AI-first does not start with the AI
The World Economic Forum has published a blueprint for AI-driven operating models. Four of its five building blocks assume something almost no company keeps: an organisation that can be read.
In 2026 the World Economic Forum published a paper that lands in board meetings: The AI-First Operating System. Not a trend report — a blueprint. It deserves to be taken seriously, precisely because in one place it assumes something it barely talks about.
What it says
The report separates two states. AI-enabled means AI makes individual tasks faster while the workflows stay as they were; take the AI away and the company carries on. AI-first means core workflows, roles and decision rights are rebuilt so that AI carries value creation at industrial scale.
Its central claim: the advantage does not come from individual copilots, and not from a collection of pilots, but from a loop in which every use becomes a learning signal. The paper describes five building blocks for that — the loop itself, an AI stack built to be swapped out, the redesign of operations, the organisation of human-AI teams, and the consequences for the business model.
Three points in it deserve particular attention:
- You own the control layers. Routing, policies, agent identity, permissions, auditing, data gateways — leaving all of that to the model provider is, the report argues, a mistake.
- The decision is per task, not per tool: AI acts on its own, AI assists, or AI stays out — where liability, strategic accountability or unmanageable risk are at stake.
- Trust is a product feature, not an appendix. Oversight should match the potential harm, uncertainty should be visible, every result verifiable.
The prerequisite that gets half a sentence
At one point the paper mentions an ontology: a machine-readable description of objects, rules, relationships and permitted actions. Half a sentence — and the rest hangs off it.
Because every step above assumes it. You can only redesign a workflow end to end if you have it in front of you as a structure. You can only decide whether a task may go to an AI if you know who answers for it today. You can only run people and AI as one team if you keep both in the same model.
That model is exactly what almost no company keeps. It has a database for its customers, one for its money and one for its code. For itself, it has slide decks.
So the bottleneck is not the AI. The bottleneck is that the organisation is not machine-readable.
What roleALPHA does here
We keep the operating model as one connected record: roles, value streams, objectives, risks, competences, policies, IT landscape — objects of a single model rather than chapters of a presentation. Kept current, queryable, and readable by people and by machines alike.
Within it, an AI agent is an organisation member. It receives no prompt but its place: the role it derives its remit from, the value stream it contributes to, the competences it uses, and the policies that bind it. Versioned, with provenance, retrievable over MCP — and every retrieval is logged.
For the three points above, that means:
Control layers. Every agent has its own identity with its own key, stored only as a verification value and revocable at any time. That access can do exactly one thing: read its own context. No write permission, no foreign data, and whatever the caller may not see does not appear in the bundle either. Anyone who has to evidence which rules were delivered to an agent at which point in time finds a trail rather than a recollection.
Graduated oversight. For each kind of object you can configure who has to agree before a change to the operating model takes effect — from “one named person decides” through “consent with a quorum” to “a single objection blocks it”. That is the graduated human oversight the paper calls for, only as a setting rather than an intention.
Limits. An agent is not an infinite employee. Throughput and budget sit side by side, with a note on which limit binds first. Where none is recorded, it says “no limit recorded” — no infinity sign and no filled bar. The difference between has no limit and we have not entered one is exactly the difference these projects turn on in the end.
What we do not do
We do not run the agents. There are runtimes enough for that, and that contest belongs to the model providers. We build the layer every runtime needs and none brings along — which is why we work with all of them. We ground agents; we do not orchestrate them.
And we do not call ourselves AI-first. The organisation model works without a language model at all, and the built-in assistant can be switched off and ships switched off. In a European company that is not reticence — it is a purchasing condition.
The question before the question
Anyone who reads the paper and comes away asking “where do we deploy agents” has read half of it. The other half is: under whose remit, by which rules — and who is allowed to say no.
Everyone builds agents. We build the organisation they are allowed to work in.