In short: At around fifty people, teams form and vocabularies drift. Sales, product and finance use the same words for different things, and AI tools amplify the confusion. The enterprise brain's job becomes meaning: shared definitions, clear decision rights, and one place where records live.
Key points
- Keep a glossary of twenty to thirty commercial terms, each with one definition, an owner, a date and a linked decision.
- Supersede old definitions with a reason instead of editing them away, so agents and analysts know what is current.
- Write one-line decision rights for recurring decisions: who decides, who is consulted, who is informed.
- Start a conflicts list with an owner and a resolution date for each contradiction between teams' records.
- Choose one company-owned home for records and require new AI tools to write or export to it.
With a few teams, each develops its own habits and its own language. "Customer" means a signed contract to finance, an active account to customer success, and anyone in a trial to marketing. "Churn" is measured three ways. Nobody notices until two dashboards disagree in a board meeting.
This is the stage at which large organisations eventually build semantic layers and ontologies. A fifty-person company does not need either. It needs the lightweight version.
A glossary with owners
A short list of the terms that matter commercially, perhaps twenty or thirty. Each has one definition, one owner, the date it was last changed, and a link to the decision that set it. When a definition changes, the old one is superseded with a reason, not edited away.
That last point matters for AI. Agents and assistants that read a glossary need to know which definition is current, and analysts comparing numbers across quarters need to know when a definition changed.
Decision rights
The second thing that breaks at fifty is knowing who decides what. Write it down for the recurring decision classes: discount approvals, hiring, roadmap priorities, vendor choices, exceptions to policy. One line each: who decides, who must be consulted, who is informed. It sounds bureaucratic. It removes more friction than almost anything else at this size.
The first conflicts
With several teams writing records, contradictions appear. Sales records that a customer was promised a feature; product records that it was deprioritised. Start a simple conflicts list, with an owner for each and a date by which it is resolved. It will be short. It will also be the most read part of the record.
The tool sprawl problem
Fifty people is when every team adopts its own AI tools: a meeting assistant here, a writing tool there, a research agent somewhere else. Each keeps its own memory. The organisation's reasoning fragments across half a dozen products, none of which talk to each other and none of which the company controls.
Decide now where records live: one place, owned by the company, that every tool writes to or can be exported into. Tools are replaceable. A fragmented record is very hard to reassemble.
In practice
- Write a glossary of the twenty to thirty terms that matter, each with an owner and a date.
- Supersede definitions instead of editing them.
- Write one-line decision rights for the ten most common decision classes.
- Choose one company-owned home for decision records, and require new AI tools to write to it or export to it.
What would prove this wrong. Fifty-person companies with no shared glossary or decision rights reporting no more definitional disputes or slower decisions than those with them.
Common questions
What breaks when a startup grows to about fifty people?
Teams develop their own vocabularies, so words like customer and churn mean different things, and nobody is sure who decides what. AI tools amplify the confusion.
Does a fifty-person company need a semantic layer or ontology?
No. It needs the lightweight version: a short glossary of key terms with owners and dates, and written decision rights.
How do you stop AI tools fragmenting a company's knowledge?
Decide on one place, owned by the company, where records live, and require every tool to write to it or export into it. Tools are replaceable; a fragmented record is very hard to reassemble.