Six steps.Understand it first. Then redesign it.
Ground
What the organisation takes in, and what it delivers.
Before anyone is asked a single question, we model what the organisation says it is — from public and client-supplied sources only, and about roles, never about people.
Read what is already public
Job adverts are the richest public source there is. Annual reports, procurement policy and tenders do the rest.
Establish what goes in and what comes out
A citizen files an objection and a decision comes back. A tenant reports a leak and the leak gets fixed. Those chains are the spine of the map; everything else is how the one becomes the other.
Count the people in every role
Descriptions, responsibilities and systems, with no names anywhere — and how many people hold each role, which is what sets the pace and the price of the whole engagement.
Compound it per sector
Grounding material is sector-specific, not client-specific. Once one municipality is done, the next is nearly free.
Job adverts are the richest public source there is, and structurally underrated. An advert describes a role in detail, names the systems it works with, identifies the adjacent roles, and gives away what is changing.
Annual reports, procurement policy, press releases and tenders do the rest — tenders in particular give away system migrations before they land.
Grounding also fixes what goes in and what comes out: the objection that arrives, the decision that leaves. We already know what an organisation does; what nobody can describe is the chain of work in between.
Ask
Fifteen minutes. Everyone. In their own words.
An AI conducts every interview, asynchronous and at the person's own desk — and each one starts from everything the conversations before it established.
The opening is written per role
Not “describe your work” but “the instruction says you close X in system Y — do you? And when it is backed up?”
Each wave knows what the last one found
Questions are generated from what is still missing: a handover with no receiving role, a step nobody has described, a chain that reaches no end. A consultant cannot rewrite the interview guide between every conversation.
Adaptive depth
Once a role's picture converges the conversation shortens. Where roles contradict each other, it lengthens.
We draw who is asked, not you
The moment a client picks the representative for a role, the selection bias is back — and that is exactly what is wrong with twenty interviews.
Twenty consultant-led interviews run one after another, through a single diary. That is what makes the answer cost a quarter. Automated conversations do not queue, and each starts from what the previous ones established — which is what turns a sample into everyone.
Participation is the whole game. At fifteen per cent response you have self-selection again, and the only advantage of the method has evaporated.
Everyone gets their role profile back to correct: the role as you expect it, the role as it runs, the variants, the handovers and the waits. It describes the role, never the person — you should recognise your work without finding yourself in it.
It is finished when everyone has been asked and every chain can be walked start to finish. A gap that survives is either a finding — nobody here knows what happens between intake and assessment — or a hole in our work, and the report says which.
Difference
Where what you expect and what happens come apart.
Expectation against reality, per process — and the contradictions between levels.
Expectation against reality
Which steps nobody has written down, and which steps you believe are happening that nobody does any more.
Handovers and waiting time
Where work leaves a department, what happens in the gap, and which roles carry knowledge alone.
Contradictions between levels
The board thinks X, the floor says Y. Until now there was no instrument that heard both at once.
Variants: same trigger, same goal, different route
One person uses the system as designed. One keeps a spreadsheet because it rejects a case type. One phones a colleague because it beats the queue. A process with many variants is a process that does not fit the work.
Shadow tooling gets its own section: the spreadsheet, the group chat, the phone call. These are not failures to be stamped out — they are the load-bearing parts of the real process.
Variants are the finding nobody else produces: anyone sampling one person per role cannot see them, and process mining cannot either, because the spreadsheet and the phone call leave no log. They are also where a redesign breaks, on first contact with the route the documentation never had.
They are reported like everything else — attached to a role, counted, never named. “Four of the eleven people in this role bypass the case system for renewals” is reportable. The same sentence about one person is not.
Direct
The investment case, ranked and bounded.
With confidence intervals instead of false precision, and an explicit account of what we could not establish.
Ranked and evidenced
Every recommendation carries the statements that support it and the number of participants behind them.
Margins, not false precision
A number that shows its own margin of error is worth more than a number that merely looks precise.
What we could not establish
Which roles under-responded, which processes stayed out of reach, and where the conclusion is thin.
Every recommendation carries the statements that support it and the number of participants behind them, so a board can see how thick the evidence is before committing money to it.
Redesign
AI designed in, not bought alongside.
If we designed this process today, knowing what a model can do, what would it look like?
Redesign
The process takes a different shape. Steps disappear rather than getting faster.
Leave alone
Some processes are good, too rare, or too judgement-heavy. A report that recommends AI everywhere is a brochure.
Delete
Some processes exist because a system from 2009 demanded them. They deserve to disappear, not to be automated.
What stays human, and why, is a mandatory part of every proposal. That is not a concession to the works council but good design: a process nobody understands any more fails on the first exception.
Build
On the process that runs, not the one you expect.
Off-the-shelf AI assumes a standard process. No organisation runs the standard process.
Built on observed reality
The solution follows the process that runs, including the spreadsheet, the group chat and the phone call.
The expensive part is already paid
Working out what the software should do normally dominates the cost. Steps one to five settled it.
What stays human, and why
A mandatory part of every proposal. A process nobody understands fails on the first exception.
We do not sell you the build
What leaves is an executable specification; who delivers it is your choice. An advisor who names his own build arm is no longer an advisor.
What leaves is an executable specification: what the software has to do, on the process as it actually runs. Who builds it is your choice — an implementation partner, a supplier you already work with, or your own team.
We do not sell you the build. An advisor who also sells the solution is no longer giving advice, and the organisations most likely to bring us in are the ones that build. Staying out of that work is what keeps us usable to them.
An interview identifies people. This is what stops it.
Not a policy but a pipeline. An interview identifies people even without a name — “I’m the only one who does the Tuesday reconciliation” points at one person in a team of twelve. So it has to be built in, not promised.
01
Transcript
Exists during the conversation, is processed afterwards, and is destroyed. Never stored.
02
Extraction
Discrete statements are lifted out of the transcript: a step, a system, a wait, a workaround, a bottleneck.
03
Generalisation
A mandatory pass that weakens or removes self-identifying detail before anything is written to storage.
04
Role tagging
Every statement hangs off a role and a process. There is no field for a person, and no session identifier that could be joined to one.
05
Threshold
Nothing reaches a report until several people have independently said the same thing. That number is called k, agreed before we start. Below it nothing is shown at all — not summarised, not hinted at.
No per-employee view. Not hidden behind a role, not switchable on request. And no archive that can be searched later.
Both follow from the pipeline above rather than from a promise, which is the only reason they are worth stating. k is set per engagement, in writing, before any conversation takes place.
What compoundsYour map gets read against every other map in your sector, not on its own.
What a large consultancy really sells is its basis for comparison. After five or six engagements in one sector we hold material nobody else does: which handover fails in every housing association rather than only in yours.
Opt-in, arranged contractually from engagement one, and suppressed below the same threshold as everything else. Being first in a sector costs nothing extra; it just means the comparison arrives with the second map instead of the first.