How an organization's working reality is gathered, made reachable, and turned into a decision
Method document. Revision 1. Six phases, twenty-two steps. What is old here, what is new, and what is unproven are named in the last two sections.
What this is
A structured process for establishing how work actually moves through an organization, converting that into a reachable context layer, and producing a determination about what should happen next.
It runs about fifty hours of gathering and fifteen of synthesis, plus one meeting.
Its output is not a recommendation. It is an artifact, and a decision that follows from the artifact. What that decision turns out to be is not known at the start and is not supposed to be.
The principle it rests on
Every improvement an AI system can make to an organization happens at a moment of interface between a person and the system. There is nowhere else for value to enter.
Those moments occur only inside the discrete acts that make up somebody's work.
Therefore the set of possible improvements is bounded by the set of acts in the organization's workflows, and that set is finite and enumerable.
Which means an organization that has not mapped its workflows cannot know its own opportunity space. Not has not got around to it. Cannot, because the space is defined by acts nobody has articulated.
What is old about this
Most of this is not new and I want to say so at the front rather than have somebody tell me at the back.
Organizations have been inventorying their information resources since at least 1988, under the name information audit, with a real methodology and a real literature. Burk and Horton's InfoMap in 1988 was the first full treatment. Orna in 1990 and 1999, Buchanan and Gibb in 1998, and Henczel's seven-stage model in 2001 built out from it. If you read those and then read this, you will recognize the shape of Phases 0 through 4 immediately. Discover the resources, map the flows, identify the gaps, evaluate against what the organization is trying to do. That is their frame and I am using it.
Buchanan and Gibb's own comparative review in 2008 named three problems that had dogged the practice: limited guidance on managing scope, an ambiguous relationship to the technology, and no standard method. Anyone who has watched an audit stall will recognize all three.
The reason audits were done once and then never maintained is simpler than any of that, and it is the thing that changed. The catalogue could not answer anything. You finished a large inventory of your information resources and what you had at the end was an inventory. Keeping it current cost real money and returned a more current inventory. So the second one never happened, and by year three the artifact described a company that no longer existed.
Two economics moved at once. The catalogue became answerable, which gives maintaining it a payoff it never had. And the gathering got cheap enough that the audit can be repeated rather than performed once as a project.
So what is actually new here is narrow, and I would rather state it narrowly than let it be assumed wider.
The reachability states. Marking each artifact as reachable, stranded, tacit or unmade, and splitting unmade into unwritten and undecided. The audit literature identifies gaps. It does not, as far as I have found, separate the four states by what it would cost to close each, and the four costs are completely different currencies.
The danger cells. Sorting questions by settled or unsettled, reachable or unreachable, internal or external, and locating the two cells that matter. That axis comes from what these systems do wrong rather than from what an organization holds.
Refusal as a designed behavior, which is discussed in Phase 5 and is not novel in itself. Corrective retrieval systems already evaluate whether retrieved material is correct, incorrect or ambiguous and act on insufficiency. What I have not seen elsewhere is the internal-versus-external distinction and the argument that an ungrounded internal answer is undetectable in a way an ungrounded external one is not.
Everything else in this document has a thirty-year literature behind it and should be read as an application of that literature to a moment when the catalogue can finally answer.
Two properties of the artifact
It is tool-independent. Models change, vendors change, plans change. A map of how work moves through a building is stable for years and every subsequent decision runs on it.
It is complete on its own. If nothing is built afterward, the organization owns a working description of itself that it did not previously have.
Phase 0 · Frame
Roughly 3 hours. Before anything is gathered.
0.1 · Scope and access
Establish which parts of the organization are in scope, who authorizes access, and whether the sponsor decides or recommends.
Output. A one-page scope note, agreed in writing. Gate. Do not proceed without knowing who decides. Discovering it in week five is the single most common cause of a stalled engagement.
0.2 · Role census
List every distinct role, not every person. Two people doing the same job is one role. One person doing two jobs is two.
Output. A role list with a name against each, marked for interview order. Failure mode. Accepting the org chart. It describes reporting, not work. Ask the person who books the meetings who actually gets pulled into things.
0.3 · Artifact request
Request what exists, in whatever state it exists. Ten categories. Do not ask anyone to prepare, summarize, or clean anything.
Output. Whatever arrives, plus a record of what was requested and did not. The absence is data. A category that produces nothing has told you more than one that produces forty files.
Phase 1 · Intake
Roughly 8 hours. Before the first interview.
1.1 · Inventory what came in
Every document, its form, its date, where it lives, whether it is current, and whether a machine could reach it as it stands.
Output. The artifact inventory, first pass, with a reachability state against each row.
1.2 · System inventory
What holds the work. Project system, file storage, communication, finance, and whatever else. Note what is authoritative and what is a copy.
Output. A systems list marked for whether each is reachable, and by whom.
1.3 · Build the provisional map from documents alone
Before speaking to anyone, construct the workflow as the documents describe it.
This is the official version and it is going to be wrong. Build it anyway, because the gap between it and what people say is one of the two most valuable findings in the whole process.
Output. A provisional flow diagram, explicitly labelled as the documented version.
1.4 · Hypothesis set
Name the three to six flows you expect to find, and write down what you expect each interview to confirm.
Why. So that a surprise registers as a surprise. Without a stated expectation, everything sounds equally interesting and nothing stands out.
Output. A written hypothesis sheet, dated. Do not revise it during the interviews. Revise it once, in Phase 3, on the record.
Phase 2 · Interview
Roughly 30 to 35 hours. One hour per role.
2.1 · The instrument
Twelve questions. Same twelve, every role, every time. Consistency is what makes triangulation possible. Where a role requires a follow-up, ask it after the twelve, not instead of one.
1. Walk me through the last piece of work you finished. Start from the moment you first heard about it. A specific recent job, never a general description. General descriptions produce the official version. A specific job produces what happened.
2. Where did it wait? Queue detection. People will name waiting readily because it is not their fault.
3. What did you have to go and find out that somebody else here already knew? Context reconstruction, measured directly.
4. Who did you have to ask? Produces the real network, which is never the org chart.
5. What did you have to explain to someone that you had already explained to someone else? Seam detection. Every repetition is information failing to travel with the artifact.
6. Where in this is there an actual rule, and where is it just how we do it? Separates decided from habitual. This question produces the undecided register and nothing else does.
7. What would you have to tell a new person that is not written down anywhere? The unwritten extractor. Ask it slowly and wait through the silence.
And note carefully what this question can and cannot reach. It reaches knowledge somebody holds and could say if asked, which has simply never been written down. That is most of what people call tacit and it comes out readily in an interview.
It does not reach knowledge that is genuinely tacit in the strict sense: the judgment a practitioner cannot articulate even when they want to, which is learned by doing and transmitted by apprenticeship. The knowledge management literature is consistent on this, and Botha, Kourie and Snyman put it plainly, that the more tacit the knowledge the less a high-technology solution helps. Chua and Lam found in 2005 that over-emphasis on systems led directly to neglecting the tacit layer.
So this question produces articulable-but-unwritten material, which is extractable, and it produces a boundary marker where the person cannot say it, which is craft and belongs to question eight. Confusing the two is the most likely way this method produces a map that is wrong about the work while being right about the routine.
8. What part of your job is not a sequence? Where are you just thinking? The craft exclusion. Marks what must not be decomposed.
9. What do you put into an AI tool now, if anything, and on what account? Asked without judgment, in a flat voice, as an inventory question. It is one.
10. What would you never put into one? People name their own line, and the lines they name are usually better than the policy would have been.
11. If you got two hours a week back, what would you do with them? The value anchor. Also the only question in the set that is about them rather than about the work.
12. Who else should I be talking to? Census correction. Run until the names stop being new.
2.2 · Recording protocol
Every interview recorded, with consent, stated at the top: this is recorded, it is not shown to anyone as an individual response, and nothing attributed to you appears anywhere.
The reason recording matters is not capture. It is that it frees the interviewer to listen.
Extraction happens downstream. In the room, do not take structured notes and do not attempt to build the map. One job only: listen.
2.3 · What to listen for
Not content. The transcript holds the content. Listen for the things that will not survive transcription.
Hesitation before an answer about who decides. Almost always an undecided rather than an unknown.
"Technically we're supposed to." The gap between the documented process and the real one, offered voluntarily.
"It depends." Followed by a pause. Either genuine judgment, which is a craft marker, or an unresolved rule, which is not. Ask which.
A handoff described differently than the person on the other side described it an hour ago. The highest-value finding available and it only appears across interviews.
A workaround, described casually. People do not flag these because they have stopped noticing.
A person named where a role should be. Single points of failure announce themselves this way.
Anything said with a laugh. Almost always a known broken thing that everyone has agreed not to fix.
2.4 · Order
Operations or traffic first, because they see the whole flow and their account is the frame everything else adjusts against.
Then the disciplines, in the order work moves.
Leadership last, not first. A leadership account taken early becomes the frame and distorts everything after it.
And at least one person who has been using AI on their own initiative, whoever that turns out to be. Frequently the most useful hour in the whole phase.
Gate. Stop interviewing when question twelve stops producing new names. Continuing past that point costs money and adds nothing.
Phase 3 · Integration
Roughly 12 hours. Where the map is actually made.
3.1 · Transcribe and structure
Every recording to text, then to a structured document with the twelve questions as the spine. Machine-assisted, checked by a person.
Output. One structured document per interview, plus a combined corpus organized by question rather than by person.
3.2 · Extract against the known targets
The targets are known in advance, which is what makes this tractable. For each interview, extract:
Acts and their sequence · queues and their causes · handoffs · context reconstruction moments · rules, marked as decided or habitual · unwritten knowledge, marked as articulable or genuinely tacit · craft, marked as not-a-sequence · tools in use, sanctioned and not · data categories touched · stated refusals.
On the sixth of those. Articulable means the person could say it and nobody asked. Genuinely tacit means they cannot say it, and no amount of interviewing will get it out, and the honest thing to do is mark the boundary and leave it inside the craft region. An extraction set that treats the second as the first will contain confident summaries of things nobody actually said.
Output. Ten extraction sets, each traceable to the interview it came from. Rule. Nothing enters an extraction set without a pointer to the source. An extraction with no provenance is an opinion.
3.3 · Triangulate
Compare every account of the same flow against every other.
Agreement is not the finding. Contradiction is. Where account and traffic describe the same handoff differently, the truth is in neither and the seam is real.
Output. A contradiction register, ranked by how much work passes through each contested point.
3.4 · Second pass on contested seams
Go back to the two or three seams where the accounts conflict most, with both parties, separately, and ask about that specific handoff.
This is not optional and it is where most of the value in the whole method concentrates. It is also the step most likely to be cut for time.
Output. A resolved account of each contested seam, or an explicit statement that it remains contested, which is itself a finding.
3.5 · Mark the craft
Go through the map and mark every act that is not a sequence.
A map that decomposes judgment into steps is wrong, and a practitioner will see it in ninety seconds. Marking these honestly is what earns the right to decompose everything else.
Output. The map, with craft regions bounded and labelled.
3.6 · Revise the hypothesis, on the record
Return to the Phase 1 sheet. Write down what was expected, what was found, and where they differ.
Output. A one-page revision note. It goes in the final artifact.
Phase 4 · Ingestion
Roughly 8 hours. Making it reachable.
4.1 · Assemble the corpus
Everything gathered, converted, structured, and organized by retrieval intent rather than by origin.
Output. A reachable corpus with structure preserved.
4.2 · Attach provenance
Every passage carries what it is, when it is from, and where it came from, inside the text rather than in metadata a system may drop.
Gate. Nothing enters the corpus without provenance. No exceptions, including for convenience.
4.3 · Build the exclusion set
From the answers to questions nine and ten, plus the governance layer, plus judgment.
Three categories: reachable, reachable behind a wall, and never. The third is removed from every denominator, because the correct coverage on some rows is zero.
Output. The exclusion list, named, with a reason against each row, ratified by the sponsor. Gate. No corpus goes live before this is signed. A context project without an exclusion list is a breach with a roadmap.
4.4 · State every artifact
Against the ten layers, each artifact marked reachable, stranded, tacit or unmade, with unmade split into unwritten and undecided.
Output. The coverage picture, computed rather than estimated, with the gap shape per layer: how many stranded, how many tacit, how many undecided, and what each would take.
Phase 5 · Iteration
Roughly 8 hours. Working the context before anyone else has to.
5.1 · Query the corpus against the real questions
Take the questions people actually asked during the interviews and run them.
Output. Per question: grounded, thin, or not covered.
5.2 · Locate the danger cells
Identify questions that are unsettled, internal, and unreachable. These are the ones where a general system will produce a fluent answer about nobody.
Output. A named list. This is the finding that makes the case, and it is made of their own questions rather than examples.
5.3 · Build two or three small things
Not a system. Small, real, from their context, at points the map identified.
Purpose is demonstration, not delivery. Show, do not describe.
Rule. Nothing shown that does not work. No mockups, no staged data, no happy path only.
Phase 6 · Determination
One meeting. What, so what, what now.
6.1 · Deliver the artifact first, separately
The full pack goes to the sponsor at least three working days before the meeting.
Reason. The meeting is a decision, not a reveal. A room seeing findings for the first time reacts rather than decides.
The pack contains. The workflow map, all flows, with craft marked. The contradiction register. The artifact inventory with states. The coverage picture with gap shapes. The exclusion set. The danger-cell list. The hypothesis revision. And what was built, working.
6.2 · The meeting
Three movements, in order, and the order is the whole structure.
What. The map, read back. No interpretation. The room corrects it. Corrections are recorded and change the artifact.
So what. Where work stalls, what it costs, and which questions currently land where the system cannot help. Still no recommendation.
What now. The room determines the next step from what is now in front of it.
6.3 · The bounded outcome set
The determination is not open-ended. Given an artifact of this shape and a meeting of this structure, the outcome falls within a known set, and the set is stated in advance so nobody suspects the conclusion was pre-written.
Fix the workflow first. The constraint is organizational and no tool touches it. In this case the honest answer is that the next work is not ours.
Do it with what you already own. The gap closes with configuration, templates, a database, and the tooling already licensed. Recommended with the specifics, and no further engagement follows.
Close the gathering gap. The map shows the cost is stranded and tacit material. The work is conversion and extraction, not building.
Build something specific and bounded. A named intervention at a named seam, with success and failure criteria.
Or not yet. Something needs deciding first, usually in the undecided register, and no build survives an unresolved decision underneath it.
And one refusal, stated in advance and held. Partial implementation on partial context produces confident output about nobody. That outcome is not offered and will not be assisted with, regardless of who asks or what they are prepared to pay.
What is not in this method
No training before the system exists. Formation of judgment about a system attaches to a system, and there is nothing to form judgment about beforehand.
No open-ended sandbox. A bounded place to experiment, with nothing of the organization's own in it, is a blank page, and nobody explores a blank page.
No policy written ahead of the intake. The exclusion set and the governance layer fall out of questions nine and ten, and a policy written before the intake describes an organization nobody has met yet.
No recommendation formed before Phase 6. A recommendation carried in from the start is a conclusion looking for evidence, and the people worth working with can identify one at a distance.
What this does not prove
One claim I want to keep out of the method, because I have watched people make it on my behalf.
The whole argument rests on the idea that an ungrounded answer about your organization displaces a better outcome, quietly, without leaving a signal. I believe that. I have not measured it, and neither has anyone else that I have been able to find. There is no study establishing how often it happens, what it costs, or whether organizations that close the gap perform differently from organizations that do not.
The absence of a literature is not evidence for a claim. It is only an absence. If displacement turns out to be rare or cheap, most of what follows from Phase 5 is less important than I think it is, and I would want to know that.
What the method does establish is smaller and is not in doubt. Which questions your people actually ask. Which of those the corpus can currently answer, and which it cannot. That is a count, made of real files and real conversations, and it is true whatever the displacement rate turns out to be.
Failure conditions
Access slips and the interview phase compresses. Raise it at the midpoint and choose explicitly between narrowing scope and extending, rather than doing thirty rushed hours.
Leadership is interviewed first. Their account becomes the frame and every subsequent interview adjusts toward it.
The contested seams are not revisited. The most valuable step is skipped, and the map records the official version.
Craft is decomposed. The map is accurate about the routine and blind about the work, and a practitioner dismisses it in ninety seconds.
The exclusion list is deferred. Everything downstream is now a liability.
Or the determination can only conclude one thing. If the meeting cannot end in an outcome that does not involve us, it is not a determination. It is a sales close with a whiteboard, and the room will know.