Why every AI agent decision needs a receipt

Picture an AI agent watching a checkout system. A conversion alert fires half an hour after a pricing engine pushes new discount rules. The agent digs through millions of session logs, pulls 50 that show cart abandonment, and confidently declares the pricing change broke things. Sounds useful. Except it skipped every question that actually matters: did conversion drop across the whole customer base or just one region? Did page load times spike at the same moment, pointing to a performance issue instead? The agent found matching records. It never measured how the underlying numbers actually moved.

That gap between retrieval and measurement is the real problem, according to a piece from The New Stack. Retrieval finds candidate evidence — logs, sessions, documents that match a query. Analytics is a different job entirely: computing a rate, comparing time windows, testing whether a shift is real or just noise. An agent that reads through abandonment logs and jumps to a causal conclusion has quietly skipped the verification step, and nobody notices until the recommendation is already wrong.

The proposed fix is what's called a bounded evidence packet — a structured response that travels with the measurement instead of just handing over a number. It includes an as_of timestamp, per-source ingest watermarks showing how fresh each region's data is, and a known_gaps field that spells out incomplete coverage instead of letting the agent assume everything's there. If EU checkout events are lagging by roughly 90 seconds, that fact needs to sit inside the packet, not buried in some runbook nobody reads during an incident. The packet also references a versioned metric definition and a parameterized query template, so when an agent reports a 22% conversion drop, someone can trace exactly which calculation produced it and rerun it later.

Approximation matters too. Plenty of analytics pipelines lean on algorithms like t-digest percentiles or HyperLogLog cardinalities — legitimate tools, but not exact ones, and the packet needs a calculation field that says so plainly rather than letting an estimate masquerade as fact. Counterchecks matter as well: if conversion fell while average order value rose and load times stayed flat, that combination undercuts a performance-based explanation and points investigators toward customer segment behavior instead. Drill-down session IDs still help humans eyeball examples, but the packet has to make clear those sessions illustrate an aggregate, not prove it.

Underneath all this is a governance argument. A trustworthy system separates what the database observes from what the language model interprets. The database returns auditable, re-executable measurements; the LLM makes the call on what they mean; a policy gate or human reviewer can then rerun the query and check whether the interpretation holds up. Skip that separation, and an agent's confidence starts looking exactly like the system's certainty — which is precisely the moment things go wrong quietly, at 2 a.m., with nobody watching.