The Cyborg CEO · An educational deep dive
Give your company a brain.
Before AI can speak for your company, something has to reconcile what you say about yourself with what you actually do. This page explains how that works: the research behind it, the disciplines it borrows from, and how to run the core loop yourself with no special tooling.
The problem has a name
Every company runs on three versions of itself
Organisational researchers Chris Argyris and Donald Schön named this gap fifty years ago: every organisation has an espoused theory, the story it tells about what it does and who it serves, and a theory in use, the one its behaviour actually reveals. Their finding was blunt. The gap between the two is normal, largely invisible from the inside, and the main reason organisations fail to learn.
What you say
The brand guide, the positioning, the ICP on the strategy slide. The espoused theory: the company as leadership describes it.
What you publish
The website, the blog, the collateral. Your published content is your theory in use: it shows who you actually write for.
What actually happens
The sales calls, the search queries, the deals you win and lose. Reality keeps its own record, whether anyone reads it or not.
Positioning expert April Dunford adds the time dimension: positioning decays by default. Most companies keep the story from the day they were founded long after the product, the customers, and the competition have moved. Drift is not a failure of discipline. It is the default state, which is why it has to be detected, not just avoided.
Why ingestion is not understanding
A pile of documents is not a brain
The obvious move is to load every company document into an AI and call it done. The research says otherwise, and it is worth understanding why before you trust any system with your company's identity.
Retrieval is a lookup
Standard retrieval-augmented AI finds passages that look similar to a question and pastes them into the answer. Microsoft Research documented that this fails on questions that need facts connected across documents, or a holistic view of the whole corpus. Knowing your company is exactly that kind of question.
Conflicts get picked silently
When two sources disagree, models tend to pick one side without telling anyone. Researchers call this a knowledge conflict, and a pile of unreconciled company documents produces them constantly. A Stanford study of premium legal AI tools, all retrieval-based, still found wrong answers 17 to 33 percent of the time.
Authority is not in the text
No model can tell which of two conflicting documents is the current one, because authority lives in governance: dates, ownership, approval, what supersedes what. That decision is organisational. Someone has to make it once, record it, and let everything else inherit it.
This problem is older than AI, and so is the fix. Data teams have spent decades building what they call a golden record: one reconciled, governed version of each business fact, assembled by matching conflicting records, applying agreed rules about which source wins, and escalating what the rules cannot settle to a human steward. A company brain is that discipline applied to identity. And the data-centric AI movement made the underlying rule plain: a model can only be as consistent as the data you show it. Improving the corpus usually beats upgrading the model.
How it works
Three layers, one agreed understanding
01 · The foundation layer
Start with who you think you are
The brand guide, the website, the blog, the collateral: everything the company has written about itself. This layer is deliberately your version of the story, because you cannot correct a story you never wrote down.
Content strategists have a name for the artifact this layer should distil into: a message architecture. Not customer-facing copy, but a short ranked hierarchy of what the company claims about itself, in priority order. The ranking is the point. When two claims compete for the same headline, the hierarchy has already decided which one wins, and every claim is supposed to carry a proof point underneath it. A claim without proof is not optional garnish; in messaging practice it is treated as unfinished work.
Feeding the foundation
Output: a ranked, evidenced statement of who you say you are.
The challenge loop
The brain
You describe your ICP as mid-market. Nine of your last twelve articles are written for enterprise buyers. Which one is true?
The ruling persists, with the evidence and the reasoning attached.
02 · The challenge loop
Then it pushes back
This is the layer with real science behind it. Machine learning calls it active learning: a system that chooses its own questions, asking the human about the case it is least certain of, reaches the same accuracy with far less human effort. And the best question is the one where the evidence disagrees with itself. Disagreement is not noise to smooth over. It is the signal that locates the most informative question.
Two more findings shape the design. Anthropic research on sycophancy showed that systems rewarded for human approval learn to agree, not to be right, so the loop must reward resolved contradictions, not pleasant answers. And a classic 1960 psychology experiment by Peter Wason showed people overwhelmingly test their beliefs with cases they expect to confirm them. Humans do not naturally hunt for their own contradictions. The loop institutionalises the test people skip.
One rule is non-negotiable: your ruling must persist. In active learning, the human's answer becomes ground truth the system retrains on. A loop that forgets its adjudications re-asks the same question forever, and burns the one resource it cannot waste: your attention.
03 · The living layer
Who you are changes. The brain keeps up.
A brand guide is a photograph. A company is a moving picture. Dunford names three triggers that quietly invalidate a stated identity: your product changes, your competitors change, or the market changes. All three fire continuously, which is why the understanding needs a live feed, not an annual review.
The feed is not exotic. Sales conversations, turned into countable themes. Search data you already own. Win/loss evidence from real buyers, collected close to the decision, because memory compresses a deal into a tidy story within weeks. Each signal either confirms the agreed understanding or opens a new challenge.
One warning from the same research: the default outcome of a review should be no change. Positioning takes a long time to establish, and a brain that reshapes the company's identity every quarter is thrash wearing a lab coat.
Live signal feed
Each signal confirms the understanding or challenges it again.
What to collect, and what it actually tells you
| Signal | What it tells you |
|---|---|
| Competitor names mentioned on calls | The shortlist buyers actually compare you against, which is your real market category, not the one in the deck |
| Recurring objection phrases | Which of your promises is not landing, in the exact words it fails in |
| Customer sentences, captured verbatim | Message and headline candidates. Paraphrasing strips the phrasing that makes copy convert |
| Buyer-stated reasons vs CRM reasons | Where the organisation is telling itself a story. Reps and buyers routinely disagree about why deals close |
| Search queries rising in impressions | Demand forming, in the phrasing the market actually uses. Exclude your brand name or it drowns the signal |
| Queries ranking positions 8 to 20 | Topics where you are almost credible. Updating those pages beats writing new ones |
| Win rate and retention by segment | The customer pattern you are actually built for, whatever the ICP slide says |
Grounding worth knowing: classic voice-of-customer research by Griffin and Hauser found 20 to 30 one-on-one customer interviews surface roughly 90 percent of the needs in a category, and that a single analyst reading the transcripts misses needs that multiple readers catch. Small, disciplined samples beat big lazy ones.
Where it pays off
The gate at the end of the loop
Serious publications do not trust the writer's assurance. In the magazine fact-checking model, a separate checker verifies a near-final draft against the source material, which must travel with the piece. Technical writing teams do the mechanical half with style guides encoded as automated rules, run on every draft.
A company brain gives your content the same two gates. The mechanical one: does this draft match the agreed voice and vocabulary? The editorial one: is it aimed at the audience you agreed, making claims your evidence supports? When a draft reads enterprise and you agreed mid-market, the gate says so before your audience notices, instead of after.
No tooling required
Run the challenge loop yourself, this week
The core of the brain is a comparison you can do manually. Both audits below are established practice; together they put your espoused theory and your theory in use on the same table. Budget an afternoon for each.
Audit one: who you actually win
- 1 Write the stated ICP as specific, checkable attributes: industry, company size, buyer title, trigger event.
- 2 Pull the last 20 to 30 closed-won deals from the CRM.
- 3 Score every deal against each attribute. Record the champion role, cycle length, and retention where known.
- 4 Ask: which segment closes fastest and stays longest? Does it match the slide?
- 5 If reality and the slide diverge, decide deliberately: retarget the execution, or rewrite the ICP. Both are legitimate. Drifting silently is not.
Audit two: who you actually talk to
- 1 Take the same stated ICP and value themes as the reference.
- 2 Inventory your last 20 published pieces: articles, talks, posts. Fixed window, no cherry-picking.
- 3 Tag each piece with the audience it actually addresses and the value theme it supports, or none.
- 4 Compute the distribution. What share of your output speaks to the ICP you claim?
- 5 Put both audits side by side. The gaps you find are the exact contradictions a company brain would have raised. Resolve each one: confirm the strategy or rebut it. Write the rulings down.
Argyris called the deeper move double-loop learning: do not just correct the behaviour to match the stated theory. Be willing to question whether the stated theory itself is wrong. Fixing the content while the ICP is stale is polishing the wrong artifact.
These two audits are Phases 03 and 04 of the full build
The complete guide takes you from a pile of documents to a versioned, agreed, evidenced Brain file: ten phases, each with the exact prompt to hand to Claude, and a monthly loop that keeps it honest.
Failure modes
Where this goes wrong
Every one of these is documented somewhere expensive. Knowing them is most of the defence.
The forgetting loop
Rulings that do not persist get re-asked forever. Data stewardship records every adjudication on the golden record for exactly this reason. A brain that forgets its decisions destroys trust faster than having no brain at all.
Rewarding agreement
If the success signal is the human accepting the output, the system learns to produce acceptable outputs, not true ones. Score the loop on whether contradictions were real and resolved, never on whether the human was pleased.
Querying everything
Surface every minor inconsistency and the review queue overflows, so humans start rubber-stamping. The whole point of uncertainty targeting is that most examples are not worth your attention.
Decorative citations
A real-looking citation is not a supporting citation. The Stanford legal study found production tools attaching genuine sources to claims those sources did not back. The test is simple: according to this source, does the claim hold?
Version blindness
Without effective dates and supersession, the 2021 positioning and the 2025 positioning are just two similar documents, and retrieval has no basis for preferring the current one. Dating and superseding are governance, not decoration.
Thrash in a lab coat
A living layer is not a licence to reposition quarterly. Establishing a position takes far longer than a quarter, so the recurring review should usually conclude: no change. The brain earns trust by mostly agreeing that you are who you said.
What a company brain is not
Not a chatbot with your logo. Retrieval over a pile of documents indexes the pile, contradictions included. Reconciliation, the golden-record work, is what turns a pile into an understanding, and no amount of model quality substitutes for it.
Not a content mill. Volume of wrong words is worse than silence. The gate exists to make what you publish more true, not more frequent.
Not a replacement for judgment. The machine finds the contradictions; the human makes the ruling. That division is not a compromise. Active learning exists because expert attention is the scarcest input in the system, and the design question is where to spend it.
Go deeper
Everything on this page traces to published work. These are the sources worth your time, in rough reading order.
-
Obviously Awesome, April Dunford · the positioning method behind best-fit customers, drift triggers, and the case against thrash
-
Theory in Practice, Chris Argyris and Donald Schön · espoused theory vs theory in use, and double-loop learning
-
Active Learning Literature Survey, Burr Settles · why the machine should choose the question, formally
-
Towards Understanding Sycophancy in Language Models, Sharma et al. · the evidence that approval-trained systems learn to agree
-
Hallucination-Free? Stanford RegLab · what retrieval does and does not fix in premium production tools
-
The Voice of the Customer, Griffin and Hauser · the interview-count math behind disciplined customer listening
-
Content Strategy at Work, Margot Bloomstein · message architecture: the ranked hierarchy that settles disputes
-
Docs as Code, Write the Docs · one canonical source, everything derived, divergence treated as a bug
This page is part of how Darren Goonawardana thinks about leading with AI. The companion essay is What Is a Cyborg CEO?, and questions or disagreements are welcome here.