Product Field Work All writing
WritingCompany knowledge

One version of the truth

AI makes it effortless to produce a strategy deck, a training guide or a forty-page plan. That’s the problem. When anyone can create a slightly different version of what the company believes, and agents act on whichever one they were given, the scarce thing isn’t content. It’s agreement.

The problem

Most of what a company knows lives in slide decks, meeting recordings, chat threads and people’s heads. It was always hard to keep aligned. Two things just made it harder. Anyone can now generate a polished forty-page version of the strategy overnight. And agents now do real work based on whatever context they were handed, which means an outdated version of the goals doesn’t just confuse a person; it quietly steers a system.

People also need to understand why the agents do what they do. If employees can’t see the reasoning, they can’t spot when it’s wrong or suggest something better.

40%of U.S. desk workers received AI-generated “workslop” in the past month: polished content without the substance to move work forward
1 hr 56 minlost dealing with each incident, decoding, checking or redoing it
61% vs 40%of a passage remembered a week later by people who practiced recalling it, versus people who just reread it
How I’d measure whether a company is in sync
  1. Share of recurring questions answered by an approved source, not a thread or a guess
  2. Time from an approved change to every person and agent using it
  3. Understanding, measured with short quizzes, not page views
  4. Duplicate content created, which should fall, not rise
Why I wrote it

I’ve been building internal deep-dive articles to help a team learn a product, its customers and its domain, and AI makes that material far better than it used to be. It also made me notice the opposite problem: when everyone can generate their own version, nobody is sure which one is real.

The ideaKeep one approved version of everything the company has decided. Teach it to people in formats they actually retain, in their own language. Sync it automatically to every agent’s skills, and ship every change like a release, with the reasoning attached. Encourage pushback on the approved version rather than new versions of it. And put a governor on new content, so creating another version is a deliberate choice.

1Five decks, one strategy

Anyone who has worked in a growing company knows this week. The CEO stays up late and writes a new strategy deck. Sales has its own version for the board meeting. Someone makes a “final” with slightly different priorities. Each is reasonable. Together they mean nobody is quite sure what was decided.

The same quarter’s strategy, before and after

Switch to see the copies collapse into one approved version, with the audience-specific versions linked to it rather than competing with it.

2The engine needs a map

When people did most of the work by hand, the pace itself was a safety net. There were meetings every week, and a wrong assumption got caught before it went far. Agents remove that net. An engine running on the wrong goal, an outdated persona or an old pricing rule doesn’t drift slowly. It goes the wrong way at full speed, across hundreds of decisions, before anyone notices, and every one of them has to be undone.

That makes approved knowledge more precious than it has ever been. It’s no longer background reading. It’s the steering.

Same destination, three ways of getting there

Illustrative. Distance from the intended direction over a quarter. Press play.

A team working by hand, corrected at weekly meetingsAn engine working from a stale versionAn engine working from approved knowledge

A Ferrari pointed the wrong way doesn’t get you there sooner. It gets you lost faster.

3Atoms and deep dives

Company knowledge comes in two sizes. Deep dives teach: who the customer is, how we sell, how we build, what the market looks like. Atoms decide: one position, one sentence, with an owner, a status and a date. “We don’t offer on-premise deployment.” “Our primary buyer is the operations lead, not IT.” Deep dives are built from atoms, so when an atom changes, the deep dives that use it know.

A searchable library of what’s been decided

Search or pick an entry to see the full atom. Every approved atom has one owner and one current version.

4Publish once, everything follows

Here’s the mechanism that matters most. When an owner approves a change, it shouldn’t be announced in a meeting and hoped for. It should flow, automatically, to every place people learn and every place agents work.

One approved change, followed downstream

The quarter’s top goal changes from “grow new logos” to “expand within existing accounts.” Press publish.

The old way: an all-hands mention, a new deck in a shared drive, and a few months before everyone, and every document, catches up. Agents never would.

5Every change ships like a release

If approved knowledge steers the engine, a change to it is a production release, and it should be handled like one. Nothing should shift quietly overnight so that an agent starts making different choices and nobody can say why. Every change gets release notes: what changed, why, everything it touches, and which skill versions move.

The reasoning travels with the change. People read it in the release note; agents read it in the updated skill. So when an agent does something differently, both the person watching it and the agent itself know what changed and why.

KNOWLEDGE RELEASE2026.09.3 · Q4 goal: expansion firstApproved by the CEO · Effective Monday · Rollback: restore v3

What changed

− Our top goal this quarter is growing new logos. (v3)+ Our top goal this quarter is expansion within existing accounts. (v4)

Why

Expansion revenue grew fastest last quarter, and retention economics beat new-logo acquisition at our stage. Two dissenting comments from sales were answered on the atom.

Who’s affected, and what they’ll get

Sales, customer success, marketing and product. A two-minute update video in six languages, a one-page brief per role, and two quiz questions on days 1, 4 and 15.

Materials updated

  • Deep dive: Q4 goals regenerated, reviewed
  • Sales playbook article 2 sections
  • Onboarding course, module 2 updated
  • Customer-facing success plan template updated
  • 3 videos re-narrated

Skills updated

sales-playbook3.4.0 → 3.5.0Prioritizes expansion plays in account plans
account-research2.1.2 → 2.2.0Flags expansion signals first
marketing-content1.8.0 → 1.9.0Customer stories over acquisition campaigns
product-context4.0.1 → 4.1.0Goal and bets updated

Each skill re-ran its evaluations before release. Each carries a short note explaining the change, so agents apply the reasoning, not just the rule.

6A governor, and a gap detector

The same logic behind a governor on AI spending, and a budget on how much product change customers can absorb, applies to internal content. We don’t need more of it. We need the right version of it.

A knowledge server that every agent can reach, using the Model Context Protocol, can sit quietly behind anything anyone creates. When a question matches approved knowledge, it guides people to that version. When a new deck would duplicate an approved one, it says so.

Just as valuable is what it learns from the questions it can’t answer. Every unanswered question is a signal. When the same one keeps coming up, from people or from agents asking for clarification, it becomes a request for new knowledge, routed to the owner who should decide it.

What the knowledge server does, in three moments

Pick a request. The responses come from the approved library, not from a model’s guess.

Signals that knowledge is missing
  • The same question asked repeatedly, across teams
  • Agents pausing to ask for clarification on the same topic
  • Two new drafts that contradict each other
  • Customers noticing inconsistent answers

Creating a new version should be a decision, not an accident. “Why does this need to exist, and does it change the original?”

7Learning that sticks

Publishing isn’t the same as people knowing. Everyone is busy, and a long document is easy to skip. AI helps here too: the same approved knowledge can become a short article, a narrated walkthrough in each person’s native language, an interactive explainer, a quiz and a one-page brief for each role, without anyone rewriting it five times.

There are two kinds of training, and both matter. Onboarding teaches the whole picture: customers, product, strategy, how we sell and build. Updates teach what changed and why, in two minutes, every time a knowledge release ships.

The research on memory is clear about one thing. Rereading feels productive and fades quickly. Being asked to recall something, a little later, and again a little later after that, is what makes it stay.

The right format for each kind of knowledge

Highlighted formats are the primary one for each. Every piece also gets a short quiz and a “what this means for your role” brief.

Remembered one week later
Reread it40%
Practiced recall61%

Roediger and Karpicke, 2006.

Spaced check-ins after a change
Day 1
Day 4
Day 15
Day 45

Two questions each, in the tools people already use.

8Pushback is the job

Approved knowledge is a living thing, and keeping it right is everyone’s job. A locked version isn’t a gag order; it’s the thing to push against. If something is outdated, wrong, or overtaken by new information, people are expected to say so.

What changes is how. Not by ignoring the approved version and writing a new forty-page one, but by pointing at the specific part you disagree with and saying why, or how to make it better. The owner answers. The approved version either changes for everyone, through a knowledge release, or the reasoning for keeping it is written down where everyone can see it.

1Learn itFrom the deep dive, video or brief, in your language.
2Push back on itComment on the specific part, with evidence or new information.
3Owner decidesChange it, or explain why not, within a set time.
4Release itA new version with release notes; the old one is marked superseded, not deleted.
5Everything followsPeople get a short update; skills get new versions and the reasoning.
Executives own directionGoals, strategy and the big bets. They approve those atoms and don’t publish competing decks.
Each function owns its domainProduct owns personas and roadmap; sales owns how we sell; engineering owns how we build; experts own the domain rules.
One team keeps it healthyAn editor role, often product or operations, watches for gaps, stale atoms and duplicates, and runs the weekly review.
CoverageRecurring questions answered by an approved atom
Sync timeApproval to every person and agent using it
UnderstandingQuiz results by team, not page views
DuplicatesNew decks or docs that restate an approved version
FreshnessAtoms past their review date, and pushback waiting on an owner
Repeat questions“Didn’t we decide this?” moments, which should approach zero

Choices

Atoms before documentsDecisions are small, owned and versioned. Documents are assembled from them, so they can’t drift apart.
People and agents from one sourceThe same approved knowledge teaches the team and feeds every skill, so what people learn and what agents do can’t disagree.
Knowledge changes are releasesVersioned, explained and scoped, so no one wonders why an agent started doing something different.
Where this can go wrong

Too much central control turns a knowledge base into a bureaucracy, and people route around it. Approval has to be fast and owners have to answer pushback, or “locked” starts to mean “ignored.” Generated videos and quizzes can be confidently wrong, so the owner reviews what’s generated from their atom. Not everything should be locked; early thinking needs room to be messy. And a server that watches what people create has to be clearly about helping them, never about surveillance.

Sources

  1. BetterUp Labs and Stanford Social Media Lab, “workslop” research, reported in Harvard Business Review, 2025: 40% received workslop in the past month; 1 hour 56 minutes lost per incident; many said leadership encouraged AI use without explaining how or why.
  2. Summary of Roediger and Karpicke, 2006: 61% versus 40% recall after one week for retrieval practice versus rereading.
  3. EBSCO Research Starters, “Testing effect”: the original experiments and delayed-recall results.
  4. Retrieval practice and spacing: why check-ins spread over days and weeks outperform one massed session.
  5. Anthropic, “Introducing the Model Context Protocol”: an open standard for connecting agents to the systems where knowledge lives.
  6. Anthropic Engineering, Agent Skills: packaging knowledge so agents load it only when a task needs it.