I thought the wiki was there to help AI.

Help AI understand the organisation as a map. Then help AI run workflows against living truth — receipts policy, deployment approach, change principles — instead of freezing those facts into a prompt that is wrong the week after you write it. Both of those jobs are good reasons to build an organisational wiki. I still believe both of them. They are just not the reason that makes this whole architecture feel inevitable once you have used it with a real knowledge worker in the chair.

The reason that lands is quieter. A marketing person stops trying to schedule the interview tour. They stay in marketing’s dialect, asking marketing’s questions, and in half an hour of conversation they have articulated the repeatable shape of past projects as an offer — not a segment-average blog post from model priors, but something only that company could publish. The staff member never left their comfort zone. The intellectual property came into the room anyway. That is not “smarter automation in the back office.” That is a person wearing a cognitive exoskeleton loaded with the organisation’s actual muscle memory.

You built the wiki to make AI smarter about the org. The product is making humans smarter about their own org — with AI as the traversal engine.

Two good stories that stop too early

The map story is the engineering story. Give the model a structured estate of claims and edges so answers are about your projects, your clients, your failures — not the internet’s average consulting firm. Without that map, every clever agent is still a tourist.

The workflow story is the operations story. Once the map exists, agents can look things up with receipts. They can show which policy page they used. They can refuse to invent a travel rule that was never written down. Hardcoding organisational truth into prompts is how you create a second, worse source of truth that nobody maintains.

Both stories treat the human as adjacent: the person who deploys the system, or the person who receives the automated output. Neither story centres the ordinary knowledge worker doing the job they already have — productising an offering, opening a blank week with a great client and nothing to sell, stress-testing an initiative before the exec meeting.

That is the inversion this book is about. The wiki is not only infrastructure for machines. It is muscle memory for people. The AI is how you walk the muscle memory at conversation speed. The human still makes the call, holds the relationship, and owns the consequence.

What this book owns

We will climb a three-step ladder — grounding truth, AI workflows, human augmentation — and then stay on the third rung long enough to make it designable. You will get three role morphologies worked end to end: marketing’s Case Library, sales’ capability join (including the humble RFP and “who was the Telstra PM?” moments), and the CEO’s parliament grounded in company history rather than generic risk language. You will get the split that does not move: AI owns connection, translation, and generalisation; humans keep taste, relationship, and accountability. You will get the expert fix — stop using Terry as a retrieval API — and the trust mechanism that makes the whole thing enterprise-grade: every material claim double-clicks to bronze.

If you want the communication-topology diagnosis — why marketing walks away with five sentences compressed along the wrong dimensions — read Your Company Speaks Five Languages — and Nobody’s Translating. We will point at that mechanism when marketing and sales dialects matter. We will not rebuild it.

A personal proof, then the org

At personal scale, the same architecture is already familiar if you have ever kept a career corpus and asked an honest question of it. Storage was never the problem. Retrieval was. The prosthetic index over intact-but-unreachable sources is the Life Wiki intuition: the archive is fine; the walk is broken. Organisations are the same pathology with more mailboxes and a worse calendar.

When the walk works, something emotionally specific happens. You make a claim about your own past work, and the system surfaces the internal email you had forgotten — the deliberation, not just the polished proposal. That feeling is not nostalgia. It is provenance. Later in this book we will turn that feeling into an enterprise requirement: dignity for the person whose work it was, auditability for the firm that must stand behind the claim. One click path. Two virtues.

The promise

After this book you should be able to design role-shaped AI assists on top of a single knowledge substrate without writing a replacement fantasy. Marketing keeps taste. Sales keeps the relationship. The CEO keeps the decision. The wiki stops being a back-office map only AI teams care about and becomes the muscle memory every role boots from.

The bot is the legs. The human still walks into the room.

Start with a boring example, because boring examples are where enterprise architecture goes to die.

Someone asks: what is our receipts policy for client dinners over $200? The amateur move is to paste the answer into a system prompt. It works for a week. Then finance updates the threshold, legal adds a regulated-industry exception, and the agent keeps citing the frozen sentence with perfect confidence. The professional move is to keep the policy in the organisational wiki and let the agent look it up — with a pointer to the page, the effective date, and the edge cases that live two links away.

That professional move is real progress. It is also only rung two.

Rung 1 — Grounding truth

Before workflows, there is orientation. The organisation’s soft data — proposals, emails, post-mortems, decks, tickets — is joined into a form AI can traverse: claims, edges, people, projects, clients. The point is not “search that works.” The point is a map that makes answers about us.

Without rung one, every later cleverness collapses into generic competence. The model is fluent. It is not yours.

Rung 2 — AI workflows

With a map, you can stop hardcoding organisational truth into prompts and tools. Agents retrieve policies, assemble context packs, and show receipts. They can answer adjacent questions the frozen prompt never anticipated. They can admit absence when the corpus does not contain a rule — which is a form of honesty most RAG chatbots never learn.

Rung two is where AI teams feel finished. The demo works. The bot cites a page. Governance relaxes slightly. Then the knowledge workers who were supposed to benefit still open the same calendar invites and still ship the same dull product pages, because nobody redesigned their prep.

Rung 3 — Human augmentation

Rung three is the same substrate used as a cognitive exoskeleton for people who never think of themselves as “AI users.” The marketer productising an offer. The salesperson with a warm relationship and a cold pipeline. The CEO about to walk a half-formed initiative into a room full of other executives.

Each of them needs different prep. None of them needs to be replaced. All of them need the organisation’s muscle memory in the conversation while the thought is still warm — reconstruction inside the half-life of the question, not a ticket that returns next Thursday.

Rungs one and two are real. Rung three is where the commercial and cultural prize usually lives.

Why programmes stall on rung two

Generic tools are excellent for individuals because they are flexible. They stall in the enterprise when they never learn the organisation’s workflows, language, or history.¹ A grounded wiki is how you stop serving segment averages. But if you only wire that wiki into backend agents, you have built a smarter machine room. You have not yet built a smarter firm.

The gap between pilots that stall and programmes that scale is often managerial, not technical.² Rung three is a managerial design problem: who gets which prep morphology, what stays human, and which expert rituals you are willing to retire.

Use the ladder as a diagnostic

On Monday, ask your programme which rung it is honestly on. If you are still arguing about chunk sizes, you are pre-rung-one. If the bot can cite policy and nobody in sales has a new motion, you are on rung two. If people in roles are walking into human moments better prepared from the same corpus — without surrendering accountability — you have touched rung three.

This book stays on rung three. The next chapter freezes the division of labour that makes rung three safe: what AI is allowed to own, what humans must keep, and why that split is constant even while the prep pack changes shape.

We have named this pattern before. The highest-leverage split between AI and human is not task-by-task delegation. It is structural: AI saturates everything before and around a high-stakes human moment; the human owns the judgment, the relationship, and the accountability. An exoskeleton amplifies the wearer. It does not replace them.

Human judgment remains load-bearing for strategy and taste precisely because preparation is now cheap and standing behind outcomes is not.³ If accountability transfers to the model, the pattern collapses into automation theatre with a human rubber stamp.

What the wiki changes: the payload, not the split

Without an organisational wiki, the exoskeleton runs on model priors plus whatever context someone pasted into the prompt. That is enough to sound professional. It is not enough to be yours. Generic marketing copy is the tell: fluent, polished, and interchangeable with any competitor’s page.

With the wiki, the exoskeleton runs on compiled muscle memory — proposals, emails, delivery notes, post-mortems — joined so a conversation can walk them. The split does not move. The payload does. Connection, translation, and generalisation now operate over company history. Taste, relationship, and accountability stay human on purpose.

AI does connection, translation, and generalisation. Humans keep taste, relationship, and accountability. Only the prep morphology changes.

The table (definitive placement)

This is the spine of the book. Later chapters apply it; they do not re-derive it.

Prep morphology is a design object

Most teams invent a chatbot and hope roles will invent their own use cases. That is how you get a bland assistant nobody trusts with real work. Morphology means you design the shape of the prep on purpose: what packet lands in the human’s hands before the moment that matters.

Marketing does not need a chatty generalist. Marketing needs a Case Library card upstream of the webpage. Sales does not need another email drafter. Sales needs the join of their exclusive client knowledge with the firm’s exclusive capability map. The CEO does not need another summary of last quarter. The CEO needs a sparring parliament that remembers the project that actually failed.

What the split forbids

It forbids AI that “decides” while a human clicks approve without reading. It forbids marketing claims nobody technical will stand behind. It forbids sales commitments the graph cannot evidence. It forbids executive theatre where risk language is generic enough to never be wrong and never be useful.

It also forbids the opposite failure: experts used as the organisation’s retrieval layer. That anti-pattern gets its own chapter. For now, hold the constant: machines may connect, translate, and generalise. Humans keep the parts that require a name on the outcome.

Bridge to the worked cases

Part II is not three new frameworks. It is this table applied three times. Same split. Same substrate. Three prep morphologies, each walked from input through wiki walk to role-shaped output. If you only remember one artefact from this book, remember the table — then watch marketing, sales, and the CEO cash it.

Strategy from the top says: productise. Build reference offerings. Stop reinventing every engagement from a blank page. Then marketing does what marketing has always done when the org has no compiled memory: book meetings, interview a project manager, walk away with five sentences compressed along dimensions the PM cared about, and publish a page so dull it could belong to any firm in the segment.

That failure is usually filed under “marketing doesn’t get it.” Sometimes it is a topology problem — serial telephone hops between technical, PM, and marketing registers — which we treat properly in a sibling piece. Here the ownership is different: once ground truth is joined, marketing gets a first-generation Case Library articulation without using the PM as a retrieval API.

Input

Composite but familiar. The firm has delivered a telco billing engagement. Internally, people are proud of a mediation-first approach that avoided a multi-year rip-and-replace. Externally, the website still says something like “helped a leading telecommunications provider modernise its billing capability.” Leadership wants an offer. Marketing owns the page. Nobody has six weeks for another interview tour.

The marketer’s question into the exoskeleton is not “write a blog.” It is: what have we actually done that is repeatable, and how do we say it as an offer?

Wiki walk

The walk hits more than the polished case study folder.

• Project page — scope, client, outcomes, people, related proposals.
• Client proposal — what was sold and constrained.
• Delivery notes — what was hard, what was true in production.
• Internal email — the deliberation: “disintermediation without rip-and-replace.” Never client-facing. Often where the reusable move actually lives.

That last artefact is the tell. Organisations archive binaries and delete source. The proposal is the compiled output; the internal framing is closer to the source code of the judgment. A meeting with the PM rarely recovers it. A soft-joined corpus can.

Role-shaped output: the Case Library card

We file plays as situation shape + move + outcome, and we promote a move toward a principle only when it recurs. That is the Case Library discipline: analogy at the right level of abstraction, not similarity spam.

Notice what this is not. It is not the final webpage. It is not a content calendar. It is the layer upstream of both: the articulation of company-specific truth as something marketable. The human still applies taste — brand voice, claim strength, what would embarrass delivery, what legal will not allow. AI connected the corpus, translated into marketing register, and generalised a reusable shape. Marketing kept judgment.

A few conversation turns in half an hour become a masterpiece of articulation — not a meeting on the calendar to start the process.

The half-hour conversation (worked)

Depth beats speed theatre. AI as a tool for thought is about the quality of the articulation, not the number of posts produced per week.⁴

Company-specific truth vs segment-average content

Generic AI, writing from model priors, produces the page any competitor could publish: leading provider, structured programme, improved outcomes. Corpus-grounded AI produces constraints, method, and proof that only your firm earned. Buyers can taste the difference even when they cannot name it. So can your delivery team — which is why they roll their eyes at the website today.

Split constant from Chapter 3, applied: connection, translation, generalisation on the machine side; taste and fair claims on the human side. Marketing stays in marketing’s chair. The exoskeleton brings the IP.

Picture the salesperson who is good at the part everyone romanticises. They have the relationship. The client takes their call. Trust is real. And this month they have nothing to sell that fits — not because the firm is empty, but because the firm’s capability map does not live in their head. It lives across roughly two hundred colleagues, a decade of projects, and systems optimised for data entry rather than read-back.

Knowledge workers already lose on the order of five hours a week waiting on colleagues or recreating knowledge that exists somewhere in the building.⁵ Sales feels that tax as empty pipeline next to a full firm.

Case A — Cross-sell from the corpus

Input

“This is my best account. What else can we honestly offer them that I have never sold?”

Wiki walk

• Ten years of this client’s proposals, emails, SOWs, support threads.
• Firm-wide projects in adjacent domains the rep has never touched.
• People nodes: who delivered analogous work; who is still available to staff it.
• Proof artefacts: outcomes, constraints, post-project notes that prevent oversell.

Role-shaped output

A shortlist of real solution shapes — not decorated generics — each with why it fits this account, who has done it, and a double-click path to bronze. The rep decides what to raise, how to sequence it, and whether the relationship can carry the ask. Negotiation and commitment stay human. The exoskeleton supplied the map the human structurally cannot hold.

Case B — RFP assembly without the Frankenstein

RFPs and RFQs are where organisations rediscover that no single person holds the whole knowledge set. The traditional fix is parallel human retrieval: Hey Terry, write the change management chapter. Hey Simone, write deployment. Hey Priya, pull three references. Everyone writes the bit they remember. Nobody wants to re-read the whole document. The voice fractures. The calendar burns.

Full agentic RFP orchestration is a larger programme — proposals that show their work end to end. The exoskeleton-only version is already valuable and deployable:

Do not undersell the Telstra question. People are nodes in the graph too. Expert-finding is the humble version of expert-replacement: sometimes the right answer is a person, and the system’s job is to route you there before the opportunity cools. It is also easier to sell politically. The tool that finds Terry is less threatening than the tool that pretends to be Terry.

“Who was the project manager on Telstra?” answered in one turn instead of three days of fumbling corporate systems.

Split constant applied

AI connects the account to firm capability, translates solution shapes into sales language, and generalises patterns that fit. The human keeps the relationship, the negotiation, and the commitment. If the system starts “owning” the client, you no longer have an exoskeleton. You have a liability with an SMTP address.

Corporate systems spent decades perfecting the write path. SharePoint became where data goes to decay. The read path stayed human-priced, so it stayed optional. (The full economics of that inversion is a sibling article; here it is enough to say: sales is one of the roles that pays the optional-read tax every quarter.)

The CEO says the quiet part: people are busy and excellent, margin is not moving, and the next initiative looks like progress if you do not look too hard. The old move is to float the idea in a room and discover the landmines live. The new move is to stress-test it first against a parliament that remembers.

Input

Composite initiative — call it Harbour: stand up a productised data-migration offer, staff it by pulling senior delivery people into a shared “factory,” and ask marketing to sell throughput. The CEO wants angles before the exec meeting: finance, HR, marketing/sales, risk. Not a slide of platitudes. A sparring partner with receipts.

Wiki walk

• Prior productisation attempts and what happened to utilisation and margin (shape of history, not fake precision).
• Campaigns that sold what delivery could not staff.
• Hiring and attrition notes that constrain the fantasy bench.
• Orion — a real internal failure in this composite world: migration factory cancelled in month seven; post-mortem names over-promised reuse and under-funded change.

The AI Think Tank pattern already argues for multi-lens deliberation with visible rejection: operations, revenue, risk, people — structured disagreement beats single-agent synthesis. What the organisational wiki adds is the difference between a textbook council and a company council. Without the corpus, every brain argues from model priors. With the corpus, Risk can open a file.

Role-shaped output: the parliament brief

The Risk brain cites the project that did fail — post-mortem attached.

Why pure translation helps the CEO least

In many organisations the CEO has already been acting as a human translation layer — expected to speak marketing, sales, finance, and delivery well enough to chair the room. Dialect translation still helps the enterprise; it may help that individual least. What they lack is not another language. What they lack is a parliament that remembers company history while they still own the decision.

That is the CEO-shaped exoskeleton: not a replacement strategist, not an autopilot board pack generator, a sparring council grounded in your past. The human keeps political capital and accountability. The machine connects history across lenses and generalises failure shapes with pointers. Human judgment remains the seat for long-horizon calls; the prep is what gets upgraded.³

Before the real meeting

The worked outcome is not a decision made by AI. It is a CEO who walks into the exec session having already heard the strongest internal objections with exhibits attached. Some initiatives die in the brief — cheaply. Some survive with sharper scope. Some get a kill criterion named in advance. That is managerial architecture, not model magic.

Split constant, applied once more: connection, translation, generalisation from history on the machine side; decision and accountability on the human side. The table in Chapter 3 just cashed its third row.

Say the sentence out loud: Hey Terry, can you write the change management chapter for this RFP?

It sounds like teamwork. Structurally it is a lookup. The organisation needs the standard change principles, the language that worked last time, the difference between current practice and last year’s template. Instead of querying a system designed for retrieval, it queries a person designed for judgment — and pays in calendar latency, incomplete recall, and a Frankenstein document nobody wants to re-read.

The organisation has been using its experts as a RAG system: slow, calendar-bound, lossy, and the results do not compose.

What experts were actually hired for

Terry’s scarce resource was never the ability to remember the boilerplate chapter. It was the ability to say: this client is regulated and hostile to change theatre; the standard chapter will lose; here is what must be different; here is what we should refuse to promise. That is judgment and novelty. That is the work worth a senior salary.

When you spend Terry on recall, you get two losses. You pay expert rates for retrieval. And you still do not get a coherent whole, because five experts each retrieve a fragment and nobody owns the spine.

After the wiki: rewrite the ask

The wiki drafts from current principles, prior RFPs, and linked bronze. Terry reviews exceptions. Simone is found in one turn when the answer truly is a person — expert-finding from Chapter 5 — not drafted into writing a chapter they could have retrieved.

The cost of retrieval-by-delegation

Hours spent waiting for colleague knowledge or recreating what already exists are not a soft culture problem. They are an architecture problem with a timesheet.⁵ Depth in knowledge work comes from better judgment loops, not from more status meetings that exist only because the archive cannot answer.⁴

This is managerial redesign. Model quality does not retire the hey-Terry ritual. A named split does: machines retrieve; humans judge. The gap that stalls programmes is often managerial, not technical.²

What changes for experts personally

Their inbox should get quieter on boilerplate and louder on hard calls. Their status shifts from “human SharePoint” to “terminal judgment on exceptions.” Their best days look like reviewing grounded drafts with exhibits attached — not reconstructing the standard chapter from memory for the third RFP this month.

If your AI programme makes experts busier on recall, you are not augmenting them. You are adding another requester to the human RAG cluster.

Start with the demo, not the theory.

Mini-demo: two clicks

Dignity for the person. Auditability for the enterprise. One architecture, two virtues.

Dignity

In companies, idea paternity is contested and lost. Success has many fathers. The deck gets presented by whoever is senior. Five years later nobody can say whose pivot it was. A provenance chain settles attribution with receipts-in-time. That is not a soft HR feature. It is how you keep experts willing to let the system surface their work: the system shows the exhibit of what they did, rather than laundering their judgment into anonymous machine prose.

At personal scale the feeling is sharp. You claim pride in a pivot; the walk surfaces the email you did not know still existed. You did not ask for the thing you did not know existed — the system treated your claim as a hypothesis and went looking for corroboration. That is a different tier from search. Org-scale, the same tier becomes idea provenance for every material contribution the firm wants to reuse.

Auditability

Enterprises do not need models that “sound sure.” They need resolvable pointers. Risk committees, legal reviews, quality gates, and regulated clients all want the same primitive: show me where this came from. A wiki that never cites its own guesses as source, and that can open bronze, is governable memory rather than opaque retrieval.

When receipts replace tenure as the way truth is established in a conversation, the culture shifts from “who said so” to “what exhibits.”

Why chat-only RAG fails this test

Similarity chunks can be right and still be untrustworthy, because there is no walkable chain from claim to artefact the human can open. Agents that walk a wiki of relationships can carry provenance as a first-class path. Soft joins across email, CRM, and project stores are what make the bronze path exist in the first place.

Make it a requirement

For every material claim in a marketing card, a sales pack, or a CEO parliament brief, require a double-click path. No path, no claim in the decision pack. That single rule does more for enterprise trust than another accuracy percentage nobody can audit.

Leave circulation dashboards and archive-repricing metrics to the sibling that owns read-path economics. This chapter’s job is narrower and load-bearing: the click path is the trust product.

You do not need another framework name. You need a wall session.

The constant, condensed

Chapter 3 owns the full table. Here is the checklist form:

If a proposed design moves a cell from the right column to the left, it is not an exoskeleton. It is automation wearing a friendly UI.

Morphology design checklist

For any role — including roles beyond the three in this book — answer:

1. Moment: What high-stakes human moment are we wrapping (call, page ship, exec decision, RFP submit)?
2. Prep: What packet would make the human dangerous — in a good way — walking in?
3. Neighbourhoods: Which corpus regions must the walk hit (clients, projects, policies, post-mortems, people)?
4. Human reserve: What must never leave the human (taste, relationship, commitment, political ownership)?
5. Double-click: What is the claim → wiki → bronze path for every material assertion?
6. Expert gate: After retrieval is automated, which judgment questions remain for named experts?
7. Entry bias: What boot profile opens the shared wiki for this role without forking a rival kernel?

Monday moves

1. Name your ladder rung (Chapter 2). If you are still celebrating chatbot citations, you may be stuck on rung two.
2. Pick one morphology and ship it. Marketing Case Library is often the fastest demo; sales join is often the fastest revenue story; CEO parliament is often the fastest executive convert.
3. Retire one “hey Terry” retrieval (Chapter 7). Replace it with a wiki walk plus a judgment question.
4. Require double-click (Chapter 8) for material claims in any pack that influences a decision.
5. Refuse accountability transfer. No design that makes the model the owner of the outcome.

One wiki, many exoskeletons.

What not to build in this programme

Do not expand this into a serial-topology treatise — that sibling already exists. Do not wait for perfect resurrection-rate dashboards before shipping a morphology. Do not fork SMB dental seed GUIs into this workstream if that is not your market. Scope is a feature. The success condition for this book is simple: design role-shaped assists on one substrate without replacing anyone.

Close

You may have built the wiki because the agents needed a map. Keep that. Wire the workflows. Then notice the reveal that pays the bills and changes the culture: the same map fits a human better than it fits a bot alone. Marketing keeps taste. Sales keeps the relationship. The CEO keeps the decision. Experts keep judgment. Machines connect, translate, generalise, and show receipts.

You built the wiki for the AI. It was for the humans.

The bot is the legs. The human still walks into the room.