Tool of the Week: Snowflake's Business Ontology and AI Gateway, and the paperwork agents still need

On November 3, at its Expedition 2026 conference, Snowflake announced two products aimed at the same problem: agents that work against company data when nobody can say what a term meant or what the agent cost. Business Ontology records approved definitions for business terms, links them to tables and views, and serves those definitions to agents. Cortex AI Gateway is a single control point for the connections, models and tools an agent uses.

Status matters, so here it is, per Snowflake's release. Cost governance, observability and Cortex Inference are generally available. Governance for tool calls over MCP is in public preview. Business Ontology and the prompt-injection guardrails are listed as public preview soon. Sandboxes for agent code, abuse detection, trace analysis and dynamic model routing are in private preview. The release gives no pricing and no dates.

Two pieces matter most for work someone would pay for. Cost governance attributes AI spend to users, teams, agents or apps and enforces hard limits. Kargo's VP of Data and AI is quoted by Snowflake saying it showed which teams and agents drove spend and provided hard budget limits. Business Ontology gives each definition an accountable owner and logs changes. Snowflake says it can migrate definitions from third-party catalogs or spreadsheets in a guided session, saved as drafts for the owner to approve.

The reason Snowflake gives: an MIT Technology Review Insights survey of 400 senior data and technology executives, run in July and August, found 55% name a lack of business context as the main cause of poor agent answers, and 45% name fragmented or inaccessible data. Snowflake sells the fix, so read that framing with that in mind.

The takeaway does not depend on Snowflake. Any agent that touches data needs three things written down: what each term means and who owns it, which models and tools it may call, and what it is allowed to spend. Most shops have none of the three. Pulling definitions out of spreadsheets, reports and stored procedures is data work that has to happen before any gateway helps.

Caveat: all capability claims here are Snowflake's, and I found no independent testing. SiliconANGLE's coverage describes Business Ontology as still in private preview, so confirm its status before you plan around it.

Quick Hits

  • Goldman Sachs limits AI coding tools to engineers. Per The Decoder, citing Bloomberg, Goldman has pulled access to Claude Code and GitHub Copilot from traders and bankers, reported November 2. Engineers keep access. A spokesperson cited productivity and risk management and said the experimentation phase is over. In July, the firm's CIO had publicly praised non-technical staff building with AI coding tools. If you work with regulated teams, who may use agentic tools is becoming a policy to write, not a default.

  • ElevenLabs says voice agents beat chat on longer conversations. In a study published November 2, ElevenLabs compared voice and chat agents across production deployments, July through September. In conversations over two minutes, voice succeeded about 32% more often (10.9 percentage points). Successful long voice conversations averaged 4.6 minutes against 7 for chat. In a 98-person simulated flight-rebooking test, 85% preferred voice. Caveats: ElevenLabs sells voice agents, and it uses conversation length as a rough proxy for complexity.

Prompt of the Week: Business Term Definition Sheet

I need to settle the definition of a business term before an AI
agent or report uses it.

Term: [e.g. "active customer"]
Definitions I have found, with where each lives:
1. [definition] (source: [report, spreadsheet, stored procedure])
2. [definition] (source: [...])
3. [definition] (source: [...])

Produce:

1. A one-sentence canonical definition, or say plainly that the
   sources conflict and a person must choose. Do not pick for me
   if the definitions are not reconcilable.
2. A table of how the definitions differ: inclusion rules, time
   window, source system, edge cases.
3. Three test queries in plain English that would return different
   numbers under different definitions.
4. The role (not a name) that should own this definition, and what
   should trigger a review.
5. What I still need to find out. Do not guess missing details.

Run it on the one term your team argues about most. The conflicts it surfaces are the work.

One last thing

Writing this is one side of the work. The other is building it: data pipelines, file and reporting workflows, integrations between systems that were never meant to talk, and the automation that takes the manual steps out.

If a process in your week still depends on someone remembering to run it, book a free 15-minute audit and I will find your top 3 time-wasters. If it is not worth automating, I will tell you that.

Scott