Company Glossary for AI Agents: Teaching Your Internal Terms
Learn how to write a short, owned company glossary for AI agents: entry fields, a step-by-step first version, ownership rules and a test before go-live.

A company glossary for an AI agent is a short, owned list of the internal terms, product names, abbreviations and status labels your business uses, each with one plain definition, the approved wording and a named owner who can change it. Write it before the agent handles live work, keep it to the terms that change what the agent does, and load it as memory the agent reads before it interprets a request.
Key takeaways
- An agent without a definition for an internal term usually guesses, and the guess reads like a correct answer.
- Start with the terms that trigger wrong actions, not with a full dictionary of the company.
- Every entry needs a definition, a "is not" line, the approved customer-facing label and one named owner.
- Store the glossary as reviewed memory that changes through a logged process, not through ad hoc chat corrections.
- Test the glossary with real requests before any agent acts on what it says.
Why agents guess at internal terms
A general language model knows the public meaning of most words. It does not know that your dispatch team calls a rush batch "the priority run," that "hold" means billing is paused rather than stock is reserved, or that an internal nickname refers to one scheduling module. When a term is missing, the agent rarely says so. It picks the most likely meaning from the surrounding text and keeps going, and the answer sounds fluent enough that nobody questions it until a customer or an employee acts on it.
This is why a glossary is a different asset from a document library. Documents hold procedures and policies. The glossary holds meaning, and it is the layer that tells the agent how to read those procedures. If a policy says "release the hold after verification," the agent needs to know what counts as a hold and what verification means in your company before it can act. When the gap is a question the documents cannot answer at all, record it so the owner can close it, as described in AI Agent Knowledge Gaps: How to Log Unanswered Questions.
Anthropic's engineering notes on context engineering make a related point: an agent works from the context it is given, so the context must be chosen deliberately rather than left to accumulate. A glossary is one of the cheapest ways to make that context precise.
What each glossary entry must contain
Keep every entry short enough that an owner can review it in a minute. A useful entry has five parts:
- The term exactly as staff write it, including capitalization and the misspellings that appear in tickets and messages.
- A definition of one or two sentences, written for someone who joined the company last week.
- A "is not" line that separates the term from its closest neighbor, such as "Return is the physical equipment coming back, not a warranty replacement."
- The approved label for customer replies, and any internal-only label that must never appear in a message to a customer.
- One named owner who approves changes and answers when two people disagree about the meaning.
Add a version marker to each entry so that a changed definition can be traced and an older one can be retired without losing the history. If an entry starts to need a page of explanation, that explanation probably belongs in a procedure document, and the glossary entry should name that document instead of repeating it. For the broader structure of how a glossary sits next to other company knowledge, see Company Glossary for AI Agents: Terms Agents Must Use.
A plan to write the first version
The first version should take a small group a few working sessions, not a quarter. Use this sequence:
- Collect candidate terms from the last few weeks of support replies, handoff notes, order records and internal chat where a new colleague asked what something meant.
- Cut the list to terms that change an action: a status that triggers a workflow, a product module that routes a request, an abbreviation that appears in customer messages, or a word that two teams use differently.
- Ask the owner of each area to write the definition and the "is not" line in their own words, in no more than two sentences each.
- Mark which labels customers may see and which stay internal, and confirm that with the person who owns customer communication.
- Load the approved entries as company memory with the owner and version recorded, and mark them as reviewed so the agent can tell them apart from drafts.
- Run a set of real requests that use these terms. Check that the agent uses the approved meaning, and that it asks a clarifying question when a term it has not been given appears.
- Publish a short change rule: anyone can propose a change, the owner approves it, and the approved change is logged with the reason.
The last step is the one most small teams skip, and it is the one that keeps the glossary true after the first month.
An illustrative scenario: one word, two meanings
*Illustrative scenario, not a real client:* a small equipment rental business uses "return" in two ways. Staff use it for a customer sending equipment back to the depot, and they also use it for a warranty replacement that is handled through the same form. An agent without a glossary receives a customer message that says "I need to make a return" and answers with the refund policy, which applies only to unused equipment that is still within the rental window. The customer has a faulty machine and is now confused about why the answer does not fit.
After the owners add two entries, the picture changes. "Return" means equipment physically coming back to the depot. "Warranty replacement" is a separate term with its own form and queue. When a message is ambiguous, the agent asks which one the customer means, and it routes warranty cases to the right team without quoting the refund policy. The fix took two entries and one test request, and the test request is what proves the fix holds.
Keeping the glossary owned and small
A glossary decays in two ways: it grows with duplicate entries, and its definitions drift away from how the business actually works. Both are controlled by ownership and review rather than by better intentions. The AI Knowledge Base Maintenance: Owners and a Review Routine guide describes how to set that rhythm for the wider knowledge base, and the same rules apply here. The NIST AI Risk Management Framework treats clear roles and documented oversight as part of managing AI systems, and a glossary with named owners is a small, practical version of that idea.
A short checklist keeps the review manageable:
- Every entry has exactly one owner and one approving role.
- No two entries define the same thing. When they do, merge them and keep the approved label.
- Retired terms stay visible as retired, with the replacement named, so the agent does not revive an old meaning.
- Customer names, personal details and staff opinions stay out of the glossary. They belong in records with their own rules.
- Each new term passes the same test as the first version: it changes an action, and an owner has written its definition.
- A review happens on a fixed rhythm the owner agrees to, and also whenever a term causes a wrong answer in a test or in production.
When a wrong answer traces back to a term, treat the glossary change as the fix and check it against the same requests that exposed the problem.
How Be Human's managed approach handles it
In Be Human's managed approach, the glossary is part of the memory design of the first engagement, alongside the role, the tools and the approval boundaries. We work with the owners of each area to draft entries in their own words, mark which labels may reach customers, and load the approved entries with owners and versions recorded. Recurring evaluations then check that the agent uses approved meanings and asks when a term is unknown. Permission review and logged changes keep edits controlled after go-live, and incident response covers the case where a wrong meaning reaches a customer.
If you are not sure which internal term would cause the most wrong answers first, the free Be Human AI diagnostic will help you find the first task worth handing to an agent.
Frequently asked questions
How long should a company glossary for an AI agent be? A useful first version is short, usually a few dozen entries that cover the terms behind real decisions and customer replies. Length is less important than coverage of the terms that change an action. If the list keeps growing, check for duplicates and for entries that belong in a procedure document instead.
Who should write the definitions? The people who own each area should write them, because they know where a term is used and where it is confused with a neighbor. An editor or operations lead can collect the candidates and check the wording for clarity, but the meaning should come from the owner. A definition written by someone outside the area tends to be accurate in general and wrong in the details that matter.
Should the glossary replace the agent's documents? No. The glossary explains what words mean, and documents explain how to do things and what the rules are. The agent needs both, and the glossary helps it read the documents correctly. Where a glossary entry needs a long explanation, it should point to the document by name instead of repeating it.
How often should the glossary be reviewed? Review on a fixed rhythm that the owners agree to, and also whenever a term causes a wrong answer during testing or in production. The rhythm matters less than the rule that changes are proposed, approved by the owner and logged. A glossary that is reviewed only when someone remembers it will drift within a few months of real use.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


