The AI Operating Model

The problem with “adding AI”

Most organisations meet AI one tool at a time: a coding assistant here, a chat window there. Each tool helps an individual. None of them changes who decides what, how knowledge is kept, or how anyone knows the work is right. Those questions decide whether a software organisation delivers well, and they are left exactly where they were.

Why it matters

When the tools get more capable, the unanswered questions get more expensive. An agent that can change a hundred files in an hour can also spread a misunderstanding across a hundred files in an hour. Without decided answers, organisations drift into one of two failures: they approve every step by hand and lose the benefit, or they let the agent run and lose control.

The AI Operating Model

An AI Operating Model is the way an organisation is run when AI is a working participant in it. Processes, controls, knowledge and decisions are designed so that AI carries defined work under governed autonomy. People set direction, own the decisions reserved to them and assure the outcomes, and performance is measured from evidence. The model spans the whole organisation: its people and roles, its processes, its knowledge, its technology and its governance.

AI Operating Model · the whole organisation

Governance

Principles, standards, risk appetite and reserved decisions

People

Set direction, own reserved decisions, assure outcomes

Knowledge

Catalogues, decision records and shared memory

Measurement

Evidence, reviews and benchmarks feed back into governance

Agentic execution layer

AI agents carry defined work end to end, inside the controls above

Three models compared

• A traditional operating model. People do the work, with software as a tool. Governance is management review and periodic audit, and change arrives through projects.
• An agentic operating model. AI agents plan and carry out multi-step work end to end, with guard rails and approval points around them. It describes the execution layer: how work gets done.
• An AI Operating Model. People and AI each do the work they suit. Controls are built into every process, human touchpoints are named, and progress is measured from evidence and improved continuously. Agentic execution is one layer inside it.

The difference matters when judging any claim about AI delivery. A demonstration of agents doing work shows the execution layer. It says nothing about whether the decisions around that work are governed, whether knowledge is kept, or whether anyone could reconstruct why a change was made.

Governed autonomy, not approval of every step

Governed autonomy means the agent acts without asking, inside limits that are enforced rather than requested. The platform report lists the decisions that stay with people: committing to build a piece of work, accepting a risk above appetite, signing an external commitment, approving a destructive operation, and ruling on product questions. Everything between those points is the agent’s to do, and the controls, not a person watching, keep it inside the lines.

How it is measured

Progress is read from evidence, not declared. For each piece of work, requirements must be matched by code, then by passing tests, then by documentation, and a claim of “complete” that no test supports blocks progress. Across the platform, 135 bounded pieces of work tracked through the phased lifecycle between February and September 2026.

What it does not yet cover

The documented application is software delivery and support. Other business functions have not been transformed, and the operating model’s own autonomy is extended one category of decision at a time, as trust is earned. Productivity against a traditional team has not been measured.

Applying it in another organisation

Start by listing the decisions that must stay with people, and name who holds each one. Then write down the rules the work must follow, and put each one where the agent acts. Decide what counts as evidence that a requirement is met. Only then choose tools. An organisation that does these four things with ordinary tools will get further than one that buys an agent first.