The collaboration model

The collaboration model is the centre of our practice. It is the written understanding of how a corporate team and the AI agents it directs divide labour, hand work back and forth, and stay accountable. Everything else we do — the role atlas, the oversight protocols, the measurement notes — hangs from it.

The five layers of the model

A workable model can be described in five layers. They are deliberately few, because a model with forty layers is read once and then abandoned.

• 01IntentThe business outcome the ensemble exists to produce, stated in plain language and owned by a named executive.
• 02PartsEvery human role and every agent role in the workflow, each with a short description of what it contributes and what it must never do alone.
• 03Hand-offsThe precise points where work passes between a person and an agent, with the state the work must be in at that boundary.
• 04SupervisionWho reviews an agent's contribution, at what depth, and what evidence records that the review occurred.
• 05ReviewThe cadence on which the model itself is re-read and revised, so it tracks the business rather than standing still.

Most organisations we meet can describe layers one and two with reasonable confidence. They struggle with three, four and five. Hand-offs are informal, supervision is assumed, and review never happens until an incident forces it. That is why our diagnostic spends most of its time at the boundaries.

The test we apply is simple. If a new team member joined tomorrow and read only the model, could they tell where their work ends and an agent's begins, and who checks the result before it moves on? If the answer is no, the model is still a draft.

What makes a model durable

Durability is not the same as permanence. A good model changes often, but it changes on purpose. We look for four properties.

• Property 1

  Legible

  A busy manager can explain the model to a peer in three minutes without notes. If it takes a slide deck, it is too heavy.

• Property 2

  Bounded

  Every agent role states its limits as clearly as its duties. Boundaries are where trust is built, not where it is broken.

• Property 3

  Evidenced

  Supervision leaves a trace. The trace is light, but it exists, so a reviewer a month later can see what was checked.

• Property 4

  Revised

  The model has an owner and a date for its next reading. Nothing we write is meant to be final.

A collaboration model earns its keep at the hand-off, not on the title page.

Where models usually break

Across engagements we see the same four fractures. First, the invisible escalation: when an agent produces something doubtful, the person quietly fixes it and nobody learns. Second, the runaway agent: a role quietly expands beyond its bounds because nobody restated the limits after a change of systems. Third, the orphan review: a supervision step exists on paper but has no named reviewer. Fourth, the stale map: the business reorganised, and the model did not.

None of these are dramatic. That is exactly why they matter. They accumulate until a board asks a simple question — who is accountable for this output? — and no one can answer cleanly.

How we compose the model with you

1. Read the current arrangement. We gather what exists: org charts, process notes, the informal rules teams actually follow, and the places agents already appear.
2. Draw the parts. We produce a first role atlas, human and agent, and mark every hand-off between them.
3. Draft supervision. For each boundary we agree depth, owner and evidence, deliberately keeping the burden light enough to survive a busy quarter.
4. Agree the review cadence. We set the date and the owner for the model's first revision before the engagement closes, so the model stays alive.

If you would like to see how the model sits inside a full engagement, continue to engagement shapes, or read the role atlas for the vocabulary we use when naming parts. When you are ready, the briefing room is open to corporate clients.