Databricks made ai_decide more accurate—but the model behind the SQL function can change again
The beta function now carries a 33% internal accuracy claim. Teams using it for routing or prioritization still need their own fixed regression set because Databricks reserves the right to replace the underlying model.
Databricks says its beta ai_decide() SQL function now answers questions 33% more accurately while keeping its previous low-latency profile. That is useful progress for teams using SQL to classify support tickets, score urgency, or route records—but it is not a stable-model guarantee.
The product contract explicitly says Databricks may change the model behind ai_decide() when another model performs better in its internal benchmarks. The function exposes a metadata.version field, but the current documented option is only version 1.0; the documentation does not promise that pinning that function version pins a specific underlying model.
What operators should treat as the contract
ai_decide() accepts text or structured data and can answer several rubric questions in one call. It supports three output forms: a probability for a true/false-style question, one selected label with probabilities, or a probability-weighted score across an ordered scale. Generated answers can vary between calls.
Databricks says input parameters are processed within its security perimeter and are not stored, although run metadata such as the Databricks Runtime version is retained. The beta is available only in supported regions, does not run on Databricks SQL Classic, and requires Runtime 15.4 LTS or later today; Databricks recommends Runtime 18.2 or later and has separately announced a future 17.3 LTS minimum for affected AI Functions.
The rollout gate
Do not promote this change solely because the vendor reports a 33% improvement. That figure comes from Databricks’ internal benchmark, not a workload-specific evaluation disclosed on the function page.
Before using the function in a consequential routing or prioritization path:
- Keep a versioned set of representative inputs and expected rubric outcomes.
- Record the full response, including probabilities, confidence,
metadata.version, runtime version and error field. - Compare both decision accuracy and threshold behavior; an unchanged top label can still hide a material probability shift.
- Put human review around low-confidence cases and any decision that changes customer access, spending or incident severity.
- Re-run the set after runtime upgrades or unexplained output drift, because the underlying model can change without a new SQL function name.
The practical reading is straightforward: Databricks has improved a useful decision primitive, but adopters own the regression boundary. A stable SQL signature is not the same thing as a stable decision model.
sources
- Databricks platform release notes — November 2026docs.databricks.com
- ai_decide functiondocs.databricks.com
comments · 0
- A new French open-data test finds architecture matters less once the local model reaches 3B parametersNov 17, 2026
- ThoughtSpot’s Data Cache is fast—but cached and live answers are not always equivalentNov 17, 2026
- Genie One scheduled tasks can now email colleagues—treat the recipient list as a data-delivery surfaceNov 17, 2026