At 6:47 a.m., an AI system notices something small.

A supplier has changed the currency used for its invoices.

It doesn’t look important.

But if nobody catches it, the change could create a $2.3 million FX mismatch over the next quarter.

This isn’t a hypothetical scenario.

J.P. Morgan has described an agentic AI workflow that can detect this kind of change, reclassify the exposure, propose a 90-day hedge, price it against counterparty quotes, and queue the trade for human approval.

And that’s where AI in finance gets interesting.

The biggest shift isn’t AI helping financial professionals make better decisions.

It’s AI starting to participate in the decisions themselves.

Finance Has Used AI for Years

Banks have been using AI and machine learning for years.

Fraud detection.

Credit risk.

Forecasting.

Customer service.

Market analysis.

But many of these systems follow a familiar pattern:

Data → Prediction → Human Decision

AI finds something.

A person decides what to do.

Agentic AI changes the equation.

The emerging model looks more like:

Data → Reason → Recommend → Act → Audit

The difference is one word:

Act.

An AI system doesn’t just tell you that something is wrong.

It can potentially start the workflow needed to address it.

And financial institutions are now figuring out exactly where that line should be drawn.

J.P. Morgan: From Detection to Action

J.P. Morgan’s treasury example shows what this transition can look like in practice.

The system notices a supplier has unexpectedly switched its invoicing currency.

Instead of simply flagging the change, the workflow can:

Detect the exposure.

Reclassify it.

Estimate the impact.

Propose a hedge.

Compare available quotes.

Queue the trade for human approval.

J.P. Morgan’s case study describes this as part of a broader move toward agentic workflows in corporate treasury.

That’s a meaningful change in the role of AI.

Traditional analytics might tell a treasury professional:

Something changed.

An agentic workflow can move closer to:

Something changed. Here’s why it matters. Here’s what I recommend. Here’s the action I’ve prepared.

The human is still in the loop.

But the human is no longer starting from zero.

That’s the transition from AI as an analyst to AI as an operator.

Fraud Is Becoming an AI Arms Race

There’s another area where this shift matters even more:

fraud.

AI gives attackers new ways to automate scams, generate convincing content, and adapt their tactics.

But financial institutions can use the same technology defensively.

J.P. Morgan describes using techniques including graph analysis and transformer models to identify relationships and patterns associated with fraudulent activity.

That’s important because sophisticated fraud rarely looks like one obviously suspicious transaction.

It looks like a pattern.

Sometimes it looks like a network.

And AI is increasingly useful at finding patterns inside networks.

So finance is entering a strange race:

AI vs. AI.

The attackers are becoming more automated.

The defenders are becoming more automated too.

The advantage may increasingly belong to whoever builds the better system around the model.

Goldman Sachs: The Problem Isn’t Just Intelligence

But there’s another problem that becomes much more important when AI starts operating inside financial systems:

What happens when the AI is wrong?

A chatbot giving you a bad restaurant recommendation is annoying.

A financial AI producing an incorrect answer about a transaction, risk exposure, or market data can be expensive.

And potentially dangerous.

In a recent discussion about building AI systems for capital markets, Goldman Sachs highlighted a fundamental challenge: institutional AI needs to be reliable, auditable, and grounded in verified sources.

The firm specifically points to auditable grounding as a way to reduce hallucinations and trace outputs back to their sources.

That’s an important distinction.

In finance, intelligence isn’t enough.

You need to know:

Where did this answer come from?

What data influenced it?

Why did the system make this recommendation?

Who approved the action?

And:

Can we reconstruct what happened afterward?

The more autonomy AI receives, the more important these questions become.

The New Bottleneck Isn’t Intelligence

This is the part I think is easy to miss.

We’re spending enormous amounts of time asking:

How intelligent will AI become?

Financial institutions may need to ask a different question:

How much authority should AI have?

Because there’s a huge difference between an AI that can:

Observe

and an AI that can:

Act.

Think about financial AI as an autonomy ladder:

Level 1: Observe

AI monitors financial data and identifies patterns.

Level 2: Recommend

AI suggests what a human should do.

Level 3: Prepare

AI prepares the analysis, transaction, or workflow.

Level 4: Execute With Approval

AI takes action after a human approves it.

Level 5: Autonomous Execution

AI acts independently within predefined limits.

We’re already seeing financial institutions experiment with agentic AI across areas such as client service, trading, wealth management, and corporate treasury.

Reuters reported in July 2026 that major Wall Street banks including Morgan Stanley, BNY, UBS, Goldman Sachs, J.P. Morgan, and Citi were increasing their use of AI agents, while executives continued to emphasize human oversight for high-stakes tasks.

And the trend isn’t limited to large Wall Street banks.

In July 2026, Reuters reported that Bank of Singapore, the private banking arm of OCBC, began rolling out an agentic AI platform for wealth-client onboarding.

The bank said the system could reduce onboarding time from more than 30 business days to 15, with AI helping streamline due diligence and credit-risk profiling.

The interesting question isn’t whether Level 5 is technically possible.

It’s:

Where should Level 5 be allowed?

You probably don’t want the same level of autonomy for:

A customer-service response.

A $10,000 payment.

A $10 million treasury transaction.

And a multi-billion-dollar trading decision.

AI autonomy needs boundaries.

The Future Won’t Look Like a Chatbot

When people imagine AI in finance, they often picture a chatbot sitting next to a banker.

Ask a question.

Get an answer.

That’s useful.

But it may not be the most important transformation.

The more interesting future is almost invisible.

AI agents embedded directly inside financial workflows.

Watching transactions.

Monitoring exposure.

Detecting anomalies.

Researching information.

Preparing decisions.

Communicating with other systems.

And, within predefined limits, taking action.

The interface may not even look like AI.

It may simply look like work getting done.

The Real Competitive Advantage

This leads to a bigger question about the future of financial institutions.

As access to increasingly capable models spreads, differentiation may shift toward everything built around them.

Better data.

Better workflows.

Better controls.

Better evaluation.

Better governance.

And most importantly:

Knowing exactly what AI should and shouldn’t be allowed to do.

That’s why the next competitive advantage in financial AI may not be model intelligence.

It may be controlled autonomy.

The banks that figure this out won’t just have AI assistants.

They’ll have something closer to an AI operating layer for finance.

And that’s a much bigger change.

One Question I’m Watching

How much financial decision-making would you trust an AI agent to handle without human approval?

Because the biggest question isn’t whether AI can make financial decisions.

It’s who gets to decide which decisions AI is allowed to make?

AI agents are everywhere right now.

But there’s a problem.

We keep talking about what agents could do.

Much less attention goes to what they can actually handle today.

An agent that can generate a plan is interesting.

An agent that can execute that plan across multiple systems, recover when something goes wrong, and know when to ask a human is much more useful.

That’s the difference between an AI demo and an AI worker.

So instead of another list of “AI agents you should try,” let’s look at where agents are actually becoming useful.

1. Research Agents

Research is one of the clearest early use cases for agents.

The traditional workflow is familiar:

Question → Search → Read → Compare → Write

Agentic research changes that to:

Define the objective → Let the agent investigate → Review the result

OpenAI’s Deep Research is a good example. It is designed for complex, multi-step research, searching, interpreting, and synthesizing information from many online sources into a detailed report with citations.

The important shift isn’t that AI can search.

We’ve had search for years.

The shift is that the system can increasingly manage the research process itself.

But that doesn’t mean the human disappears.

The agent can collect and synthesize information.

The human still needs to decide whether the evidence is good enough to trust.

AI Wide Open Reality Score

Autonomy: 7/10

Production readiness: High

This score is our editorial assessment, not an industry benchmark.

Where it shines: Well-defined research questions with lots of information to process.

Where it breaks: Source quality, conflicting evidence, and questions where the answer depends heavily on judgment.

Verdict: One of the strongest early agents use cases because it removes a large amount of mechanical research work without requiring full autonomy.

2. Coding Agents

Coding is where the shift from AI assistant → AI worker becomes easiest to see.

Consider a traditional coding assistant.

You ask it to write a function.

It gives you code.

You review it.

An agent can take a much larger objective:

Fix this bug.

Then inspect the repository, identify relevant files, make changes, run tests, interpret failures, and iterate.

OpenAI’s Codex is built around this kind of workflow. It can write features, answer questions about a codebase, fix bugs, and propose pull requests for review.

There is an even more extreme example.

In an OpenAI engineering experiment, a team built and shipped an internal software product with zero lines of manually written code. The application logic, tests, CI configuration, documentation, observability, and internal tooling were all written by Codex. OpenAI estimated the product was built in about one-tenth the time it would have taken to write the code by hand.

That’s the important shift.

The question isn’t:

Can AI write code?

It clearly can.

The question is:

How much of the software development loop can an agent own?

AI Wide Open Reality Score

Autonomy: 8/10

Production readiness: High

Editorial assessment, not an industry benchmark.

The limitation isn’t usually generating code.

It’s understanding context.

An agent may produce a technically valid solution that is completely wrong for the architecture, product requirement, or business constraint.

The code compiles.

The system still shouldn’t ship it.

Verdict: Probably the most mature example of AI agents doing meaningful professional work today.

3. Computer-Use Agents

This category is different.

Instead of interacting with an API, database, or codebase, the agent interacts with software the way a human does.

Click.

Type.

Navigate.

Read.

Choose.

Repeat.

Anthropos’s computer-use announcement describes Claude’s ability to interact with a computer through screenshots, mouse movement, clicks, and keyboard input.

That matters because an enormous amount of business software is still designed for humans to operate directly.

Think about:

Open website → Find information → Copy data → Fill form → Submit → Check result

For years, humans have been the glue between these systems.

Computer-use agents could become that glue.

But this is also where the gap between capability and reliability becomes obvious.

A human sees a redesigned webpage and adapts.

An agent may see a completely different interface and fail.

A human notices that a form is asking for something unexpected.

An agent may continue following the original plan.

Anthropic itself describes current computer use as still imperfect, slow, and error-prone, and recommends starting with lower-risk tasks.

That’s why I see computer use as powerful, but still relatively fragile.

AI Wide Open Reality Score

Autonomy: 5/10

Production readiness: Emerging

Editorial assessment, not an industry benchmark.

Best fit: Predictable, repetitive workflows.

Biggest risk: Unexpected interfaces and actions where one mistake has real consequences.

Verdict: One of the biggest opportunities in agentic AI, but not yet something I’d trust with unlimited authority.

4. Business Workflow Agents

This is where agents become much more interesting commercially.

Consider a customer-support request.

An agent could:

Read the message.

Look up the customer’s account.

Check the relevant policy.

Search the company’s knowledge base.

Decide what action is appropriate.

Respond.

Escalate if necessary.

That’s no longer:

AI answering a question.

It’s:

AI executing a workflow.

And we already have measurable examples.

Anthropic’s Kodif case study reports that Kodif uses Claude-powered AI agents to automate 65% of Dollar Shave Club’s support tickets and achieve 90% automation for Trust Wallet’s crypto support.

Those numbers matter more than the label “AI agent.”

Because the useful question isn’t:

“How intelligent is the agent?”

It’s:

“How much of the workflow can it actually take over?”

But there’s a second question companies have to answer:

“What is the agent allowed to do?”

Should it issue a refund?

Change an account?

Approve an invoice?

Modify a database?

The more consequential the action, the more important the permission layer becomes.

AI Wide Open Reality Score

Autonomy: 6/10

Production readiness: High for structured workflows

Editorial assessment, not an industry benchmark.

Verdict: This could become one of the largest enterprise applications of agents, but only if companies solve permission, monitoring, and accountability.

5. Long-Horizon Agents

This is where I think the most interesting shift is happening.

Give an agent a goal.

Let it figure out the steps.

Let it execute.

Let it recover.

Then come back later.

That’s a very different model of software.

And we’re beginning to see users delegate work that would previously have required hours of human attention.

OpenAI’s June 2026 analysis of Codex usage found that by May 2026, 70.2% of sampled individual users had made at least one Codex request estimated to correspond to more than one hour of human work.

Even more striking, 25.6% had made at least one request estimated to correspond to more than eight hours of human work.

The important signal isn’t the exact number.

It’s the direction.

People are starting to delegate longer-horizon work to AI.

But longer tasks introduce a nasty reliability problem.

If an agent has a 95% chance of getting each individual step right, a 20-step workflow doesn’t have a 95% chance of succeeding end-to-end.

Reliability compounds in the wrong direction.

So autonomy isn’t binary.

It’s a spectrum:

One step → Multiple steps → Recovery → Long-running execution

The real breakthrough will come when agents can move further along that spectrum without becoming less reliable.

AI Wide Open Reality Score

Autonomy: 5/10

Production readiness: Emerging

Editorial assessment, not an industry benchmark.

My take

This is probably the area I’ll be watching most closely.

Not because today’s long-horizon agents are already reliable enough for everything.

They aren’t.

But because this is where the economic value of agents could change dramatically.

An assistant that saves you ten minutes is useful.

An agent that can own a four-hour workflow is something else entirely.

The AI Agent Reality Score

Looking across these use cases reveals something important.

The smartest agent isn’t necessarily the most useful one.

What matters is whether it can reliably own a piece of work.

Agent category Autonomy Production

Research 7/10 High

Coding 8/10 High

Computer Use 5/10 Emerging

Business Workflows 6/10 High for structured workflows

Long-Horizon 5/10 Emerging

These scores are AI Wide Open’s editorial assessment, based on the capabilities and real-world evidence discussed above.

They are not standardized industry benchmarks.

But the framework behind them is simple.

When I evaluate an agent, I look at five things:

1. Task Completion

Can it actually finish the job?

2. Reliability

How often does it succeed without intervention?

3. Recovery

What happens when something goes wrong?

4. Permission

What is it actually allowed to do?

5. Escalation

Does it know when a human needs to take over?

And the last two may matter more than intelligence itself.

A highly capable agent with the wrong permissions can be more dangerous than a less capable agent with tightly controlled authority.

Likewise, an agent that cannot recognize when it is out of its depth isn’t truly autonomous.

It’s simply unsupervised.

The Real Shift

This is why I don’t think the future of AI agents is simply:

Humans → replaced by agents

It’s more likely:

Humans → set goals and boundaries → agents execute → humans handle exceptions

The human role doesn’t necessarily disappear.

It moves up the stack.

Humans decide:

What should happen?

What is the agent allowed to do?

What requires approval?

What happens when confidence is low?

The agent handles:

How do we execute it?

That’s a much more interesting future than simply asking whether AI will replace jobs.

The Bottom Line

AI agents are no longer just a research concept.

They’re already performing meaningful work across research, coding, computer use, and business workflows.

But we’re still early.

The biggest bottleneck may not be model intelligence anymore.

It may be reliability, permissions, and trust.

The companies that solve those problems won’t just build better agents.

They’ll build the systems that allow agents to become part of how real work gets done.

And that may be the real AI agent opportunity:

Not building agents that can do everything.

Building agents that can reliably do the right things.

The next question is where these capabilities actually create value.

Next Tuesday, I’m taking the same lens into finance:

Where is AI already changing financial work, what’s actually working, and where is the hype getting ahead of reality?

More AI beyond the demos, less AI hype.

For months, Mike Busch knew something wasn’t right.

He felt pressure in his chest.

He had balance problems.

He wasn’t feeling like himself.

At 77, he finally went to Mayo Clinic.

His doctor ordered a routine ECG.

A 10-second test.

But hidden inside that test was something doctors couldn’t easily see.

An AI model flagged a strong signal for cardiac amyloidosis.

Its score?

98 out of 100.

Further evaluation confirmed the diagnosis.

The AI didn’t diagnose Mike on its own.

It did something more practical.

It pointed the doctors toward something they might otherwise have missed.

That may be one of the most important ways AI changes healthcare.

Not by replacing the doctor.

By helping the doctor see more.

Healthcare Has an Attention Problem

Doctors already have access to enormous amounts of information.

The problem is that no human can process all of it at once.

One patient can generate:

• Medical history

• Lab results

• Imaging

• ECGs

• Medications

• Clinical notes

• Previous diagnoses

• Symptoms across multiple visits

The data keeps growing.

Human attention doesn’t.

That’s where AI becomes interesting.

It can search through large amounts of information, recognize patterns and surface signals that deserve a closer look.

Not:

“Here is the answer.”

But:

“Look here.”

That difference matters.

Because the final decision still belongs to the clinician.

AI doesn’t necessarily need to become the doctor.

Sometimes, it just needs to become a better second set of eyes.

Then There Is the Work Nobody Sees

A patient finishes an appointment.

The conversation ends.

The doctor still isn’t done.

Now comes the documentation.

The conversation has to become a clinical note.

The relevant details have to be organized.

The medical record has to be updated.

The doctor moves from patient care to paperwork.

AI is starting to change that workflow too.

At Austin Regional Clinic, Suki reported a 97% engagement rate among onboarded clinicians.

ARC also reported an 18.5% reduction in documentation time per patient encounter.

Source: Suki / Austin Regional Clinic

That sounds like a productivity metric.

It isn’t.

In healthcare, time is part of the care.

Every minute a doctor doesn’t spend typing can potentially go toward listening, explaining, examining or thinking.

And that’s the bigger opportunity.

AI doesn’t only have to improve the medical decision.

It can improve everything that surrounds the decision.

The Best Healthcare AI May Be the AI Patients Barely Notice

The popular version of the AI healthcare story is easy to imagine:

AI diagnoses the patient.

AI recommends the treatment.

AI replaces the doctor.

But some of the most useful systems being deployed today are much quieter.

They:

• Listen.

• Summarize.

• Search.

• Flag.

• Organize.

• Surface possibilities.

• Reduce the work around the work.

And that may be where some of the biggest gains come from.

The best AI systems may not be the ones making the biggest decisions.

They may be the ones quietly making humans better at making them.

Healthcare doesn’t necessarily need more decisions made by machines.

It needs better decisions made by humans.

But There Is a Line AI Shouldn’t Cross

Healthcare is different from most AI applications.

A bad recommendation isn’t just an inconvenience.

It can affect someone’s health.

So the goal shouldn’t be to remove humans from the loop.

It should be to make the human in the loop more capable.

AI can flag an unusual signal.

The clinician investigates.

AI can draft the note.

The clinician reviews it.

AI can surface a rare possibility.

The medical team decides what it means.

That distinction is critical.

AI provides leverage.

Humans remain accountable.

The Real Opportunity

The question is no longer whether AI belongs in healthcare.

It already does.

The bigger question is what happens when it becomes part of the everyday workflow.

The future of healthcare may not be:

Doctors vs. AI.

It may be:

Doctors with AI
vs.
Doctors without it.

The difference won’t necessarily be a machine making decisions for one doctor.

It may be one doctor having:

• More information at the right moment.

• Less paperwork.

• More patterns surfaced.

• More time with patients.

• More attention for the difficult cases.

That is a very different vision of AI in healthcare.

Not automation for the sake of automation.

Augmentation.

Not removing the human.

Increasing what the human can see, process and act on.

The biggest impact of AI in healthcare may not come from replacing the person in the room.

It may come from giving that person back something healthcare has been taking away for years:

Time.

Attention.

And the ability to focus on the patient instead of everything around the patient.

The most interesting healthcare AI may not be the system that says:

“I know what’s wrong.”

It may be the system that says:

“Look here.”

And then gives the doctor more time to figure out what it means.

That might be the healthcare AI story worth watching.

And healthcare is only one part of the story.

On Thursday, I’ll look at AI agents doing real work and what they can actually handle today.

AI agents are moving beyond chat.

They can plan tasks, use tools, access information, interact with software, and coordinate multiple steps to accomplish a goal.

But an AI agent is not just a powerful model.

The Stack at a Glance

A simple way to think about the stack is:

Model → Agent → Tools → Data → Other Agents → Human Oversight

The model provides the intelligence.

The agent determines what to do.

Tools allow it to take action.

Data gives it context.

Protocols help it connect with other systems and agents.

Human oversight provides control when the stakes are high.

Here are some of the projects and technologies worth knowing if you want to understand where this new software stack is heading.

1. The Builders

The first layer is where developers actually build agents.

LangGraph

LangGraph is a low-level orchestration framework for building stateful, long-running agents and workflows.

It gives developers control over how an agent moves through a task, including persistence, human-in-the-loop interaction, and durable execution.

That makes it useful for agents that need to operate across multiple steps rather than simply answer a question.

Why it matters:

Agents need more than intelligence.

They need a way to manage complex workflows reliably.

CrewAI

CrewAI focuses on multi-agent systems where different agents can take different roles and collaborate on a larger task.

One agent might research. Another might analyze.

Another might produce the final result.

The idea is similar to building a team of specialized AI workers.

Why it matters:

As tasks become more complex, developers are exploring whether multiple specialized agents can outperform a single general-purpose agent.

But more agents do not automatically mean better results.

The real question is whether the additional coordination creates enough value to justify the complexity.

Microsoft Agent Framework

Microsoft Agent Framework is an open framework for building,

orchestrating, and deploying AI agents and multi-agent workflows across Python and .NET.

It is designed for teams moving agents from prototypes toward production, with support for orchestration, observability, governance, and human-in-the-loop control.

Why it matters:

The industry is moving beyond experimental agent demos toward systems that need to operate inside real organizations.

That means reliability and governance become part of the architecture, not something added later.

OpenAI Agents SDK

The OpenAI Agents SDK provides building blocks for creating agent applications with tools, handoffs, guardrails, and other runtime behavior.

It also includes tracing capabilities for understanding and debugging agentic workflows.

Why it matters:

AI platforms are increasingly providing infrastructure for systems that can take actions, not just generate content.

The interface between AI models and software is becoming just as important as the model itself.

Google Agent Development Kit

Google’s Agent Development Kit, or ADK, is designed to help developers build and orchestrate AI agents.

Its importance goes beyond the framework itself.

Google, like other major AI companies, is competing to become part of the infrastructure behind the next generation of AI applications.

Why it matters:

The competition is expanding beyond the model layer.

The agent-development layer is becoming a strategic part of the AI stack.

2. The Connectors

Agents become much more useful when they can interact with the systems around them.

That is where protocols become important.

Model Context Protocol (MCP)

The Model Context Protocol provides a standardized way for AI applications to connect models with external tools, data, and services. Instead of building a custom connection for every integration, developers can use a common protocol.

This may sound like an infrastructure detail.

It isn’t.

If agents are going to become useful at work, they need standardized ways to interact with the systems people already use.

Why it matters:

An intelligent agent that cannot access the right information or tools is still limited.

Agent2Agent (A2A)

Agent2Agent, or A2A, focuses on communication and interoperability between independent AI agents.

It allows agents built with different frameworks or by different organizations to discover capabilities and collaborate without exposing their internal state or implementation.

The bigger idea is interesting:

The future may not be one giant AI agent.

It could be networks of specialized agents working together.

Why it matters:

As AI systems become more specialized, communication between agents could become as important as communication between applications is today.

A simple way to think about it

MCP = Agent ↔ Tools, Data, and Services

A2A = Agent ↔ Agent

MCP helps agents interact with the world around them.

A2A helps agents interact with each other.

3. The Reliability Layer

Building an agent that works in a demo is one thing.

Building one that can safely operate inside a company is another.

That creates another part of the stack focused on reliability, evaluation, observability, and control.

PydanticAI

PydanticAI is a Python framework for building AI applications and agents with an emphasis on structured outputs, validation, and production-oriented development.

Its ecosystem also includes capabilities for tools, MCP, guardrails, code execution, file access, and sub-agent orchestration.

Why it matters:

Production AI systems need predictable interfaces and software engineering discipline, not just impressive responses.

Evaluation and Observability

As agents become more autonomous, developers need to understand what they are doing.

Did the agent choose the right tool?

Did it follow the correct workflow?

Did it make a wrong decision?

How often does it fail?

These questions become much harder when an agent takes dozens of actions instead of generating one response.

That makes evaluation and observability an increasingly important part of the agent stack.

Why it matters:

You cannot reliably deploy an AI system if you cannot measure how it behaves.

Capability ≠ Reliability

A capable agent is not necessarily a reliable agent.

The more actions an agent can take, the more important it becomes to evaluate, observe, and control its behavior.

4. The Human Layer

The final layer may be the most important one.

Not every decision should be made autonomously.

An agent might be allowed to draft an email automatically.

But sending it may require approval.

It might analyze a transaction.

But approving the transaction may require a human.

It might modify code.

But deploying that code may require a review.

This is where permissions, guardrails, security, and human-in-the-loop systems become critical.

The more capable the agent becomes, the more carefully we need to define what it is allowed to do.

Why it matters:

The difficult question is no longer only:

What can the AI do?

It is also:

What should we allow it to do?

Putting the Stack Together

The different technologies in the agent ecosystem are not isolated pieces.

They are beginning to form layers of a broader software stack.

Models provide intelligence.

Agent frameworks provide orchestration.

Tools and data give agents access to the world around them.

Protocols allow agents to connect with tools, services, and other agents.

Evaluation and observability help developers understand and measure behavior.

Security and human oversight provide control.

Together, these layers create the infrastructure that allows AI systems to move from generating responses to participating in real workflows.

What the Stack Tells Us

The interesting development is not that there are suddenly dozens of AI agent frameworks.

It is that a new software stack is beginning to form around AI.

For years, software was primarily designed around humans interacting with applications.

Now we are beginning to see software designed around AI systems interacting with other software.

That changes the architecture.

Models become one layer.

Agent frameworks provide orchestration.

Protocols connect agents to tools and data.

Evaluation measures performance.

Observability helps teams understand behavior.

Human oversight provides control.

Together, these pieces create something much bigger than a chatbot.

They create systems that can participate in work.

The Big Shift

The shift is not simply from chatbots to agents.

It is from software that humans operate to software that AI systems can increasingly operate themselves.

That means the model is only one part of the equation.

The surrounding infrastructure determines how useful, reliable, and controllable that intelligence becomes.

What to Watch Next

The AI agent race may increasingly move away from:

Who has the smartest model?

and toward:

Who can build the most reliable system around the model?

That means watching:

• Agent orchestration

• Tool use

• Memory

• Multi-agent collaboration

• Evaluation

• Observability

• Security

• Permissions

• Human-in-the-loop systems

• Standards for agent communication

The models may be the brains.

The stack is what turns that intelligence into action.

For decades, factories have followed a simple model:

Machines do the work.
Humans make the decisions.

A machine produces a part.
A camera checks it.
A technician fixes it.
A manager looks at the numbers.

The system can be highly automated without being truly intelligent.

AI is beginning to change that.

Not simply by adding another software layer to the factory floor, but by giving industrial systems new capabilities:

They can increasingly see what is happening, recognize patterns, predict problems, and help people decide what to do next.

That shift matters because manufacturing has something most software businesses don’t:

the physical world.

A wrong decision can stop a production line.
A missed defect can reach a customer.
A machine failure can disrupt an entire operation.

So, the real story of industrial AI isn’t about putting a chatbot in a factory.

It’s about making the factory itself more aware.

First, AI Gives the Factory Eyes

Modern factories already generate enormous amounts of data.

Machines produce operational data.

Sensors measure temperature, vibration, pressure, speed, and energy consumption.

Cameras capture images.

Production systems track processes and output.

Maintenance teams record failures and repairs.

The challenge is turning all of this information into useful decisions.

A human operator may recognize that a machine is behaving differently. A quality inspector may notice a defect. An engineer may identify a pattern in maintenance data.

But people cannot continuously monitor every machine, component, and data stream.

This is where AI becomes useful.

Not because humans suddenly become unnecessary.

Because AI can help them see more of what is happening, faster.

BMW Group is a good example

BMW Group Plant Regensburg uses AI-assisted quality inspection as part of its GenAI4Q pilot. Photo: BMW Group.

At BMW Group Plant Regensburg, the company has been developing an AI system called GenAI4Q to create customized quality inspections for individual vehicles.

The plant produces around 1,400 vehicles each workday, and vehicles can differ significantly in their configurations. BMW says its AI system delivers tailored inspection recommendations for each vehicle based on its individual production context.

The AI isn’t simply looking at a picture and saying “good” or “bad.”

It is using production context to help determine what should be checked and when.

The factory is gaining something closer to situational awareness.

And BMW has been working on the data problem behind these systems as well.

In 2022, BMW published SORDI, the Synthetic Object Recognition Dataset for Industries. The dataset contains more than 800,000 photorealistic synthetic images across 80 categories of production resources. BMW says it was created to accelerate the training of AI models for manufacturing applications.

Why does that matter?

Manufacturers don’t always have enough real-world examples to train AI systems, especially when the situation they are trying to detect is rare.

Synthetic data can help fill that gap.

Instead of waiting for thousands of real examples, manufacturers can create simulated environments and generate labeled data.

The result is a new capability:

The factory can start to see.

But seeing is only the first step.

Then AI Starts Predicting

A factory doesn’t want to know that a machine has failed.

It wants to know that something is changing before the failure becomes expensive.

This is the idea behind predictive maintenance.

Machines generate signals as their operating conditions change. AI and machine-learning systems can analyze those signals and identify patterns that may indicate emerging problems.

Siemens Senseye Predictive Maintenance is designed around this approach, using machine learning and condition monitoring to help maintenance teams identify potential issues and prioritize where attention is needed.

The workflow starts to look different:

Traditional:
Machine fails → production is disrupted → technician investigates.

Predictive:
Machine data changes → AI identifies a potential issue → team investigates → maintenance can be planned.

That doesn’t mean AI can predict every failure perfectly.

It means manufacturers can use data to make maintenance decisions earlier and with more information.

And that matters because downtime can affect production schedules, workers, inventory, logistics, and customer deliveries.

The value isn’t simply that “AI predicts failure.”

It’s that AI can help people move from reacting to problems toward anticipating them.

The Factory Becomes a Digital Model

The next step is even more interesting.

What if you could test changes to a factory before making those changes in the physical world?

That’s where digital twins come in.

A digital twin is a virtual representation of a physical product, process, machine, or system that can be used to simulate and optimize its real-world counterpart.

BMW uses digital twins and virtual factory planning to simulate and optimize production systems before changes are made in the physical world. Photo: BMW Group.

BMW has already applied this idea to factory planning.

For its Debrecen plant in Hungary, BMW used virtual production planning well before series production began.

Using NVIDIA Omniverse, BMW created real-time digital-twin simulations to evaluate factory layouts, robotics, and logistics systems virtually.

Instead of:

Build → test → discover the problem → rebuild

the process can increasingly become:

Model → simulate → optimize → build

The physical factory still has to work.

But more experimentation can happen before the physical system is changed.

Digital twins create a bridge between the digital and physical worlds.

Then AI Gets a New Interface

AI is also changing how workers interact with industrial systems.

Consider a maintenance engineer dealing with a machine problem.

Traditionally, finding an answer could involve manuals, documentation, maintenance records, software interfaces, and specialists.

Generative AI introduces a different interface:

Ask the system.

Siemens Industrial Copilot is designed to support industrial workers and engineers with tasks including troubleshooting, machine instructions, code generation, and access to industrial information.

In March 2025, Siemens also announced an expanded generative-AI maintenance offering. Siemens reported that initial pilot use cases showed an average 25% reduction in reactive maintenance time. That is a Siemens-reported pilot result, not a universal industry benchmark.

That distinction matters.

Industrial AI should not be judged by impressive demos.

It should be judged by what happens inside real operations.

Does it reduce diagnostic time?

Does it improve quality?

Does it reduce unnecessary downtime?

Does it help an experienced worker make a better decision?

Those are the metrics that matter.

From Automation to Intelligence

Factories have been automated for decades.

Automation itself is not new.

A traditional automated system might follow a predefined rule:

If X happens, do Y.

AI introduces a different capability.

It can work with patterns, uncertainty, large amounts of data, and changing conditions.

That creates a progression:

Automation
Follow predefined instructions.

AI vision
Recognize what is happening.

Predictive AI
Identify patterns that may signal what happens next.

AI copilots
Help people understand information and decide what to do.

AI agents and automation
Potentially coordinate actions across connected industrial systems.

The important part isn’t any single technology.

It’s what happens when they connect.

The Factory That Can See, Understand, and Act

Imagine a production environment where these systems work together.

A camera detects an anomaly.

The system identifies the affected component.

Production data provides context.

Machine data shows that the equipment responsible has also started behaving differently.

Predictive maintenance software flags a potential issue.

An AI copilot helps the technician understand what may be happening.

The system checks maintenance records and available resources.

A human reviews the recommendation.

The maintenance workflow is initiated.

The machine is serviced.

Production continues.

This is not a fully autonomous factory today.

But the individual pieces already exist in real industrial deployments, while the broader closed-loop vision is developing.

In May 2025, Siemens announced that it was developing industrial AI agents designed to connect different copilots and automate workflows across the industrial value chain. Siemens described this as a move from assistants that respond to queries toward agents that can proactively execute processes.

That points toward a larger operational loop:

See → Understand → Predict → Decide → Act → Learn

The closer these capabilities become to one connected loop, the more fundamental the change becomes.

Intelligence Comes with a Catch

There is a reason industrial AI will not move at the same speed as consumer AI.

The consequences are different.

A chatbot can produce a bad answer.

A manufacturing system can make a bad decision that affects a machine, a product, a worker, or an entire production line.

That means the hardest part of industrial AI may not be the model.

It may be everything around the model:

Data quality.
System integration.
Security.
Reliability.
Human oversight.
Clear escalation paths.
Trust.

A model that works in a controlled demonstration still has to operate inside a messy physical environment where conditions change and failures have real consequences.

That’s why the path toward autonomous manufacturing is likely to be incremental.

The goal isn’t to remove humans from every decision.

It’s to give humans better information, better tools, and eventually better automated systems for decisions that can safely be delegated.

What Actually Matters?

The future of manufacturing isn’t simply about putting more AI tools into factories.

The deeper shift is that factories are becoming increasingly observable, connected, and adaptive.

AI can help them see.

Predictive systems can help them anticipate.

Digital twins can help them simulate.

Copilots can help workers understand and act on information.

And emerging AI-agent systems may eventually connect these capabilities into larger operational workflows.

The question is no longer:

“Will manufacturing use AI?”

It already does.

The more interesting question is:

How much of the manufacturing system can become intelligent while keeping humans in control of the decisions that matter most?

That is the shift worth watching.

Because the factory of the future may not simply be a place where machines work faster.

It may be a system that can increasingly:

See. Understand. Predict. Decide. Act.

And learn from what happens next.

AI Wide Open

There is a lot happening in AI.

New models. New tools. New announcements. New predictions.

But not all of them deserve your attention.

Every Tuesday, AI in the Real World looks beyond the hype to understand how AI is actually changing an industry, a business, or a real-world process.

Because the goal isn’t to cover everything happening in AI.

It’s to help you understand what is worth paying attention to.

AI is moving faster than most of us can keep up with.

Every week brings a new model.

Every day brings a new tool.

Every headline claims something has changed forever.

And somewhere between all the launches, demos, updates, and predictions, it is becoming harder to answer one simple question:

What actually matters?

Because not every new model is a breakthrough.

Not every AI tool will change the way we work.

And not every headline deserves our attention.

But beneath all that noise, something much more interesting is happening.

From Experiment to Reality

AI is moving from something we experiment with to something we build into the systems, decisions, and workflows that shape the real world.

A manufacturer using AI to predict equipment failures.

A healthcare company using AI to support earlier diagnosis.

A financial institution using AI agents to automate complex processes.

These aren't just interesting experiments.

They are signs of where AI is heading.

Welcome to AI Wide Open

AI Wide Open takes a wider look at artificial intelligence.

Not just the latest model.

Not just the newest tool.

Not just the biggest announcement.

We'll explore the tools, ideas, applications, and shifts shaping the world of AI.

What You'll Find Here

AI Tools & Technologies→

New tools, models, and capabilities and what they actually enable.

AI Across Industries→

How healthcare, manufacturing, finance, retail, transportation, and other industries are putting AI to work.

Business & Startups→

How AI is creating new opportunities, changing business models, and reshaping competition.

Agents & Automation→

How AI is moving from answering questions to taking action.

Trends That Matter→

Separating meaningful shifts from short-lived hype.

Real-World Applications→

Looking beyond what AI promises to what people and organizations are actually doing with it.

The goal isn't to cover everything happening in AI.

It's to help you understand what is worth paying attention to.

There is a lot happening in AI.
Let's focus on what matters.