Ration the Depth

Every scoping conversation I have had this year about assistants and copilots opens with the same question, and it is the wrong one. Which processes do we start with. The list comes back at thirty or forty, somebody sorts it into tiers, and the room feels productive, because a list looks like a decision.

The decision that actually determines whether the programme survives contact with a calendar sits one axis over. Not which processes. How far down each one you are willing to go, and what you are prepared to leave undescribed.

I went back to a proposal I worked on for a European travel and tourism group that was standing up a new central digital function. The structural blueprint was finishing and the question underneath it was whether the thing could be run on a Monday. Around thirty processes had been identified as needing description. The commitment was process flows with roles for all thirty, at one level of detail, inside a three-month plan that also carried the organisation design and the governance work. Five processes, and only five, were named for description at the level below that.

Two numbers on one page, thirty and five, and the gap between them is the whole method.

Breadth and depth are not the same commodity

The reason people scope process work badly is that the two levels sound like adjacent rungs on one ladder. They are not. They are different activities with different inputs and different economics.

A flow at the shallower level says what the process is, what its main steps are, and who owns each one. It can be assembled. Somebody who already runs the thing tells you what happens, you reconcile that against a reference shape, and you edit the difference. It is reconciliation work, and reconciliation work compresses well.

A description at the next level down says what happens on each step. Which exception fires when the supplier does not confirm, who is allowed to override, what the handover carries, what the system does when the field is empty. That cannot be assembled from an interview, because the person you are interviewing does not hold it as a description. They hold it as a set of habits. Getting it out requires a room, a whiteboard, and several people who disagree with each other, and the disagreement is the point.

Twenty-five days to take five processes to that level is what that room costs. The proposal says so, in a provision it prices separately from everything else. Five days a process is our division of its number rather than a rate the document quotes, and it is the only place in the pack where days are attached to a count of processes at all.

The arithmetic that makes the choice for you

Take the document's own number, divide it out, and run it across the estate. Thirty processes at five days each is one hundred and fifty days of effort. The committed team, three people across the quarter covering the organisation design, the governance design, the process work and the senior oversight, was quoted at one hundred and fourteen days.

Detailing everything properly would have cost more than doing everything else.

That is not a scoping subtlety, it is a wall. And once you have seen it written down, the thirty-to-five ratio stops looking like a compromise and starts looking like the only shape the work could have taken. Breadth across the estate so that nothing is invisible. Depth in five places so that five things can actually be operated.

Depth was held outside the committed price

The part I find most instructive is where the deep work sits commercially.

The committed plan names five processes for the deeper level. The same document then holds a separate provision of twenty-five days to describe five key processes in detail, priced on its own line, with a confirmation date attached. Whether those are the same five is not something the document lets me settle. What it does show is that the most expensive kind of description in the pack was the one thing not folded into the headline number.

That is a discipline worth copying. The work you cannot size in advance gets its own price, its own scope and its own expiry date, rather than an open change mechanism and a difficult conversation in month two. It also forces the decision early. Somebody had to say yes or no three weeks into a fourteen-week engagement, which is early enough that the answer is a judgement about the shape of the work rather than a report on it.

I would rather make that call three weeks in with a price attached than six months in with a slipped date attached.

How the five were chosen

The sequencing rule for the sweep is stated plainly: start by describing the complex and critical areas, and engage the process owner in the design work. The examples named are financial processes, capacity management, standing up a delivery centre, gating, and end-to-end product management. Not the easy ones. Not the ones with a willing sponsor and a tidy handover.

Confirming the process owner and the sponsor appears as its own item on the plan rather than as an assumption inside the design activity. So do the words workshop, refine and validate, each carrying its own bar. That tells you what the deeper level actually is: scheduled argument, with named people obliged to turn up.

The candidates flagged for the deepest treatment are described in the document as new processes, and the examples given include customer experience, service management and idea intake. Those have no incumbent to interview and no existing flow to reconcile against, which is why they are the ones a reconciliation sweep cannot absorb.

~30

Processes identified and committed at one level of description

5

Processes named for the level below

25 days

Priced provision to describe five key processes in detail

150 vs 114

Implied days to detail the whole estate, against days quoted for the committed team

Why this is an AI scoping problem now

Retrieval-augmented systems have made breadth extremely cheap. Point a pipeline at the document estate, embed it, put a vector store behind it, and a copilot can talk credibly about all thirty processes by the end of a fortnight. That capability is real and it is worth having.

It is also the shallower level, restated. It tells people what the process is. It does not tell a system what to do when the exception fires, and it cannot, because nobody has written that down anywhere for it to retrieve.

Everything that makes an assistant safe to put in front of a customer lives at the deeper level. The evaluation set is built from the exceptions, and the exceptions live there. The handover point where a person takes over is a step on a flow, and if the flow does not exist, the handover is a hope. Lineage, the question of which system of record a value came from and what it means, is answered at that level or not at all. The same is true of the early, carefully scoped autonomous experiments people are running: the boundary of the experiment is a set of process steps, and an undescribed boundary is not a boundary.

The regulatory direction reinforces it. European rules on AI are converging on a documentation regime, and the documentation people will eventually have to produce is process documentation with a model sitting inside it. Thirty processes described at the shallower level is cheap insurance against that. Five described properly is where the programme actually earns anything.

Breadth is a list. Depth is a room with people in it, arguing about what happens when the booking fails. One of those you can buy thirty of. The other you can buy five of, and the honest question is which five.

The rule I would write down

Describe everything once, so nothing in the estate is invisible and the next candidate is always identifiable. Then take five things deep, and pick them by asking a single question about each one: does the extra detail change a decision somebody makes, or a step a system executes. If it does neither, the shallower flow is finished work, not unfinished work.

Budget the deep five separately, with a date on the decision, and be honest that depth is the line that does not compress. Confirm an owner before the room is booked, because a deep description without an owner is a document nobody will defend six weeks later.

That is the block we run at the front of a RealAI Platform engagement, and it is why we ask which processes you intend to operate differently before we ask which model you want to use. Answer it and you have built nothing yet. You have a sorted estate: thirty things you can see, five things you can act on, and an arithmetic reason for the boundary between them rather than an argument about it.

The modelling has never been cheaper. Pipelines are assembly work, evaluation tooling is improving monthly, and the retrieval layer stands up in days. Depth is the one input that has not fallen in price, because it is still people in a room disagreeing, and it never will.

Figures are as written in a phase-two proposal for the operating model of a new central digital function at a European travel and tourism group: its scope statement, its three-month activity plan and its priced commercial offer. The one-hundred-and-fifty-day figure and the five-days-a-process rate are our arithmetic on that offer's twenty-five-day provision, not numbers the document states. That document produced a scoped and priced plan, not delivered results. The placement of individual activities within the plan's calendar is not recoverable from the source and is not claimed here. Reading the thirty-to-five split as a rule for scoping AI work is ours.