The cheapest AI safety feature in 2026: “I don’t have current data.”
Most AI features I see at scaleups ship without a mechanism for the agent to refuse to answer. The pipeline runs. The agent responds. Whether the source data is fresh or two weeks stale, you get the same confident answer.
What I’ve watched ship this year:
- An agent answering “yes” when the source database returned null. Table hadn’t refreshed in two days. No alert.
- RAG returning Q1 2024 documents because the embedding pipeline silently truncated docs over the token limit.
- A copilot recommending action based on a metric redefined in dbt three weeks earlier. The agent’s prompt still referenced the old definition.
The fix is small and usually skipped because it feels embarrassing: build a guardrail that lets the agent return “I don’t have current data” when that’s true.
In practice:
- Every source the agent queries has a freshness SLA in its metadata.
- Before responding, the agent checks freshness on every source it touched.
- If any source breaches its SLA, the agent returns “I don’t have current data for X” with the timestamp.
An agent that admits uncertainty once is more trusted than one that’s confidently wrong twice.
Does your AI feature have a way to say “I don’t know”?