|
On August 28 Databricks announced a batch of Genie One additions: a desktop app you can call up from anywhere on your computer, and drafting and commenting on reports inside the platform, so the numbers never detour through screenshots. It also takes in more context now, through file uploads and Genie Ontology, the layer that collects your company’s business rules and metric definitions. Databricks has been pitching exactly this to finance teams; an August 25 post shows capital-markets finance tracing funding and profitability shifts through Genie One. Put plainly: the AI’s answer to “how much did we spend on X” is now something a customer relies on or a CFO forwards. Which raises the stakes, because a money answer is binary. There’s no partially right.
The stakes already have a price tag. In February 2024 a Canadian tribunal decided Moffatt v. Air Canada: the airline’s website chatbot invented refund terms that didn’t exist, on a page that linked to the real policy. Air Canada’s defense, as the tribunal read it, was that the chatbot was a separate entity responsible for its own words. The tribunal called that “a remarkable submission,” ruled the company must stand behind every representation on its site, and ordered it to pay CAD 812.02. The payout is lunch money. The doctrine is not: your AI’s answer is your company’s answer. Regulators have kept pulling the same thread. The CFPB, America’s consumer-finance watchdog, warned in 2023 that inaccurate chatbot answers about money can violate consumer financial law. And FINRA, the US brokerage regulator, shows in its 2026 oversight report what regulated firms actually do about it: they point generative AI at internal work first. The same report names AI acting without human sign-off as a risk in its own right.
So can you trust the number at all? Here’s the thing: Databricks has quietly shipped a full toolkit for this. It just never presents the pieces as one system. We went through the docs and counted eight instruments, each verified against the current documentation this week.
1. Metric views. The metric is defined once, in Unity Catalog, the platform’s central library of data definitions, and the dashboard tile and the Genie answer read from that same governed definition. This is the single highest-value move: “spend” stops meaning four different things in four teams’ reports. The catch is that nothing forces a question through it; anyone querying the underlying tables directly goes around the whole idea.
2. Certification tags and lineage. Unity Catalog has an official certification stamp for tables, views, dashboards and Genie Agents that only designated curators can set, plus lineage, an automatic trace of where every column came from. A label and a paper trail, not an enforcement. But the label is what your curators bless and the trail is how you prove what fed an answer.
3. Trusted assets. Curators register pre-approved query templates and functions; when a question is answered through one, Genie marks the response as a verified answer. This is the closest thing to “the AI may only quote approved arithmetic.” The catch matters though: that verified label is documented for the Genie Agent’s own chat. The docs are silent on whether it survives agent mode, a multi-agent setup like Agent Bricks, or a Genie One conversation. The trust label is least documented exactly where the product line is heading.
4. Benchmarks. Each Genie Agent holds up to 500 test questions where you supply the correct answer query yourself, and grading compares the actual numbers returned, forgiving only row order and rounding past four significant digits. Deterministic, exactly what a money answer needs. Except in agent mode, where another AI model grades the answers instead. Sit with that for a second: the mode with the most freedom gets the least deterministic grader.
5. Inspect. In preview since March: Genie double-checks the query it just wrote by running smaller verification queries against it, and rewrites it if they disagree. Genuinely clever. Also still the model checking the model’s homework.
6. MLflow 3 evaluation. MLflow is Databricks’ testing and monitoring framework for AI: suites of questions with known-correct answers run automatically before each release, and the same checks can spot-check live answers once you are in production. The gap: there is no documented way to drive Genie’s own 500-question test bank from that machinery, so connecting the two is a workaround, not a product path.
7. The review loop. Every answer carries “Is this correct?” with an escalation to a curator, plus a monitoring tab and a weekly digest. Read the docs closely and you find the honest sentence: behavior does not change based on user feedback alone. Thumbs inform a human. They don’t train anything.
8. The audit trail. Genie events land in the account’s audit log, and the actual query behind every answer lands in query history with who ran it and when. When a dispute arrives, you can reconstruct who asked, what ran, and what data it touched. That is precisely the evidence Air Canada didn’t have.
Now the uncomfortable part. Databricks’ own engineers say a Genie agent should pass at least four of every five test questions before real users touch it, and their case study needed five rounds of tuning to get there. The launch marketing quotes a rosier score, but it comes from a small internal test, so file it under vendor confidence. The gap between demos and reality is well measured, too: when researchers rebuilt the clean academic exam from real corporate databases, the very model that aced the original managed roughly one right answer in ten. For a question about money, even the passing bar leaves one confidently wrong answer in five.
Our conclusion after a week in these docs: trust is not a property of the model. It’s a property of the path the answer took. A number you can stand behind travels a governed metric, a certified asset, a verified query, a tested agent, and an audited run. Every step you skip is where the wrong number gets in. Databricks built all eight instruments because regulated buyers demand them; an August 26 post maps these instruments onto Japan’s banking security guidelines. The instruments are there. Nobody makes you play them.
|
