Skip to content
Sentynel

#technical4 min readBy Sentynel

Define the term before the agent reads it: a glossary with owners

Snowflake now serves agents only approved business definitions. Here is how to write the glossary your own agents need, with one owner per term.

Share

An agent that answers "what was net revenue last quarter?" has to pick a meaning for "net revenue" before it picks a number. If your company has never written that meaning down, the agent will infer one, and two agents will infer two. The fix is a short glossary of official definitions, each with a named owner, that agents read before they act.

What Snowflake announced

On 3 November 2026, at its Expedition event, Snowflake announced Business Ontology, part of the Horizon Catalog. Its press release describes one governed place to record what business terms officially mean, and says agents are served only approved definitions. Teams can approve official definitions for core concepts such as net revenue and link them to the underlying tables and views.

Three details in the release are worth borrowing, whatever platform you use:

  • "Every business definition needs an official version and an accountable owner."
  • When a company migrates definitions from a spreadsheet or another catalog, everything is marked as a draft for the owners to approve.
  • Every change is logged, so an auditor can see what a definition meant at a given time and who approved it.

The release also says human-approved definitions take precedence over what the platform infers. Snowflake lists the feature as arriving in public preview soon, so treat it as an announcement, not a tool you can adopt today. The method below needs no particular product.

Why this is an operations problem

A definition is a decision. "Active customer", "resolved", "qualified lead" and "on time" each carry a rule about who counts and when. People in different teams usually carry slightly different versions in their heads, and the differences stay invisible because humans reconcile them in meetings.

An agent does not attend the meeting. It applies whichever version it finds. The result is a number that looks right and answers a different question from the one the business meant.

Build the glossary in five steps

  1. List the terms your processes already depend on. Take the three reports or decisions the agent will touch and underline every noun that carries a rule.
  2. Write one definition per term. One sentence for the meaning, one for what is explicitly excluded, and the field or table it comes from.
  3. Name an owner. One person, not a team. The owner is whoever answers when the definition is challenged.
  4. Mark everything as a draft until the owner approves it. Unapproved terms are not served to agents. This is Snowflake's migration rule, and it works on a spreadsheet too.
  5. Version every change. Record the old wording, the new wording, the date and who approved it, so a past number can be explained.
From a term in someone's head to something an agent can read
  1. 01Term foundUsed in a report or decision
  2. 02Draft writtenMeaning, exclusions, source
  3. 03Owner approvesOne named person
  4. 04Agent reads itApproved terms only

A changed definition returns to its owner and is versioned.

What goes in each entry

Field Example question it answers
Term What do we call it?
Definition What counts, in one sentence?
Exclusions What looks similar but does not count?
Source Which table, field or system holds it?
Owner Who decides if it changes?
Last approved When, and by whom?

Rules for the agent

  • Approved beats inferred. If the agent can both read an approved definition and guess one from the data, the approved one wins.
  • No definition, no answer. When a term is missing, the agent says so and routes the question to a person, instead of choosing a meaning.
  • Quote the term in the output. An answer that says which definition it used can be checked in seconds.

A Sentynel example

In the Rodo operation, the reported figure is that 82% of live-created enquiries reached Resolved, with a median of about one hour. That sentence only means something because "Resolved" is a stage in the pipeline with a defined place in the flow, and because the figures describe pipeline records, not all conversations. The source for this case is Rodo's operational report of 24 September 2026, and it covers a first measured window with no comparable earlier baseline. Writing the term down is what made the number comparable over time.

Rodo · first measured window
of live-created enquiries resolved
~1 hmedian to resolved

Source: Rodo operational report, 24 September 2026. The cohort covers 31 August–24 September and describes pipeline records, not all conversations.

A test for this week

Pick the one report your team trusts most. Ask three people to write down, without comparing notes, what its headline metric includes. If the three sentences differ, you have found a definition an agent would have to guess.

Write the official version, give it an owner, and make it the first entry in the glossary. Which term in your operation would you least want an agent to interpret on its own?

Sources

  1. https://www.snowflake.com/en/news/press-releases/expedition-2026-business-ontology-ai-gateway/
Share
← All sygnals