When an AI agent writes SQL for you, the scary part isn’t the syntax. Models are good at syntax. The scary part is that the agent has to guess what your words mean.
A client of mine, a 14-person logistics firm in Hamburg, ran straight into this last year. Two dashboards, one database, and they disagreed on open receivables by about €40,000 every single month. It took us an afternoon to find the cause. One query counted an invoice as open until its due date; the other counted it as open until payment actually landed. Both numbers looked plausible on a screen. Nobody was careless. The definitions just weren’t written down anywhere, so every tool, human or AI, quietly made up its own.
That’s roughly the problem NVIDIA took a swing at this week. If you’re already letting AI agents run queries against your business data, or you’re close to it, this release deserves ten minutes of your attention.
What shipped
On November 2, 2026, NVIDIA released Auto Ontology, an open source library under the Apache 2.0 license. It generates a semantic layer from enterprise metadata. That layer grounds the SQL your AI agents write in certified definitions, so a plain-language question resolves to the definition you certified rather than whatever the agent infers from column names. Auto Ontology serves this context through MCP, a CLI, and an API.
Why SMBs should care
Small businesses hit this problem in admin work first. Receivables. Month-end reporting. “How many orders shipped last week?” The moment you point an agent at your database and ask in plain language, it starts inferring what open, active or net means from your schema. Sometimes it infers correctly.
The expensive case is when it’s confidently wrong and the number lands in a report your accountant, your bank or a customer actually reads. The honest workaround today is double-checking: someone on your team verifies every agent-built answer against a query a human trusts. That’s 20 to 40 minutes of admin time per report, and it quietly eats most of the speed you wanted from the agent in the first place.
A semantic layer moves that checking up front. Certify the definitions once, and every question reuses them. In theory, that’s the difference between an agent you babysit and one you just run. If you’re not doing text-to-SQL yet, treat this as a “watch,” not a “do.” And I’ll admit a limit up front: I don’t know how well the auto-generation behaves when metadata is a mess, which for a lot of small businesses it is. That’s the first thing I’d test. Grounding is really a reliability question, the same territory we cover in workflow reliability engineering.
How to take advantage
Three moves for this week, all low-risk:
- 1. Write down what your words mean before touching the tool. Pick the ten to fifteen terms your admin reports lean on: open invoice, active customer, net revenue, overdue. One or two lines each, plus which table and column backs them. The result: you’ve produced the metadata Auto Ontology builds its layer from, and you’ve done the one job no software can do for you. The announcement doesn’t detail which metadata formats it ingests, so check NVIDIA’s official docs before assuming yours fits.
- 2. Run a first look against a copy of your data, never production. Install the library, generate the semantic layer, and expose it to an agent through MCP. The result: your agent’s next queries get served your certified definitions instead of guessing at raw schema. Setup walkthroughs weren’t part of the announcement, so pull the install steps from NVIDIA’s docs rather than from a random tutorial.
- 3. A/B one report for a week. Pick a recurring admin report, let the agent rebuild it with the semantic layer active, and keep producing it the old way in parallel. The result: you’ll learn whether your certified definitions actually cover the questions your team asks, before anyone stops double-checking anything.
Cost and effort
The license costs €0 per month. Apache 2.0 means you can use it without paying anyone, and no paid tiers, hosting or support offerings were announced, so I won’t invent any. What it aims to remove is the verification pass after every agent-built query, plus the chore of hand-writing query definitions for agents in the first place.
The real price is time. With metadata in reasonable shape, a technical person can probably get a first look running in half a day. Writing and certifying the definitions adds another day or two, spread across whoever owns your numbers. No one technical in-house? Budget a few external hours on top. For the wider picture of what SMB automation costs once you add everything up, we keep an AI automation budget breakdown for 2026 up to date.
One honest pitfall
The word doing all the heavy lifting here is certified. Auto Ontology can generate a layer, but only a human who knows the business can certify that the definitions are true. Skip that step and you’ve built a machine that produces confidently wrong numbers with an official-looking stamp on them, which is worse than a wrong number people already distrust.
Then there’s decay. Every new field, every process change, every renamed table needs the layer updated, or it drifts out of sync with reality without telling anyone. In the Hamburg story, the fix was never software. It was two people agreeing on what “open” means. Software can enforce that agreement. It can’t create it. My other gripe: it ships as a library you drive through a CLI and an API, which tells you it assumes a technical comfort many small teams don’t have yet.
If you want a hand with this
If untangling definitions, agents and metadata alone sounds like a bad evening, book a consultation with us and we’ll look at your setup together, no slides, I promise.
Related reading
- AI Automation for SMBs: the no-hype playbook
- Automation platforms compared: Zapier, Make and n8n
- About Damian, who writes this blog
Talk soon,
Damian
P.S. If you actually try Auto Ontology this week, tell me how the metadata import went. Genuinely. I still don’t know.
