Stop Fighting the Wrong AI War

The argument nobody's having

While everyone debates model quality, almost nobody asks the more consequential question: what happens to the intelligence an organization generates every single day? The judgment calls, the deal patterns, the operational fixes, the industry-specific knowledge that lives in a company's best people.

For most companies, the honest answer is: nothing happens to it. It doesn't get captured. It doesn't compound. It doesn't make the next interaction smarter than the last one. It evaporates into meeting notes, Slack threads, and exit interviews.

McKinsey's 2025 State of AI report puts a number on this gap. Nearly nine in ten organisations now use AI regularly — but most haven't embedded it deeply enough into workflows to realise enterprise-level value. Only 39% of respondents reported enterprise-level EBIT impact from their AI investment. Knowledge management has become one of the most common AI use cases, which confirms the intent is there — organisations are trying to activate their internal knowledge. They just haven't operationalised it yet.

"That's the real competitive gap. Not which model a company is renting this quarter, but whether it's building anything that's actually its own."

Why the two "obvious" answers are both wrong

Companies that do notice this gap tend to reach for one of two moves, and both tend to be worse than doing nothing deliberately.

Buy the public model and prompt harder. This treats the symptom, not the disease. Every session starts from zero context. And the exposure is real and measurable: research covered by TechRadar found 55% of UK employees use unapproved AI tools at work, with 10% knowingly sharing sensitive company data. A separate LayerX study found 45% of enterprise employees use generative AI tools, 77% of them paste data directly into those tools, and 22% paste PII or payment-card information. That's not a hypothetical governance gap. That's an organisation's context, judgment, and IP leaving the building every day, uncounted.

Build a private model from scratch. This usually trades the ownership problem for a worse one — falling behind the frontier. Most organisations underestimate the infrastructure, the retraining cadence, and the ongoing cost of staying current. It's part of why enterprise language in 2026 has shifted from "build your own model" toward sovereignty, model portability, and control over data placement — owning the intelligence layer without owning the burden of chasing frontier capability alone.

There's a third failure mode hiding behind both of these: letting individual teams build their own agents and workflows with no central orchestration. Forrester, reported by ITPro in June 2026, found three-quarters of enterprise leaders are adopting agentic AI — but only a small minority are operationally beyond "agentish" chatbots. The gap isn't ambition. It's orchestration, control, and trust.

What actually moves the needle

McKinsey's own review of more than fifty agentic AI builds found that value comes from redesigning workflows, not from deploying more agents — evaluation, step-level verification, reusable agents, human oversight, and common orchestration frameworks are what separate the programs that work from the ones that stall. BCG's research reaches a similar conclusion from a different angle: employees with strong strategic clarity report measurably greater impact than employees with tool access but no direction — a company can hand out every AI seat license it wants and still get nothing back without the operating model to match.

There's already a live example of what centralised, governed orchestration looks like in a regulated environment. In April 2026, Citi launched Arc, an internal platform described as a centralised operating system for AI agents — letting employees use frontier models inside one secure system, with the ability to monitor agent behaviour and stop tasks if needed. It's a real-world instance of exactly the shift this argument is describing: access to frontier capability, under an orchestration layer the institution controls.

The fight that actually matters

The model layer is converging. Every serious lab is racing toward the same capability ceiling, and the gap between the leading models shrinks every quarter. What won't converge is the depth and governance of the intelligence sitting underneath the model layer at each individual company. That's the layer that actually compounds — and it's the one almost nobody is building.

This isn't a call to abandon the tools teams already use well. It's a call to stop treating model selection as the strategy, and start treating an organisation's own knowledge as the infrastructure it actually is.

A company's competitors can subscribe to the same model it does. They cannot subscribe to what that company already knows. The fight worth having isn't over which AI is smartest this quarter — it's over who's still standing when the leaderboard resets for the tenth time.