How – Inside the AI OS: running governed agents on your own cluster, liveRegister →
Back to Insights

Enterprise AI Is Not a Bigger Copilot

The market defines enterprise AI as a scope. Scope is not a standard — what makes AI enterprise-grade is that a process can be trusted to it.

Nathan Bijnens

Ask three vendors what enterprise AI is and you get the same sentence back, more or less: artificial intelligence integrated into business operations at scale. As a description of where the technology sits in a company, it is accurate. As a definition it is close to useless, because it describes a scope — how much of the organisation the software touches — and scope is not a standard. You cannot procure against it. You cannot fail an audit against it. It tells a CIO nothing about whether the system in front of them survives contact with a process that has a regulator attached.

A more useful definition starts at the moment that actually matters: the point at which an organisation is willing to let software run part of a process rather than help someone do it faster. That is the bar, and it is why we treat enterprise AI as an engineering property rather than a market category. Four things have to be true before a process can be trusted to a machine:

What has to be true before a process can run itself
Grounded
Reasons over one model of the business — entities, relationships, state, rules — not over whatever documents happened to be in the index
Governed
Every agent has an identity, a policy that says what it may touch, and an audit trail that settles before the answer is returned
Orchestrated
Work crosses people, agents and systems in one flow, with hand-offs, approvals and the ability to stop
Sovereign
Runs where the data already lives, under your keys and your jurisdiction, including where there is no network at all

Nothing in that list is a model capability. That is the point.

Why copilots plateau

A copilot is a good product, and the productivity is real. Someone drafts a reply in a third of the time, gets through a legal review before lunch, finally writes the tests. Every one of those wins is genuine, and every one of them is personal.

The trouble starts when you try to add them up. A thousand people each working slightly faster is a thousand private improvements; the process running between those thousand people takes exactly as long as it did before. When the tab closes, the context goes with it. The next person starts from nothing. The assistant knows what was pasted into it and nothing about the approval rules, the last time this case ran, or why the exception was granted in March.

So the plateau is structural, not a sign that the model needs replacing. It is the difference between personal AI and organisational AI, and it shows up in four places at once.

Personal AI — the copilotOrganisational AI
Unit of improvementOne person, one task at a timeOne process, end to end
ContextLives in a tab and dies with itHeld in the model of the business, across runs
AuthoritySuggests; a human retypes the result into the systemActs on the system of record, under a policy
EvidenceA chat log, if anyone kept itAn audit trail produced by running
What it changesHow fast someone worksHow long the process takes, and what it costs
A thousand private improvements do not add up to a faster process. The right-hand column is not a bigger version of the left one, and it does not arrive by buying more seats.

Most pilots do not fail on the model

The pattern is familiar enough to be boring. The demo works. Everyone is impressed. Then someone asks the four questions that decide whether it ships — and the pilot has no answers, because none of them are model problems.

Provenance
Where did that fact come from, and was it current? Answered by a model of the business, not by a larger model of language.
Identity
Whose identity did the agent act under? A named principal with permissions that follow the work, never a shared super-user.
Permission
What was it allowed to touch? Policy enforced where the workload cannot switch it off, rather than instructions in a prompt.
Evidence
What record exists afterwards? An audit trail produced as a by-product of running, not reconstructed once someone asks.

Grounding is a data problem before it is an AI problem. An ontology — the entities, relationships and rules your business actually runs on, mapped onto the systems that hold them — is what lets an agent retrieve meaning rather than plausible text, and lets it write back to an operational system without guessing at what a record means. It is unglamorous work and it is the work.

Governance is the second. An agent that acts on real systems needs the same treatment as a member of staff: a named identity, permissions that follow the work rather than a shared super-user, policy enforced where the workload cannot switch it off, and an audit trail produced as a by-product of running rather than reconstructed later. That is what AI governance has to mean in practice, and it is why bolting a review step onto a chatbot does not produce it.

Orchestration is the third, and it is the one most organisations discover last. Real processes are not one prompt; they are a sequence of steps across several systems with people in some of them. Something has to hold that sequence, resume it, escalate it, and know which human owns which decision. That runtime is what we call an AI OS — the layer between the models and the business, doing for agents roughly what an operating system does for programmes.

Swapping in a stronger model moves none of these. It is the platform around the model that decides whether the pilot becomes a process.

Take it with you

The four questions above, expanded into about thirty — and a score per dimension.

Enterprise AI to production: the checklist

About thirty questions to put to your own team, grouped under grounding, governance, orchestration and sovereignty. Tick what is true for you today and see a score per dimension. One page, yours to print.

We use it to answer, and to know which organisation is asking. Nothing else. We handle your details as described in our Privacy policy.

What European organisations have to design for

The American explainers on this topic are competent and they all stop at the same place: benefits, risks, a stack diagram, a product call to action. What is missing is the set of constraints that European organisations have to design for from the first architecture meeting, not retrofit before go-live.

What appliesWhat it puts on youWhat it changes in the architecture
EU AI ActObligations on the deployer, not only on the companies building models: classification, meaningful human control, logs, informing the people affected, use within the terms the system was placed on the market underKnow which of your systems are high-risk before anything is built, keep the human decision point in the flow, and retain the record by default
DORAAI services treated as ICT third-party dependenciesA register, resilience testing and an exit plan for every model and service the process leans on
NIS2An agent's access to operational systems is a security-governance question, with accountability that reaches managementAgent identity and permissions have to be auditable facts, not conventions in a runbook
Cloud and AI Development Act
Commission proposal, 3 June 2026 — not law yet
A cloud sovereignty framework with four assurance levels, from processing inside the Union up to demonstrated control over the whole supply chain, plus a joint procurement framework for public administrationsIf you are a public body or sell to one, the assurance level your architecture can honestly claim turns into a procurement criterion — and it is decided by where the workload runs and who controls it, not by a clause
Data that cannot leave the buildingA meaningful set of European workloads, with no network path outAir-gapped operation as a property of the whole system — including how models are updated and how evidence is exported
CADA still has the Parliament and the Council in front of it, so treat that row as direction rather than obligation. The rest already apply — and every one of them is cheap to design for and expensive to add afterwards.

Sovereignty is not a badge you apply at the end; it is a property of where the thing runs.

The honest part

Most organisations we meet have capable people, processes that are broken in ways everyone can describe, and data they cannot use. Our position on that has not changed: We fix two out of three. Deliberately.

That is a scope statement, not modesty. The processes and the data are engineering problems, and we will take them: the ontology gets built and owned, agents get identities and policies, the audit trail becomes a by-product of running. The third one stays with you. Who owns which process, which decisions a human keeps, what Human+AI actually means in your operating model — none of that ships in a licence, and any vendor implying otherwise is selling you the pilot rather than the production system.

The cost is real and worth stating plainly. This is a platform decision with a data project inside it, and the first year buys fewer, deeper things than a copilot rollout does. What it buys instead is a process that compounds.

If this is the argument on your desk right now, we are running it live on 13 October: From personal AI to organisational AI: why copilots plateau, and what an AI operating system changes — 45 minutes, slides and a working system, and the organisational side as well as the technical one. If you would rather start on your own, the checklist above turns the four properties into the questions to put to your own team — tick what is already true and it scores each dimension — before anyone writes a business case.

Newsletter

Get the next one in your inbox

If this was worth your time, the next one will be too. We write when something changed in European AI sovereignty, or when we learned something building for it. A few times a year, no drip campaign.

  • Analysis, not announcements
  • Field notes from defence, government and finance deployments
  • First seat at the live briefings

A few times a year. No drip campaign, unsubscribe in one click. Privacy policy

Rather talk it through? Book a 30-minute briefing