Your AI needs the business dictionary

Good morning! Ask three teams for “active customers” and you may get three perfectly confident answers. Give their spreadsheets to an AI agent and the disagreement can travel much faster. Today’s lead is about giving that agent a shared dictionary.

Three highlights:

Plus, three data tools, a five-term glossary exercise and a reusable prompt. Finally, a meeting where arguing about definitions is actually the assignment.

At its November 3 Expedition event, Snowflake announced Business Ontology and Cortex AI Gateway capabilities intended to connect shared business meaning with agent controls.

Business Ontology is designed to give core concepts approved definitions and accountable owners, linked to the underlying data. Its availability is public preview soon, according to the release’s footnotes. It is not generally available today.

Cortex AI Gateway’s inference, observability and cost-governance capabilities are marked generally available. Gateway Tools is in public preview; several other capabilities, including agent abuse detection and dynamic model routing, are in private preview.

For founders and operators, the central problem is familiar. “Revenue” might mean invoices issued, payments collected or revenue recognized. “Active” might mean a login, a purchase or an ongoing subscription. The right answer depends on the question and the agreed definition.

Putting an agent in front of this ambiguity does not resolve it. A useful starting point is a small glossary that records each term’s meaning, calculation, exclusions, time period and owner. Link the definition to the actual source that implements it.

For example, a customer-success team could define an active account by a specific product event during a rolling period, while finance uses a different, explicitly named subscription measure. Distinct names make those differences visible rather than forcing one misleading number.

For data and IT teams, ask how approved definitions reach dashboards, analyst queries and agents, and how changes are recorded. For business teams, appoint the person who can settle a dispute.

Hume’s November 2 announcement introduces Audio Decision API: provide audio, questions and answer choices, then receive probabilities. Examples include conversation dynamics and background sounds. The announcement directs users to a playground and API-access waitlist. Perceptual judgments such as frustration reflect interpretation, not direct knowledge of somebody’s internal state. Any performance comparisons in the launch are Hume’s own results.

Anthropic’s November 2 research release introduces FlashProve, an open-source prototype for cryptographically checking that a committed model produced particular output. The proof aims to verify the computation without revealing the weights. This does not prove an answer is true. The research has important limits, including short contexts and an integer model with greedy decoding; it is not a finished verification feature for everyday chat.

OpenAI announced the program on November 2 to develop arrangements for future compute campuses to reduce electricity demand when grids are strained. Georgia is the starting point. Its planned Effingham County campus is still in development. The practical distinction: these are commitments and future operating arrangements, not evidence that the campus is already providing grid support.

Established tools for making definitions inspectable. These are not new product launches.

Catalog data assets, ownership and business context so people can find the definition behind a number. Useful for data teams maintaining a shared business glossary. The open-source project can be self-hosted; Collate offers a separate managed service with a free tier and paid Premium and Enterprise options. Project; managed-service plans.

Run analytical SQL against a small approved dataset to see whether a proposed business definition produces the expected result. Useful for analysts comparing inclusion rules and date boundaries. DuckDB is free under the MIT license and runs within a host process by default; it does not require a separate database server. Official overview and license.

Explore database tables and queries with a graphical interface. Useful for an analyst checking which fields implement a business definition. Community Edition is free and open source; PRO is commercial and adds capabilities. Use an approved connection and appropriate permissions when inspecting business data. Community and PRO comparison.

Pick one recurring report that people actually use, then follow these five steps.

The output can be a simple table. Its value comes from the agreement behind it, not the software containing it. A glossary that nobody owns is just another document waiting to become historical fiction.

Announced: November 3. Money raised: $43 million. Round stage: undisclosed. Valuation: undisclosed.

The company’s syndicated financing release names participants including QBE, Mucker Capital, Motivate Ventures and State Farm Ventures. Lazarus develops enterprise AI for documents, cybersecurity and legacy modernization, with deployment options including customer infrastructure. It says the money will support partnerships, research and product development.

Our read: deployment control is becoming part of the product conversation. Buyers should still examine model-change approvals, evidence, support obligations and total operating costs. A financing announcement establishes funding; it does not independently validate the vendor’s accuracy or cost claims.

Two correct calculations can still be answering two different questions.

Today’s fictional cartoon is dedicated to the person who asks “what do we mean by revenue?” just as everyone thinks the meeting is finished. They may be doing the most useful work in the room.

Which business term causes the most confusion in your team? Reply with one example, and share this briefing with the colleague who owns that number.