi will actually go back and forth and making these deliveries on medication. Patients can come in and leave faster than they could before, which is incredibly important for them.

FG: And with the rise of pharmacogenetics, I can just see Moxi is going to be doing even more running back and forth because every patient’s medicine is going to be different.

VC: Exactly. Currently, most medicines are in secure dispensing cabinets. But as more and more medicine is personalized, it has to be more ad hoc. And then how do you get that around the hospital?

^{This article contains general information only and Deloitte is not, by means of this article, rendering accounting, business, financial, investment, legal, tax, or other professional advice or se2rvices. This article is not a substitute for such professional advice or services, nor should it be used as a basis for any decision or action that may affect your business. Before making any decision or taking any action that may affect your business, you should consult a qualified professional advisor. Deloitte shall not be responsible for any loss sustained by any person who relies on this article.}

AI transformation isn't a tech project. Stop treating it like one.

Wed, 10 Jun 2026 14:01:20 GMT

Editor’s note: To our subscribers who have been following the publication for any length of time, you may notice some changes to the look and feel of our site and newsletter. It’s part of an effort to refresh the publication and make it more relevant to practitioners building the enterprises of tomorrow. What isn’t changing is our commitment to bringing you the perspectives you need to make sense of the latest technology trends. And so without further ado, here’s today’s post from Deloitte’s chief technology officer, Bill Briggs

AI transformation isn’t a tech project. Stop treating it like one.

Too many enterprises treat AI like any other technology implementation. What they’re missing is the crucial role process redesign and change management play.

When I look at the AI programs running inside most large enterprises right now, I see a familiar pattern. Substantial investment has gone in. Pilots have been stood up. A center of excellence exists somewhere on the org chart. And yet, when you ask leaders to point to something that has fundamentally changed about how their business operates, the answers tend to get vague.

There’s a reason for that. We’ve been treating AI transformation as a technology project, and technology projects don’t transform organizations. People do. Process does. Leadership alignment does.

The investment imbalance

Here’s a data point that surprised me: our research shows that organizations are putting 93% of their AI investment into technology — the models, the chips, the deployment infrastructure — and only 7% into the people, workflows, and cultural change required to actually use it.1 When I came across that figure, my first reaction was disbelief. My second was recognition. I had felt that imbalance in conversations with clients across industries for months, but I hadn’t been able to quantify it until now.

What passes for “transformation” in a lot of organizations right now is really just procurement. And when the technology doesn’t deliver value, leaders often conclude that the technology failed them. In most cases, the technology wasn’t the problem.

Weaponizing inefficiency

There’s a pattern I’ve seen repeat itself across every major technology wave of the past 30 years, and AI is no exception. The path of least resistance is always to take the new technology and apply it to the way you’ve always worked. It’s fast, it’s familiar, and it avoids the hard conversations about process redesign and organizational change.

The problem is that it doesn’t work. Worse than that, with AI, it can actively make things worse.

If you take an inefficient workflow and automate it with agents, you haven’t eliminated the inefficiency. You’ve accelerated it. You’ve turned what was once a slow, human-paced problem into a fast, machine-paced one, and now you’re paying for it per clock cycle.2 I call this “weaponizing inefficiency.” You meant to build a competitive advantage and instead you built a cost machine.

We saw something similar during the cloud migration era. Organizations moved their legacy processes to the cloud and declared themselves transformed. Then the bills arrived. What looked like modernization was actually just a more expensive version of the old way of working. The same dynamic is unfolding right now with agentic AI, and the stakes are even higher because agents operate with a degree of autonomy that makes course corrections harder.

Grace Hopper, one of the founding figures of modern computing, had a phrase for this kind of thinking: “The most dangerous phrase in the language is ‘we’ve always done it this way.’”3 4That phrase has derailed more transformations than any technical failure I’ve ever seen.

The trust gap hiding in plain sight

Beyond the investment imbalance, there’s a structural problem with how most organizations have approached AI rollout, and it shows up in the data on workforce trust.

We conducted a study on AI trust levels across organizational hierarchies and found that managers and leadership report 26% and 40% higher trust respectively than the average across employee levels, while Staff score 53% below that average.

That gap isn’t a communications problem. It’s a design problem. Organizations built their AI strategies in the executive suite, announced them from the top down, and expected adoption to follow. But frontline employees — the people closest to the actual work — were never part of the conversation about how AI would be used, what it would change, or what it would mean for their roles. So they didn’t trust it. And in many cases, they quietly worked around it.

The irony is that those frontline employees are exactly the people best positioned to identify where AI can actually create value. They know where the friction is. They know which tasks are genuinely tedious and which ones require judgment that can’t yet be automated. When you leave them out of the design process, you don’t just lose their trust. You lose their knowledge.

There’s also a shadow AI problem emerging as a direct consequence of this trust gap. According to Deloitte research, 43% of workers with access to approved AI tools admit to using unauthorized alternatives instead — tools they describe as easier to use and more accurate than what their employers provided.5 That’s not a cybersecurity footnote. That’s a signal that the official AI program isn’t meeting workers where they are.

HR for machines

One area where I think the framing needs to shift dramatically is how we govern AI agents. Right now, most governance conversations are about guardrails — what can the tool do, what is it prohibited from doing, how do we audit outputs. Those are important questions, but they miss something bigger.

As organizations move from deploying a handful of AI tools to running hundreds or thousands of agents across their operations, the governance question starts to look a lot less like a compliance problem and a lot more like a workforce management problem.

Some of the most sophisticated organizations we work with are beginning to treat AI agents almost like employees. They have onboarding processes that define what data the agent can access, what decisions it’s authorized to make, and what actions require escalation. They have performance management policies that evaluate whether the agent is behaving as expected, how its patterns have drifted, and whether it need to be retrained or retired. And they have offboarding policies that determine who inherits accountability for an agent’s decisions once it’s decommissioned.

I’ve raised what might seem like an absurd hypothetical, but it points to a real problem: imagine a human employee creates an agent, that agent spins up five more agents, and a decision made by a fifth-generation agent causes harm. Who’s accountable? What’s the appropriate response?6 These are governance questions, and right now most organizations don’t have answers.

The companies getting ahead of this are the ones treating their AI agents like a silicon-based workforce — thinking seriously about what an HR operating model looks like for digital workers, not just for human ones.

What it actually takes

I’ve said this in boardrooms and I’ll say it here: the technology is ready. Right now, today, without a single additional line of code being written, there is enough proven capability in existing models and agentic frameworks to fundamentally transform how most organizations operate.7 The limiting factor is not the technology. It’s whether leadership is willing to do the harder work.

That work has a few components.

It starts with the CEO making clear, publicly and consistently, that the organization is going to reimagine how it works, not layer AI on top of old workflows, but genuinely rethink them. Enterprises need to have the CEO saying that the organization is going to reimagine everything. This is a moment in time for us to really rewire the nervous system of the entire organization.8

It also requires the executive team to operate in genuine alignment — not just alignment on the aspiration, but on the specific tradeoffs. What does the organization stop doing? How does it handle the workflow disruption that real redesign creates? How does it invest in helping workers whose roles are most affected? According to Deloitte’s Tech Trends 2026 report, virtually every organization we studied is undergoing some form of operating model change as a result of AI — only 1% reported no changes at all.9 The question is whether those changes are being driven by strategy or by accident.

And it requires patience with the learning curve. The organizations I’ve seen succeed with AI aren’t the ones who moved fastest to deploy. They’re the ones who built genuine feedback loops, structured ways that allow them to understand what’s working, what isn’t, and how to improve. That sounds obvious. It’s remarkably uncommon.

The message to anyone waiting for the right moment: paralysis is not a strategy. No matter how much traffic there is, the sooner you leave, the sooner you can get there.”10

Join our community to get emerging tech insights every week.

Subscribe now

Have you seen similar trends in your industry? Share your thoughts in the comments below.

Leave a comment

If this post resonated, forward it to colleagues who would also find the insights helpful!

-- Bill Briggs | Chief Technology Officer | Deloitte Consulting.

^{This article contains general information only and Deloitte is not, by means of this article, rendering accounting, business, financial, investment, legal, tax, or other professional advice or se2rvices. This article is not a substitute for such professional advice or services, nor should it be used as a basis for any decision or action that may affect your business. Before making any decision or taking any action that may affect your business, you should consult a qualified professional advisor. Deloitte shall not be responsible for any loss sustained by any person who relies on this article.}

Tech Trends 2026 | Deloitte Insights

Deloitte CTO Bill Briggs: Reimagining AI Adoption—Beyond Tech Hype to Trust, People and Real Enterprise Impact - Digital CxO

Bill Briggs, quoted in "Deloitte's Bill Briggs on the Future of Enterprise AI Governance and Control," Tech Talks Network, March 2026.

Hulme, Digital CxO, February 2, 2026 (see note 3).

Deloitte's CTO: companies are spending 93% on tech and only 7% on people and that has to change | Fortune

Lichtenberg, Fortune, December 15, 2025 (see note 6).

Hulme, Digital CxO, February 2, 2026 (see note 3).

Bill Briggs, quoted in "Deloitte Tech Trends 2026: Moving from AI Experimentation to Enterprise Impact," Digital CxO, December 15, 2025.

Deloitte, Tech Trends 2026 (see note 1).

Lichtenberg, Fortune, December 15, 2025 (see note 2).

How I learned to prompt AI through the lens of recipe development

Deloitte OCTO — Wed, 03 Jun 2026 14:01:06 GMT

When I first started using generative AI, I made a simple mistake: I treated it like a search engine. As AI is increasingly integrated into our daily lives, effective prompting is now a critical skill. However, effective prompting carries a steep learning curve. The mere-exposure effect states that people develop a preference for what is most familiar to them, hence treating AI like a search engine. However, effective prompting requires a different kind of interaction. It asks for context, judgment, and iteration. What finally broke down the proverbial wall for me was bringing AI to a place I’m familiar with but never would have thought AI would apply: my kitchen.

Step One: Setting the Table/Context

As an avid cook and one who likes to make my own recipes, I use AI to get inspiration for dishes. The first step when prompting is to always give the context that the AI needs to complete your tasks. For cooking purposes, context would be type of meal. What ingredients do I want to emphasize? Am I cooking what is considered fine dining or a simple meal? How many people am I trying to serve with my dish? In both cooking and prompting, giving the most context sets up your prompts to deliver the best results. Given the same ingredients, knowing whether it is for breakfast, lunch, dinner, or a snack wildly changes the thought process for the type of dish you are preparing.

Step Two: Evaluate/ Seasoning to Taste

What makes generative AI so powerful is that it excels at pattern recognition. In my recipe development, it gives me inspiration for how to use certain ingredients and shows me how they are utilized across dishes from many cuisines. By analyzing usage patterns across established dishes, AI condenses what would take me hours of perusing cookbooks and blogs into a curated list. However, what AI cannot do is taste for me. When following recipes online, one of the last steps is always to season to taste. We as individuals all have very different palettes. Something too sweet to my taste buds may taste just right to someone else, and it takes experience to know what works for you. This very much applies to the output of AI models. It’s important to evaluate the output of your prompts and see what is missing. By scrutinizing the output, you’ll begin to see how relatively minor tweaks to your prompts can create major changes in the end product.

Step Three: Iterate

When working on recipes, I rarely perfect my dish on the first try. Maybe it needs more brightness, a different texture, or a simpler approach. Prompting works the same way. Like a failed dish, you need to refine your outcomes based on your evaluations until you achieve your desired results. A useful technique I use in my refinement is called tree of thought prompting, which is to prompt the AI to explore multiple solutions to your prompt. For an activity as interpretive as cooking, it’s useful to have multiple distinct paths to approach a recipe centered on a certain ingredient or concept. For example, that might mean asking for three different ways to build a dish around mushrooms: one rustic, one dinner-party worthy, and one fast enough for a Tuesday night. By selecting from a set of solutions, it helps enhance the refinement process because selecting one solution will tell our AI to ignore the other solutions from the refinement, potentially shortening the time to achieve your desired outcome.

So if prompting still feels awkward or abstract, I would not start with AI at all. I would start with something you already know deeply. Familiarity will not solve everything, but it can give you a bridge. Once I crossed my bridge, prompting became just as familiar as cooking any dish.

Enjoying this post? Subscribe now so you never miss the next one.

Subscribe now

Share your thoughts in the comments below.

Leave a comment

Help spread the word — forward this to your network.

Justin Chen | Consultant | GPS Consulting

Revolutionizing Routine

Deloitte OCTO — Wed, 27 May 2026 14:03:03 GMT

When most people think of artificial intelligence, they picture self-driving cars, human robots, or perhaps conversational chatbots. Yet, some of AI’s most transformative work is happening behind the scenes in the back office, where the core of business operations quietly takes place.

This is where AI transforms essential tasks like invoice processing, supply chain management, employee onboarding, and data management. These functions are fundamental to business continuity but have traditionally been resource-intensive and repetitive, often limiting productivity and innovation. Today, back-office AI is not just streamlining operations; it’s unlocking new levels of efficiency and business value.

What is back-office AI?

The back office is the operational backbone of any organization. It is the domain where transactional processes, reconciliations, and administrative duties keep the business running. Historically, these tasks relied on manual input and repetitive workflows.

Modern AI solutions are changing that. By automating certain tasks, AI is enabling teams to focus on strategic analysis and high-impact projects. Imagine finance teams spending more time generating insights instead of chasing spreadsheets or HR driving culture-building initiatives rather than drowning in paperwork.

AI adoption in the back office can be visualized as a maturity curve:

Basic automation: Rule-based bots handle repetitive tasks such as document routing and notifications.
Process automation: AI capabilities extend into more complex workflows, like invoice processing or automated onboarding.
Predictive analytics: Machine learning analyzes trends, forecasts demand, and informs smarter decision-making.
Prescriptive AI: Tools recommend optimal actions based on predicted outcomes.
Generative AI: Advanced models create content, draft reports, and design new business processes.

Each level builds on the last, transforming the back office from a cost center into a hub of innovation and competitive edge.

Agentic AI: What’s next

Building on these fundamentals is agentic AI, a new level of capability where autonomous digital agents are capable of making and executing context-aware decisions. Agentic AI dynamically responds to changing business conditions without manual intervention. This is the future of operational resilience: AI that continually adapts and optimizes core processes.

Subscribe now

Why does this matter? The pace of business is accelerating. Rising costs, changing workforce expectations, and persistent talent gaps are pushing organizations to work smarter and faster. Back-office AI isn’t merely automating the routine—it’s about freeing human potential to focus on innovation, strategy, and value creation.

The real opportunity lies not just in improved speed, but in better outcomes and more empowered employees.

AI is making the transition from back-office support to a core driver of enterprise strategy. This shift has far-reaching implications for leadership and organizational growth. The conversation around AI has matured and is no longer a technical novelty, it’s a strategic imperative, especially in back-office operations.

A new perspective: AI as a business priority

Just a few years ago, back-office AI was perceived as supplementary, an experiment to improve efficiency here and there. Today, it’s a boardroom priority. CFOs, COOs, and business leaders are acting with a sense of urgency.1 They see the data. They understand that in a competitive landscape, AI can be a decisive differentiator.

Why the shift?

1. Data-driven decision making: The back office contains a treasure trove of data—financial transactions, supply chain metrics, HR analytics. AI analyzes this data for valuable insights, enabling faster, smarter decisions.

2. Cost and efficiency pressures: As margins tighten and expectations rise, organizations leverage AI to automate labor-intensive work, minimize errors, and redirect talent from routine operations to strategic work.

3. The talent crunch: With skilled professionals in high demand, AI fills critical gaps, helping teams to do more with less available resources and empowering leaders to focus on growth.

What sets industry leaders apart is how they embed AI directly into their enterprise platforms, from ERP systems to HR management tools. This integration elevates forecasting, enables real-time supply chain adjustments, and strengthens workforce planning, making the back office a strategic powerhouse.

When AI is embraced at scale, it becomes a catalyst for agility and collaborative decision-making. Teams gain richer insights, leaders navigate through uncertainty with greater confidence, and employees are empowered to contribute creatively.

Transitioning to back-office AI isn’t without obstacles. Integrating with legacy systems and managing change effectively are essential. Success in any business requires a thoughtful approach that builds trust and ensures technology aligns with daily operations.

Leave a comment

What’s ahead

In our next blog, we will showcase real organizations that are unlocking tangible value through back-office AI. Their success stories offer actionable lessons and practical insights for any business ready to begin its AI journey.

Stay connected! Transformation is the beginning.

CFO Signals™ 4Q 2023 | Deloitte US

—Richard Solari | Managing Director | Deloitte Consulting LLP

This article was originally published on November 20, 2025.

The digital workforce is here. The governance isn’t.

Deloitte OCTO — Thu, 14 May 2026 13:50:03 GMT

The question used to be whether enterprises would actually embrace AI agents at scale. Somewhere in the last six months, that question quietly resolved itself. Walking the floor at Google Cloud Next in Las Vegas last month, it was clear the conversation had moved on. The new question is what happens when the agents start showing up to work before anyone has built HR.

That’s not a metaphor. Agents are now acquiring privileged access to systems like databases, email, financial systems, and operational tools. They are executing tasks across enterprise environments with minimal human oversight, often coordinating with other agents in multi-step chains that no single person fully monitors. The infrastructure to deploy them is maturing fast. The infrastructure to govern them is, at best, nascent.

Photo by Accuray on Unsplash

From pilot to pressure

Most enterprises are now piloting or deploying multi-agent architectures. That sounds like progress, and it is, but, in many cases, enthusiasm has outrun execution. The failure patterns are becoming well-documented and increasingly expensive. Agents may misinterpret ambiguous instructions leading to confident misalignment. Errors compound as tasks pass between agents in a chain, creating semantic drift. As context gets lost at handoff points, the final agent in a sequence executes based on degraded information.

These misinterpretations frequently trigger infinite reasoning loops, where an agent sensing its output does not match the goal and repeatedly attempts to self-correct. Without deterministic circuit-breakers to limit such behavior, this often leads to excessive data retrievals, massive token cascades, and recursive API spikes that can exhaust budgets in minutes. These loops are often exacerbated by poorly implemented RAG systems where an agent retrieves irrelevant or contradictory data and enters a hallucination loop trying to reconcile the noise.

Multi-agent systems currently fail in production at high rates not because the agents are incapable, but because the scaffolding around them is immature. 1 Central to this scaffolding is rigorous data governance. It is the only way to ensure agents have a single source of truth and well-defined semantic meaning to combat semantic drift. In fact, half of all generative AI projects are abandoned before reaching production because the underlying data isn’t clean, consistent, or accessible enough to build on. Even the most sophisticated agentic platform in the world runs on data. If the data is broken, the agents are broken.

Infrastructure is now a workforce decision

One of the subtler shifts happening right now is that infrastructure decisions are being reframed in workforce terms. This isn’t about which cloud is fastest or cheapest in traditional benchmarks. It’s about what compute fabric can support millions of concurrent agent operations with near-zero latency. A pizza chain’s voice ordering agent that responds in under 200 milliseconds is more than a fun demo. It’s a concrete proof point about what this infrastructure actually needs to deliver.

The billing model is changing along with technology. Some customers are already pushing back on traditional API-call pricing. If an agent runs continuously on behalf of a business, treating each inference as a discrete transaction starts to feel like charging a salary by the word. The industry is moving, haltingly, toward execution models that treat agents more like employees who are persistent, always-on, and measured by outcomes rather than API calls.

The platform vendors understand this. New tooling for agent-to-agent communication, agent identity management, and policy enforcement at the agent layer has started appearing in earnest. Open protocols designed to let agents from different vendors interoperate are gaining traction. But open protocols and a functioning governance architecture are not the same thing, and right now enterprises mostly have the former without the latter.

The question nobody has answered

The sharpest conversations happening right now are about accountability more than ability.

For example, what happens when an agent makes a decision that causes harm? This could be an erroneous financial transaction, a workplace incident involving a robot operating on AI-generated instructions, a customer communication that creates legal exposure. Who is responsible? The organization that deployed the agent? The vendor whose model powered it? The integrator who built the workflow? In a multi-vendor agentic stack, which is the only kind most enterprises will actually run, the liability chain is genuinely unclear. The current answer in most organizations is a set of disclaimers that effectively stall deployment. That’s not a long-term solution.

Closely related, there is currently no reliable way to audit a decision made by a multi-agent system. An agent acting in good faith on instructions from another agent acting on instructions from a human prompt three steps back doesn’t leave a clean paper trail. For most enterprise use cases — finance, healthcare, legal, HR — the ability to reconstruct why a decision was made isn’t a nice-to-have. It’s a compliance requirement. That capability doesn’t exist yet at production scale.

None of this means the technology isn’t real or that enterprises should wait. But it does mean that the organizations moving fastest right now are accumulating governance debt alongside capability gains. The companies that will come out ahead aren’t necessarily the ones who deploy the most agents in 2026, rather they’re the ones who build the accountability architecture to operate them reliably at scale.

What comes next

Many compare the current state of enterprise agent use to the early days of cloud adoption, when IT teams were spinning up infrastructure faster than security and compliance could catch up. Eventually, governance caught up, but only after expensive lessons in data exposure and shadow IT sprawl. We are at a similar crossroads today, but with one critical difference: organizations cannot afford a decade-long learning curve.

The agentic governance problem is an order of magnitude more complex than the cloud governance problem. Cloud resources are passive. They don’t make autonomous decisions, coordinate with each other without being explicitly told to, or act with privileges scoped to a human employee’s identity. Agents do all three. There are many bright spots that are emerging such as agent identity management and agentic gateways. Much like API gateways of the past, these new control planes act as smart policy enforcement points for the digital workforce where agents’ intent is understood and checked before allowing execution. It also ensures that every agent has a unique, verifiable ID and that its actions are filtered through a security layer that can block unauthorized prompts and transactions in real-time. This shifts the requirement from static compliance to real-time observability and deterministic enforcement. The old governance playbook, designed for static services and human-triggered events, needs a substantial rewrite, not just a digital update.

The 2027 outlook: The need for transparent agentic computing

As we look toward 2027, the winners will likely be the organizations that move beyond fragmented tools and pilots to build robust agentic platforms. To achieve true scale, agentic platforms must anchor on three mature pillars:

- Real-time observability: Enterprises will move past basic logging to live traceability, utilizing high fidelity integrated agentic monitoring tools and dashboards to visualize agent reasoning paths as they unfold. This includes the emergence of supervisor agents designed to circuit-break runaway processes and a comprehensive array of kill switches that are required to give leadership the peace of mind to let agents operate autonomously.

- Universal semantic layer: The data problem will be resolved by a mature semantic layer acting as a universal translator. This requires revamping our understanding of data governance and data architecture to create a foundation suitable for transparent agentic computing where every agent shares a singular, governed understanding of corporate logic, effectively neutralizing the semantic drift that plagues today’s pilots.

- The immutable audit trail: Future architectures and agentic platforms must produce step-by-step immutable records of every autonomous decision. This creates provenance of agentic thought that will help in meeting the demands of stringent compliance and legal requirements.

The infrastructure is ready for a digital workforce. The question is whether the enterprise is.

Enjoying this post? Subscribe now so you never miss the next one.

Subscribe now

Have you seen similar trends in your industry? Share your thoughts in the comments below.

Leave a comment

Help spread the word — forward this to your network.

Mamoun Hirzalla | Alphabet CTO at Deloitte

Why Half of GenAI Projects Fail: Avoid These 5 Common Mistakes

How privacy works in the age of AI personalization

Deloitte OCTO — Wed, 13 May 2026 14:01:39 GMT

A few weeks ago, I found myself mindlessly hovering over a skin care product page. I didn’t click or save anything, just paused for a second, lost in the usual scroll. Fast-forward to later that evening, a totally different app on my phone buzzed with a notification featuring content eerily close to what I’d just browsed. Different app, different session, zero obvious connection. Suddenly, it didn’t feel like helpful personalization anymore. It felt like someone was peeking over my digital shoulder taking notes.

When we talk about personalization, most people imagine apps politely reacting to what we explicitly tell them. But in reality today’s AI is more like a super-observant friend who catches every subtle cue, the pause before a click, the stuff you nearly choose, the things you ignore. Sure, it’s technically mind-blowing that these systems can guess our preferences from digital breadcrumbs. But from the human side, it’s a bit uncanny, like the algorithm knows secrets we never meant to spill.

Most of us shrug off the idea that some anonymous server is collecting our data. The real gut-punch comes when an algorithm nails a recommendation so perfectly that it feels like it read your mind. Bad suggestions are easy to dismiss, but those bullseye moments make you pause and think about how it knows, what it assumes, and whether you have any say in your digital doppelgänger.

What is contextual integrity and how can it help users?

This is where the idea of Contextual Integrity1 comes in. The concept comes from Cornell professor Helen Nissenbaum. It’s less about keeping personal data secret or maintaining full control over its use and more focused on ensuring data flows conform to prevailing norms within a given context. Privacy isn’t just about keeping secrets, but about making sure your story is told on your terms.

However, today, these systems are making calls about who we are, often without our input or consent. All this digital sleuthing has become nearly invisible. Old-school systems relied on things you could actually see, your shopping cart, your search history, your saved playlists. But now it’s abstract, almost ghostly. Deep learning models chew through thousands of micro-signals, like how fast you scroll, where your mouse lingers, even your hesitation on a certain photo. Maybe you hover over hiking boots for half a beat, and suddenly the AI predicts you’re planning a backpacking adventure or going for a jog, even if you never type a single word about it. These insights are born from behavioral whispers, not shouts. Sometimes, even the engineers building these systems can’t explain exactly how they connect the dots. That’s the imbalance. The system is always watching and learning, but you’re left in the dark about what it really knows. When you can sense that something’s happening but can’t see the gears turning, trust starts to wobble.

Why transparency is no silver bullet for AI privacy

Whenever AI and privacy come up, someone inevitably says, “We just need more transparency!” Better explanations, clearer consent buttons, more peeks behind the curtain. And sure, that helps, but it’s not the whole story. Even if an AI gave you a perfect rundown of every data point it used, would that make you feel better? Sometimes, it just makes things weirder. What people really crave is that these systems behave with a little social savvy respecting unspoken boundary, letting awkwardness and ambiguity exist. Transparency is nice, but the real test is whether an interaction actually feels right.

So, where’s the line? How smart is too smart? The myth that “more accuracy is always better” just doesn’t hold up. When recommendations cross into uncanny valley territory and become so relevant they’re almost creepy, people pull back. Sometimes, there’s real value in holding back better to be generally helpful than weirdly spot-on and invasive. A little restraint can go a long way.

Businesses measure personalization in clicks, conversions, and hours glued to a screen. But trust is a different conversation. It doesn’t fade gradually but rather can vanish in a flash thanks to one awkward notification. Sure, regulations like GDPR give us rules, but the real challenge is much messier.

The case for respecting context in personalization

How do we build systems that never trigger that “ick” feeling to begin with? Maybe the real question isn’t “How much CAN we personalize?” but “When does personalization cross the line from cool to creepy?” That line shifts for everyone, depending on mood, context, and what’s at stake. But one thing’s universal: people want to feel seen, not surveilled. Recognized, not reduced to a bundle of data points by an algorithm they can’t control.

The best systems aren’t the ones that try to know it all. They’re the ones that know when to take a step back. In the end, people want more than just eerily accurate suggestions. They want space, agency, and the simple dignity of defining themselves on their own terms.

Join our community to get emerging tech insights every week.

Subscribe now

If this post resonated, forward it to colleagues who would also find the insights helpful!

Share your thoughts on these insights in the comments below.

Leave a comment

By Jeshwini Gopinath | Consultant | GPS Consulting

"Privacy as Contextual Integrity" by Helen Nissenbaum

Tech Trends 2026 update: New compression tricks could change how enterprises think about AI infrastructure

Deloitte OCTO — Wed, 06 May 2026 14:02:51 GMT

The popular story about AI progress has mostly been a story about scale. Bigger models, more data, more powerful chips. That story is still being written. But challenges around how AI applications use memory are increasingly driving how businesses actually experience AI.

Specialized, high-speed GPU memory is extraordinarily expensive and physically limited. That’s one of the reasons why some businesses are rethinking their infrastructure decisions. As we described in this year’s Tech Trends report, as they scale AI applications across the organization, enterprises are discovering their existing infrastructure strategies aren’t designed for AI’s demands. The big driver is the computational work an AI model does every time it responds to a query, known as inference. Recurring AI workloads mean near-constant inference, and when using a cloud-based AI service, this can lead to frequent API hits and escalating costs. Some organizations are seeing monthly AI bills climb into the tens of millions of dollars.

Why do AI computations require so much memory? Every time an AI model processes your input, it needs to hold an enormous amount of information in fast-access memory. The component at the center of this is something called the key-value (KV) cache. This essentially functions as the model’s short-term memory. It’s a working record of the conversation so far, the documents you’ve shared, and the context the model needs to make sense of your question. The longer or more complex that context gets, the more GPU memory it demands.

The good news is that developers are working on solutions that use existing memory more efficiently to help AI models run faster and improve their capabilities.

New KV cache compression techniques make AI more resource-efficient

Many solutions focus on shrinking the footprint of the KV cache. Known as KV cache compression, this helps optimize memory usage and computational efficiency during inference. A few KV cache compression techniques exist. One of the most prevalent is quantization — the practice of representing data with less numerical precision. For many practical purposes, the less precise measurement is good enough, and it takes up far less memory.

Traditional quantization methods have a catch, though. They introduce a kind of bookkeeping overhead, requiring extra data to track how the compression was applied. That overhead can eat up a significant chunk of the memory savings.

A new algorithm reduces that overhead substantially through a two-step process. Developed by Google researchers, TurboQuant uses an approach that converts data into a more compact mathematical form. Then it applies a single “error-correction” bit using a separate technique that catches and eliminates any small inaccuracies introduced in the first step.

According to Google’s benchmarks, this approach can result in an up to 8x speedup in certain computations and a 6x or greater reduction in memory usage, while maintaining full accuracy across a range of AI tasks, including question answering, summarization, and code generation.1

Meanwhile, an algorithm known as ChunkKV reimagines the basic approach to compression. Typically, to save memory, LLMs delete certain tokens in the KV cache that it deems unneeded. Developed by researchers at the Hong Kong University of Science and Math, ChunkKV deletes semantic chunks rather than isolated tokens as the basic unit, an approach that preserves linguistic context more faithfully under aggressive compression, improving throughput by 26.5%.2

Elsewhere, FlashAttention is not a KV compression approach at all, but an algorithm that’s integrated into popular machine learning libraries. It optimizes how AI models read and write memory during attention computations. The latest version, FlashAttention-4, achieved a 20% speedup on typical AI infrastructure.3 A separate research line, KVTC, uses a combination of dimensionality reduction, adaptive quantization, and entropy coding to achieve up to 20x compression while maintaining reasoning and long-context accuracy.4

The field is moving fast, and the momentum is real.

Practical applications

Tech Trends 2026 highlights a significant shift underway in how enterprises are thinking about where to run their AI workloads. Cloud-first approaches are giving way to more hybrid strategies, partly for cost reasons, but also because of data sovereignty concerns, latency requirements, and the need for resilience. Real-time AI workloads demand proximity to data sources, especially in manufacturing environments and autonomous systems, where network latency prevents real-time decision-making.

The problem with running powerful AI models on local or edge infrastructure has always been that those environments have far less memory and compute to work with. Efficiency techniques change that calculus. A model that previously required a large GPU cluster to run might become viable on more modest hardware once its memory footprint shrinks by a factor of six or more.

This matters in a world where the majority of enterprises’ data still resides on premises, and organizations increasingly prefer bringing AI capabilities to their data rather than moving sensitive information to external AI services. Smaller, faster models are central to making that vision practical.

What this means for AI infrastructure

The compression techniques described above are just single data points in a larger story about the AI industry’s growing focus on efficiency. For the past few years, progress in AI has been all about scaling through bigger models, more data, and more compute. That approach has delivered remarkable results, but it has also run into physical and economic ceilings that are becoming increasingly hard to ignore.

As we noted in Tech Trends, inference costs have plummeted 280-fold over the last two years, yet enterprises are experiencing explosive growth in overall AI spending because usage has dramatically outpaced those efficiency gains. Emerging compression algorithms may drive a similar pattern. When technology advances in efficiency and capability, people often use more of it, not less.

For enterprises wrestling with AI infrastructure decisions right now, the key takeaway isn’t necessarily to wait for emerging compression techniques. It’s that the software side of AI is advancing rapidly alongside the hardware side. Decisions about infrastructure strategy should account for both. A hybrid architecture that looks barely feasible today, constrained by memory and compute requirements, may look quite different once another generation of compression and optimization techniques filters down from research labs into production systems.

Join our community to get emerging tech insights every week.

Subscribe now

Have you seen similar trends in your industry? Share your thoughts in the comments below.

Leave a comment

If this post resonated, forward it to colleagues who would also find the insights helpful!

Ed Burns | Editor | Office of the CTO

TurboQuant: Redefining AI efficiency with extreme compression

NeurIPS Poster ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference

We reverse-engineered Flash Attention 4

NVIDIA Researchers Introduce KVTC Transform Coding Pipeline to Compress Key-Value Caches by 20x for Efficient LLM Serving - MarkTechPost

You can't manage what you don't measure: A scientific approach for organizations to get a handle on their tech value realization

Deloitte OCTO — Wed, 29 Apr 2026 14:03:17 GMT

Many organizations today are obsessed with adding—new AI pilots, tools, platforms, and more spending on tech. And to be fair, the logic makes sense. Data shows that increasing investments in AI, data, and IT could significantly improve financial performance, sometimes even doubling earnings per share (EPS) in modeled scenarios.

But what many organizations may have missed is something quieter, almost invisible. Better data can amplify AI, and stronger infrastructure can reduce friction. But technical debt can slow everything down, and tech capacity cuts can delay change in ways that aren’t immediately visible.

This is not a story about “AI wins” or “cost cuts lose.” It’s a story about what happens when decisions interact, especially when they’re layered onto aging foundations. Technology decisions don’t work in isolation; they behave more like ecosystems. Over time, each decision creates ripple effects, some that accelerate progress, and others that introduce drag.

Five years from now, the most interesting thing about high-performing organizations will be what they’ve stopped ignoring rather than what they’ve added.

A tale of two S&P 500 organizations: “Average” vs. “Ambitious.”

To see how this plays out in the real world, our team considered two hypothetical organizations of similar size and starting earnings profile:

An average S&P 500 organization: A typical S&P 500 company with median IT spend, moderate digital maturity, fairly strong (but not exceptional) organizational change capability, and limited AI activity beyond pilots. Its strategy is to stay broadly aligned to market averages. In the simulation, the organization starts at EPS of US$2.00 in year one and grows to US$5.17 by year five, reflecting typical market performance.
A more ambitious competitor: The organization has a similar starting point, but intentionally tests targeted interventions—scaling AI, advancing data maturity, adjusting tech investment, optimizing workforce capacity, and strengthening change execution—to assess how different choices translate into different outcomes over time.

We then applied systems dynamics to model how technology, organizational, and financial factors interact over time within our hypothetical enterprises.

How do the two companies perform?

At first, tech decisions like AI investments, infrastructure modernization, data strategies, and workforce shifts may not seem connected. But over time, they begin to interact. Some reinforce each other and drive performance gains, while others create friction or only become visible later, when the organization struggles to keep up.

When tech decisions align, value compounds

What became clear over time wasn’t just that investment drives growth, but also how uneven that growth becomes when decisions aren’t connected.

In our model, the difference between “average” and “ambitious” enterprises comes down to a manageable set of choices that reinforce each other. When investment in infrastructure, data, and AI moves together, the system builds momentum. A sustained 5% annual increase in IT budget (aimed at modernization, capability building, and adoption) lifts EPS to US$6.42 by year five, a 24% increase over the baseline path.

Scaling AI further accelerates this effect. A 35% increase in AI maturity raises EPS by 45% over five years, translating to up to US$269 million in annual earnings per company and around US$135 billion in collective gains across the S&P 500.

In the future, growth won’t likely come from how much organizations invest, but from how well those investments reinforce each other.

Technical depth isn’t just a cost, but a growth constraint

The “average” organization can look stable from the outside. Systems are running, investments are being made, and new capabilities are layered in. But underneath, value quietly leaks through friction.

Technical debt is rarely treated as a quantified growth constraint, yet it impacts how fast an organization can actually move. Deloitte’s 2026 Global Technology Leadership Study estimates that technical debt can account for 21% to 40% of IT spending, which helps explain why adding new tools on top of aging foundations often feels like “more effort, same speed.”

The model highlights how reducing friction reshapes the system. Targeted infrastructure modernization can reduce tech debt by up to 10% in year one and by up to 18% over five years compared to an “average” organization. At the same time, a 35% increase in data capabilities over 18 months reduces latent potential by 8.5% in two years and 52.5% by year five.

You can’t “AI your way out” of technical debt. The real advantage will come from reducing drag and unlocking what already exists because the decisions made today will shape performance for years to come.

Transformation moves at the speed of digital readiness—not technology

If there’s one place where the system breaks down fastest, it’s here. Technology can be deployed quickly, but organizations take longer to absorb change.

In our model, we saw that a one-time 20% reduction in tech headcount creates an “invisible backlog,” with digital capability dropping 5.7% below average. Even when paired with a 35% increase in data and AI maturity, digital readiness declines in the short term—4.5% in year one and 6.2% in year two. Over time, the system begins to recover slowly, and the readiness gap narrows by year three.

Digital capability can scale quickly, but digital readiness takes time because sustaining change depends on people, processes, and execution, and not just technology.

The future belongs to systems that work as one

If you take a step back and look at these decisions, a pattern begins to emerge. What once felt like separate choices—AI investments, infrastructure changes, cost decisions, workforce shifts—starts to behave as part of a larger system.

Looking ahead, the advantage will likely come from how well systems work together in addition to how much tech organizations deploy. The ones that pull ahead will be those that reduce friction, align their decisions across data, AI, and infrastructure, and build the readiness to sustain change. Because over time, how the system evolves will matter more than any individual moves your organization makes.

Join our community to get emerging tech insights every week

Subscribe now

If this post resonated, forward it to colleagues who would also find the insights helpful!

Share your thoughts on these insights in the comments below.

Leave a comment

Ireen Jose | Analyst - Center for Integrated Research

From Mood Board to Move-In: AI’s Growing Role in Home Renovation

Deloitte OCTO — Wed, 22 Apr 2026 14:02:21 GMT

Last year, I stood with my friend, Rachel, admiring her new fireplace feature, complete with custom shelves that we had envisioned and built together. Taking it in, I couldn’t help but reflect on how different that journey would have been just a few years earlier. The roadblocks we ran into would have slowed us down at best or nudged us to outsource the project entirely. That experience was one of the moments that clarified for me how artificial intelligence (AI) has quietly woven itself into everyday problem-solving. It’s an experience that shows how, with the right tools and prompts, people can take on new tasks and responsibilities.

This is equally true in today’s workplace as it is in the home. I’m an industrial engineer by training, with minimal experience in construction or power tools. Like many people, I’d used a drill to mount a TV or hang something on the wall, but I wouldn’t mistake myself for a contractor. Precisely because my starting point is familiar, I wanted to explore how AI can help inspire creative minds to nurture a budding idea and do more than they previously felt capable of doing.

Why design and build it yourself?

People take on do-it-yourself (DIY) projects for a range of reasons: creative control, personal satisfaction, and the opportunity to develop new skills. For many, the primary driver is much more practical: cost. A study surveying homeowners, found that more than 60% of people begin DIY projects primarily to save money, reflecting a common priority: delivering meaningful improvements without expanding the budget.1

Done thoughtfully, DIY can significantly reduce costs by eliminating labor expenses and minimizing the premiums often built into contracted work. And while hiring a contractor or designer can be the right decision for complex or time-sensitive projects, it can also increase overall spending substantially, sometimes nearing twice the cost of doing it yourself. 2

The current barrier: why don’t more people take on DIY challenges?

Even when the motivation is there, many people hesitate to start a DIY project because the commitment can feel high and the margin for error feels low.¹ It tends to be less about ambition and more about risk, capability, and credibility of the result.

This is where AI can act as a practical companion, helping with everything from early-stage concepting to the granular methodology for tackling an idea. Whether you need a spark of inspiration or a full scaffold of support, AI can help carry the work to the finish line.

At a high level, AI can provide:

Concept development to turn a vague idea into a clear plan
Methodology and sequencing so you know what to do first, next, and last
Materials and tool recommendations tailored to the specific job you’re tackling
Troubleshooting support when you hit a snag mid-project

When I was helping to renovate Rachel’s living room and wanted to incorporate built-ins framing for a wood-burning fireplace, AI served as a steady reference point, helping me think through structural considerations, validate appropriate materials and tools, and translate a big ambition into clear, manageable phases for ease of execution.

It sharpened the more tactical decisions: recommending appropriate types of wood for shelving to best balance durability and finish, advising how to engineer the shelving structure to carry heavier loads, and helping to guide what joinery techniques would improve the strength and stability of the frame. As we considered resurfacing the fireplace in a lighter material, AI helped us evaluate material options that could perform safely near high heat surfaces.

Many people who are using AI in the workplace are seeing similar benefits. A recent Gallup pole found high numbers of employees using AI to generate ideas and learn new skills.3 People are using AI’s broad capabilities to take on new responsibilities that may not have been a part of their job description in the past.

Looking forward

Research suggests that completing a DIY project can lift confidence, deepen happiness, and leave a lasting sense of satisfaction—benefits that linger long after the final reveal.4 And while the backdrop may be a home renovation, the impact translates into life and work alike. To set a goal, stay with it through uncertainty, and work patiently through obstacles is to forge a more durable kind of confidence. It affirms a truth many professionals rediscover again and again: growth is often waiting just beyond the hard part.

Each morning, Rachel enjoys her coffee in a room she brought to life, an everyday proof that we can build what we once only imagined. It’s a quiet reminder to greet the next challenge with open hands. Looking ahead, AI can help us step into projects and experiences we might have hesitated to try before by making the starting point clearer and the steps less intimidating. The real alchemy is when intrinsic motivation is joined with AI’s ability to sharpen, support, and bring your goals to life.

Enjoying this post? Subscribe now so you never miss the next one.

Subscribe now

Share your thoughts in the comments below.

Leave a comment

If this post resonated, forward it to colleagues who would also find the insights helpful!

Emily Schildmeyer | Customer and Marketing Consultant

DIY Statistics

House Renovation Costs: A Complete Breakdown for Homeowners & Investors

AI Use at Work Rises

Home DIY Projects Improve Confidence and Happiness | Engoo Daily News

Agentic commerce: Why value and trust may become the new scorecard of business

Deloitte OCTO — Wed, 15 Apr 2026 14:02:58 GMT

Key insights

Agentic commerce is emerging quickly: AI agents are increasingly helping consumers research, compare, and purchase products online.
“More value for the price” is becoming a critical competitive signal: Consumers are prioritizing brands that deliver the strongest value exchange, not simply the lowest price.
Predictive customer insights may outperform traditional loyalty metrics: Signals that track shifts in customer perception relative to competitors can forecast growth months ahead.
Trust is becoming a foundational requirement for AI adoption: Reliability, transparency, capability, and humanity all influence whether people trust AI-driven decisions.
Businesses may need to organize around outcomes: Companies that focus on delivering measurable customer outcomes, not just selling products, may be best positioned for the era of AI-mediated commerce.
What is agentic commerce and why is it important?
First web revolutionized commerce, then mobile did the same. The next transformation may be driven by AI agents. These tools are increasingly becoming part of the buying process, helping people research options, compare prices, evaluate product reviews, and sometimes even complete purchases.
This emerging model, often referred to as agentic commerce, could reshape how brands compete online. Instead of consumers manually searching and evaluating products, AI systems increasingly act as intermediaries by analyzing data and recommending the best option based on the customer’s goals. In this environment, the brands that win may not be the loudest or most visible. Instead, they may be the ones that consistently demonstrate value and trust.
Why agentic commerce is on the rise
Economic uncertainty and shifting expectations are pushing consumers to rethink how they spend. But the shift is about more than finding the lowest price. Consumers are increasingly evaluating value. Premium brands can still deliver strong value if the experience, quality, or outcomes justify the price. Conversely, even discounted products may struggle if they fail to meet expectations.
These dynamics are already changing consumer behavior. For example, higher-income shoppers who historically avoided discount retailers are increasingly visiting value-oriented stores when they perceive a stronger value equation.1 The implication for companies is significant: customer loyalty is becoming more fluid and consumers are increasingly willing to switch brands when they find better value.
The ways in which companies used to connect with customers and develop relationships with repeat buyers are becoming less relevant in the current data-driven world.
Why agentic commerce demands new metrics
For decades, companies relied on customer satisfaction scores, loyalty indicators, or net promoter score to gauge brand health. But these indicators often measure past experiences rather than future momentum. What matters more may be how customer perception is shifting relative to competitors.
When organizations analyze large-scale customer feedback over time, they can often identify moments when perceptions begin to diverge. Such signals may predict changes in revenue growth or market performance months later. In this sense, predictive customer insight can act as a forward-looking navigation system, helping leaders see what may be coming rather than simply measuring what already happened.
The rise of AI agents amplifies the importance of value signals. AI systems evaluating products do not respond to marketing messages the same way humans do. Instead, they rely heavily on data signals that indicate which product delivers the best outcome. This could fundamentally reshape marketing and brand competition. If AI agents prioritize measurable customer outcomes and verified experience data, brands that consistently deliver real value may rise faster, while those relying primarily on advertising or brand perception could struggle.
Trust becomes the foundation of AI-driven commerce
For agentic commerce to scale, however, trust will become even more important than it currently is. Consumers and enterprises are still cautious about AI systems making decisions on their behalf. Trust depends on several elements working together, including reliability, transparency, capability, humanity, and purpose. Brands that can demonstrate that they’ve centered their products and services around these ideas are likely to perform better with consumers.
Another emerging factor is the quality of the data behind AI systems. As AI models increasingly generate synthetic data, human-verified insights may become critical for grounding decisions and improving trust. Without trustworthy data, even powerful AI systems may struggle to gain widespread adoption.
Agentic commerce shifts organizations toward outcome-driven goals
Taken together, these trends suggest companies may need to rethink how they operate. Historically, many organizations structured themselves around products, sales channels, or geographic markets. But as AI mediates more decisions, those structures may be less relevant. The companies that succeed may be those that organize around a simpler question of whether they created meaningful outcomes for their customers. Because when AI agents evaluate options, compare experiences, and recommend products, they will likely prioritize brands that consistently deliver strong outcomes. That could redefine the scorecard of business.
This post is an adaption of Deloitte’s new FWD: Technology podcast with special guest Rob Pace, Founder and CEO of Hundred X. Hosted by Bill Briggs (Chief Technology Officer), Brett Davis (Chief Innovation Officer), and Simona Spelman (Chief Human Officer), the show brings clarity to a world of complexity, helping you move from buzzwords to business impact. Each episode features stories, insights, and hard-earned lessons from builders who know that innovation only works when people know what matters…and why.
Click here to hear the full interview and catch up on previous episodes.
Help spread the word — forward this to your network.
Share
Join our community to get emerging tech insights every week.
Subscribe now
^{This article contains general information only and Deloitte is not, by means of this article, rendering accounting, business, financial, investment, legal, tax, or other professional advice or services. This article is not a substitute for such professional advice or services, nor should it be used as a basis for any decision or action that may affect your business. Before making any decision or taking any action that may affect your business, you should consult a qualified professional advisor. Deloitte shall not be responsible for any loss sustained by any person who relies on this article.}
^{As used in this document, “Deloitte” means Deloitte Consulting LLP, a subsidiary of Deloitte LLP. Please see www.deloitte.com/us/about for a detailed description of our legal structure. Certain services may not be available to attest clients under the rules and regulations of public accounting.}
^{Copyright © 2026 Deloitte Development LLC. All rights reserved.}

Rob Pace, interview with Deloitte, March 5, 2026.

Field Notes from the Future

Agentic commerce: Why value and trust may become the new scorecard of business

Key insights

What is agentic commerce and why is it important?

Why agentic commerce is on the rise

Why agentic commerce demands new metrics

Trust becomes the foundation of AI-driven commerce

Agentic commerce shifts organizations toward outcome-driven goals

Real-World Wins: How back-office AI is delivering results

Real-world use cases

Strategies that support success

Lessons for all

Beginning your AI journey

Key capabilities to access

Essential strategies

The bottom line

The new AI edge is discernment, not enthusiasm

Context is king: Optimizing infrastructure for inference

Open-weight AI models are all about what you want to manage vs. what you don’t

When measuring the wrong thing is worse than not measuring at all

Quantum's awkward middle years are finally getting interesting

A date on the calendar

The clock that’s already ticking

Value while you wait

The takeaway

How technology is reshaping music production

What a Recording Session Used to Be

The Revolution That Started Before AI

What AI Actually Adds

Robots are ready. But where do you start?

Safety stopped being a sidenote

The debates are everywhere. The answers are not.

Finding the focus

From Code to Conversation: How We Stopped Building and Started Talking

The Three Eras of Creation

It Is Not Just About Code

What This Means for Organizations

The New Scarce Resource

A Word of Caution

The Bigger Picture

Navigating hospitals with some help from physical AI

AI transformation isn't a tech project. Stop treating it like one.

AI transformation isn’t a tech project. Stop treating it like one.

The investment imbalance

Weaponizing inefficiency

The trust gap hiding in plain sight

HR for machines

What it actually takes

How I learned to prompt AI through the lens of recipe development

Step One: Setting the Table/Context

Step Two: Evaluate/ Seasoning to Taste

Step Three: Iterate

Revolutionizing Routine

What is back-office AI?

Agentic AI: What’s next

A new perspective: AI as a business priority

What’s ahead

The digital workforce is here. The governance isn’t.

From pilot to pressure

Infrastructure is now a workforce decision

The question nobody has answered

What comes next

The 2027 outlook: The need for transparent agentic computing

How privacy works in the age of AI personalization

What is contextual integrity and how can it help users?

Why transparency is no silver bullet for AI privacy

The case for respecting context in personalization

Tech Trends 2026 update: New compression tricks could change how enterprises think about AI infrastructure

New KV cache compression techniques make AI more resource-efficient

Practical applications

What this means for AI infrastructure

You can't manage what you don't measure: A scientific approach for organizations to get a handle on their tech value realization

A tale of two S&P 500 organizations: “Average” vs. “Ambitious.”

How do the two companies perform?

When tech decisions align, value compounds

Technical depth isn’t just a cost, but a growth constraint

Transformation moves at the speed of digital readiness—not technology

The future belongs to systems that work as one

From Mood Board to Move-In: AI’s Growing Role in Home Renovation

Why design and build it yourself?