The AI Operating Model

Most companies do not have an AI problem.

They have an operating model problem.

That is the part of this conversation that still gets skipped.

By now, nearly every leadership team can point to something with AI attached to it. A summarization tool. A chatbot inside Slack or Teams. A copilot embedded in the workflow. Maybe even a promising agent pilot.

On paper, that looks like progress.

In practice, a lot of it is still theater.

The hard part is not getting a model to produce a decent answer. The hard part is changing how work actually moves through the business. Until that happens, AI stays trapped as a layer on top of the old system instead of becoming part of a better one.

That is why so many organizations can demo AI and so few can scale it. NVIDIA's 2026 State of AI report puts it plainly: 64% of enterprises are now actively using AI, and among large companies that number climbs to 76%. But McKinsey still finds that roughly two-thirds have not scaled AI to enterprise levels. The gap between adoption and impact is where most organizations live right now.

The real bottleneck is not model quality

The common assumption is that enterprise AI stalls because the models are not good enough.

I think that is increasingly the wrong diagnosis.

Models are getting better. They are getting cheaper. Access is broad. Vendors are shipping new capabilities almost weekly.

And yet a huge number of companies are still stuck in the same place: lots of interest, lots of experimentation, very little durable operational change.

Why?

Because production AI asks a much harder question than pilot AI.

Pilot AI asks: Can this work?

Production AI asks: Who owns it, how is it governed, what systems does it touch, how is success measured, what happens when it is wrong, and what process changes because this capability now exists?

That is not a prompt engineering problem. That is an operating model problem.

The companies pulling ahead understand this. JPMorgan Chase is spending nearly $20 billion on technology in 2026 — and the money is not going to more pilots. They have doubled their generative AI applications in the past year, put AI tools in front of 150,000 employees, and are projecting $1.5 to $2 billion in annual AI-driven business value. Operations teams handle 6% more accounts per employee. Fraud costs per unit are down 11%. Software engineering productivity is up 10%. Those numbers do not come from experimentation. They come from rewiring how work actually gets done.

What scaled organizations do differently

The organizations that get past the pilot stage tend to share a few traits.

1. They redesign the workflow, not just the interface

A lot of AI projects are still wrappers.

Someone identifies a painful process, adds a chatbot or copilot to one step, and hopes the rest of the system somehow adapts around it. Usually it does not.

The best teams start in a different place. They ask a more useful question:

If this capability were reliable enough to use every day, how would we redesign the process around it?

That leads to better decisions.

Maybe the handoff disappears. Maybe the approval chain gets shorter. Maybe the agent handles intake, triage, and data gathering before a human steps in. Maybe the work is broken into smaller decisions so confidence thresholds are clearer.

Siemens proved this at their electronics factory in Amberg. When they introduced AI into X-ray quality assurance, the goal was modest — skip the QA step for 5% of products where inspection added no value. They hit 30%. But only because they let the shop floor teams redesign the workflow around the capability, not just bolt AI onto the existing process. As their head of factory digitalization put it: the barrier was not technical, it was cultural.

PepsiCo took the same approach with digital twins. Working with Siemens and NVIDIA, they converted US manufacturing and warehouse facilities into physics-level 3D simulations. AI agents now simulate and refine system changes before any physical modification occurs, catching roughly 90% of potential issues in the digital model first. The result: 20% throughput increase on initial deployments and 10 to 15% reduction in capital expenditure. That is not AI making the old process faster. That is AI enabling a fundamentally different process.

2. They assign real ownership

Every serious AI deployment needs an owner.

Not a steering committee. Not a vague executive sponsor. Not a cross-functional working group that meets every other Thursday.

An owner.

Someone who is accountable for the workflow, the result, the adoption, and the decision to expand, fix, or kill the effort.

This is where a lot of promising work dies. Pilots live in a gray zone. They are interesting enough to fund, but not important enough for anyone to truly own. So they linger.

Eni, the energy company, did not let that happen. They have deployed roughly 300 AI use cases across exploration, operations, and production — including a 35% reduction in drilling time through machine learning. That kind of breadth does not happen through committees. It happens because specific leaders own specific outcomes and have the mandate to move.

Similarly, o9 Solutions found that agentic AI in enterprise planning could reduce investigative time for supply chain root-cause analysis by as much as 80% — but only when the deployment was designed with deep enterprise knowledge baked in, not just bolted on. Their agents combine the scalability of large language models with the precision of structured decision models. That kind of architecture requires ownership at the intersection of business process and technology, not a handoff between the two.

Scaled organizations do not tolerate ambiguity about who is driving. They treat AI initiatives like products and operating capabilities. There is a named leader, a business metric, and a timeline.

If those things are missing, it is not a transformation initiative. It is an experiment.

3. They build governance into the workflow

Governance is still misunderstood in too many AI conversations.

It gets treated like a control function that shows up late and slows everybody down.

Good governance should do the opposite.

It should define the rules early enough that teams can actually move.

What decisions can an agent make on its own? What requires a human approval? What data can it access? What gets logged? What is the rollback path? How is performance monitored over time?

When those questions are answered up front, teams can scale with more speed and less drama.

When they are left unresolved, the organization creates friction at exactly the moment it should be trying to expand adoption.

JPMorgan is a good case study here too. Even as they scale AI to 150,000 users, they have bolstered model validation teams and embedded responsible ML checkpoints across pipelines. They are reducing operations headcount by 4% while growing client-facing roles by 4% — but the total headcount stays flat at roughly 318,000. That is governance enabling transformation rather than blocking it: clear rules about where AI replaces work, where it augments work, and how the workforce rebalances as a result.

Deutsche Bank and Goldman Sachs are now piloting agentic AI for autonomous trading desk surveillance. Deutsche Bank, working with Google Cloud, monitors staff communications across more than 40 channels and has already cut surveillance false positives by over 25%. That kind of deployment is only possible because governance was designed into the system from day one — not layered on after the fact.

That is why the most mature teams do not just build agent architectures. They build trust architectures.

4. They think in systems, not one-off use cases

One of the clearest differences between companies that scale and companies that stall is whether they treat each AI project as a custom exception.

If every use case requires new tooling, new approvals, new patterns, new integrations, and new debates about risk, the organization will never move fast enough.

The companies getting traction are building shared foundations:

common tooling patterns
reusable evaluation methods
standard security controls
shared connectors into systems of record
clear release and monitoring practices

ServiceNow is a good example of what this looks like at scale. Their platform orchestrates 80 billion workflows and 6.5 trillion transactions annually for 85% of the Fortune 500. AI is not a separate initiative — it is native to the infrastructure. The results their customers are seeing tell the story: AstraZeneca reclaimed 30,000 hours annually, Pure Storage resolves cases seven times faster, and Siemens handles 210,000 tickets autonomously every month.

In other words, they are building a delivery system, not chasing isolated wins.

That matters because the long-term advantage in AI will not come from one brilliant pilot. It will come from the ability to repeatedly turn useful ideas into governed production workflows.

The shift leaders need to make now

For CIOs and operators, the challenge is not deciding whether AI matters.

That part is over.

The real question is whether the organization is willing to make the operational changes that AI requires.

The numbers say the window is closing. Capgemini reports a decisive shift from pilots to platforms, with 38% of organizations now operationalizing AI use cases. Early movers are compounding their advantage — gains in productivity, faster cycle times, cost savings, new revenue streams. NVIDIA's data shows 88% of enterprises reporting AI-driven revenue increases, with 30% seeing gains above 10%. Companies like Klarna are projecting $40 million in annual profit improvement from a single AI assistant that handles 2.3 million customer conversations a month — work that previously required 700 agents.

That means moving the conversation away from novelty and toward execution:

Which workflows are worth redesigning?
Which leaders own them?
What are the success metrics?
What controls are required?
What reusable platform capabilities need to exist so the second and third deployment move faster than the first?

This is a leadership problem before it is a technology problem.

Because once the model is good enough, the bottleneck shifts to everything else: process design, governance, data access, change management, incentives, trust.

That is where most organizations are actually stuck.

A practical 90-day move

If I were advising a leadership team right now, I would keep the next 90 days simple.

Pick a small number of workflows that matter. Assign a real owner to each one. Define a measurable business outcome. Establish clear human-in-the-loop thresholds. Use shared architecture where possible. And force a decision at the end of the period: scale, redesign, or stop.

What I would not do is launch ten more pilots just to say the company is active in AI.

Activity is not progress.

The organizations that win this next phase will not be the ones with the most experiments. They will be the ones that build an operating model that can absorb AI into the real machinery of the business.

That is a much less glamorous story than product demos and launch headlines.

It is also the story that will matter.