The Pre-Work Contract

The most expensive hour in any AI programme is the one where senior people sit in a room and discover they are not talking about the same thing. One of them means a copilot in a browser tab. One means retrieval-augmented search over policy documents. One means a loop with a harness that can call systems and act on its own. The agenda said the session was about decision rights. The opening stretch goes on vocabulary instead, and nobody wrote that time down anywhere, so nobody is accountable for it.

An executive committee governance programme I worked on for a consumer and SME banking group operating across several European markets refused to pay that hour. The preparation pack states its reason on the first line: the workshop has to focus on decision-making rather than information consumption, so every participant completes a set of foundational activities first. What follows is not a suggested reading list. It is a contract, with time budgets, hard output caps, a distribution schedule and a completion dashboard.

The room is the scarce resource, so stop spending it on context

Eleven sessions were scoped, each with a fixed duration, a stated objective, a required input and a required artefact as its exit condition. Sixty minutes to settle the target operating model and assign decision rights between a central function and the business units. Ninety minutes to classify systems against the EU AI Act's risk categories and name mandatory controls. Ninety minutes to draft the governance committee's mandate down to voting versus advisory membership, escalation paths and quorum. Ninety minutes on governance for agentic systems specifically, separate from general AI risk, covering graded autonomy, kill-switch design, and the choice between keeping a human in the loop and keeping one on it.

Add the durations up and the programme buys roughly fourteen hours of executive attention. That is the whole budget. Every minute of it spent explaining what an agent loop is, or what the Act's risk tiering actually says, is a minute not spent settling who signs off a high-risk deployment.

So the context was moved out. Ten to twelve hours of it, per person, in front of the session rather than inside it. That ratio is the part worth sitting with: the preparation load is very close to the room time it protects. The design treats those two hours as interchangeable currency and deliberately buys the cheaper one.

Three tasks, three different kinds of prior knowledge

The pre-work was not one activity repeated. It splits into three streams that do different work, and the split is the craft in it.

The reading stream is bounded and mixed on purpose: an executive summary of AI frameworks, two peer-bank case studies of eight to fifteen pages each, and the latest regulatory update. Two to three hours, three documents. One conceptual input, one comparative, one compliance. An executive who has read all three arrives able to argue by analogy and by statute, which is the register a governance debate actually runs in.

The second stream is a readiness self-assessment, ten questions in fifteen minutes, asking each participant to rate their own confidence in AI concepts, name their top three opportunities and their top three concerns, and score where they think the organisation currently stands. The instrument was built with the bank's own team rather than imported. That matters more than it sounds. A generic maturity questionnaire asks about an abstraction and gets answered defensively. One written against the live initiative list asks about projects the respondent can name, and the answers come back specific enough to be useful before the room opens.

The third stream is a written reflection: thirty minutes, one page maximum, against a fixed prompt asking what AI governance success would look like for the group at a stated horizon a couple of years out. Not an open question. A dated, concrete one, answered in writing, against the bank's existing list of AI initiatives.

That third task is the one I would steal for any programme. A fixed written prompt turns divergence in an executive team into an artefact you hold before the session. You find out on Tuesday that two of your executives mean incompatible things by governance success, instead of finding out live on Thursday with the whole committee watching and the clock running.

Distribution is part of the instrument

A workload nobody can start on time is a workload nobody completes. The pack schedules the logistics as carefully as the content: the materials are ready to send one week before, distributed through a secure portal with a minimum lead of forty-eight hours, and completion is due twenty-four hours before the session opens. The forty-eight hours is the floor; the week is the plan.

Attached to the distribution is a personal note from the sponsor. It costs almost nothing and it does the heaviest lift on the slide. Ten to twelve hours of preparation from people whose calendars are defended by other people does not happen because an outside team asked. It happens because the person who called the session asked, by name, in writing.

What a completion percentage is for

The pack shows a completion dashboard updating daily, at sixty percent when the snapshot was taken. Each of the three pre-work cards also carries its own tag and state label, so the aggregate is backed by per-item state rather than being a single number somebody typed in. The readings card reads ready to send. The written reflection reads pending review.

Sixty percent is not a result and should not be read as one. It is an in-flight reading of a design that had not finished running. What it buys is a specific follow-up. An aggregate alone tells a sponsor that some unnamed people are behind on something. Per-item state tells you which of the three tasks is stalling, which is a different conversation and a much shorter one. Reading is usually fine. Written reflection is where it jams, which on this snapshot is exactly the card still marked pending review, because it is the only item that cannot be done in the back of a taxi.

There is a second use, and it is the one I care about. The completion figure is a live forecast of what the room will be able to do. If it sits low the day before, the honest response is not to run the planned session and hope. It is to change the session, because you already know the executives are arriving without the shared floor the agenda assumed.

10-12 hours

Mandatory pre-work per participant, specified across two days

~14 hours

Facilitated room time across eleven decision sessions

48h / 24h

Minimum distribution lead, and the completion cut-off before start

60%

Pre-work completion on the daily dashboard at the time of the snapshot

The same contract, applied on the way out

The programme runs the mechanic in reverse after the room closes. A first deadline at forty-eight hours: distribute the decisions and action items, launch the feedback survey, debrief the sponsor. Then the final report, a detailed action plan, a thirty-sixty-ninety roadmap, and signed commitment cards from every executive. Then, by a stated deadline, calendar dates locked for the thirty, sixty and ninety-day reviews with agendas already written.

At the time of the snapshot the execution panel reads zero percent, awaiting workshop completion. That is the correct number for a plan that has not started, and the fact that the panel exists at all before there is anything to report is the point. Somebody built the meter before the thing it measures, which is exactly the order in which you should build a meter.

Nobody schedules the opening stretch of an executive session for levelling everyone up on vocabulary. They lose it anyway. This programme took that time out of the room, priced the replacement at ten to twelve hours, and made someone responsible for whether it got spent.

What this generalises to

Strip the banking specifics and the design is a claim about decisions in general: a decision meeting is a rendering surface, not a computation. The computation happens beforehand, in reading, in written positions, in a self-assessment nobody enjoys filling in. If that work is unscheduled, it still happens, just badly, in the first third of the session, at the highest hourly rate in the building.

Three things make the contract hold, none of them exotic. Bound every task with a time budget and a hard cap on output, because a senior person finishes a one-page assignment and does not finish an open-ended one. Attach the request to the sponsor personally, because obligation is social before it is procedural. Instrument completion per item and read it daily, while there is still time to act on it.

This is the part of a governance engagement that clients rarely ask us to price, and it is the part that decides whether the rest of it works. RealAI's Consult team writes the pre-work contract before it writes the agenda, for the same reason we specify evaluation sets before we specify a model: the artefact that constrains the work has to exist before the work does.

Details are as recorded in an executive committee information pack and workshop snapshot prepared for a consumer and SME banking group operating across several European markets: its pre-work specification, its session schedule and its follow-up plan. The time budgets, output caps and deadlines are the programme's design requirements, not measured outcomes; the completion figures are in-flight dashboard readings taken while the programme was still running. Reading the preparation contract as the load-bearing part of the design is ours.