# Why AI in production needs observability from day one

_A model that acts on behalf of your business will drift, and you will not notice from intuition. A senior PM's rule for autonomous agents: intuition and a few test cases are not enough once the system is deciding for real people._

AI features drift, and you cannot catch drift with intuition. An anonymized story from a senior PM building autonomous agents on why evaluations and observability belong on day one, plus how Density Labs bakes them into the AI Readiness Assessment.

# Why AI in production needs observability from day one

Here is the uncomfortable thing about a deployed AI feature: it can get worse without anyone touching it. The model provider updates something, the data shifts, a prompt that worked last month stops working, and the output quietly degrades. If you are relying on intuition to notice, you will notice through a customer complaint.

That is why observability is not a phase-two nicety. It is a day-one requirement, and the teams building serious AI features treat it that way.

## The PM who stopped trusting intuition

A senior product manager building autonomous agents at a large software company told me why her north star shifted to evaluations. Once a model is acting on behalf of an enterprise and its customers, she said, intuition and basic testing are not enough. You cannot eyeball whether an agent that handles a case from creation to closure is still behaving the way it did in the demo.

Her team's answer was strong evaluations and golden datasets: a curated set of examples with known-good behavior that you run the system against continuously. She compared it to behavior-driven development, but for AI behavior instead of code. The point is not to test once and ship. It is to keep confirming, over time, that the system stays within its guardrails and behaves as intended even as everything underneath it moves.

She also drew a distinction that matters for observability design. Human-in-the-loop is a blocking dependency, a person who has to approve before the system proceeds. Human supervision is oversight while the system makes autonomous progress. You cannot supervise what you cannot see, which means the monitoring has to be built in before you loosen the human's grip, not after.

The lesson underneath all of it: a demo shows you one moment. Production is a thousand moments you are not watching. Observability is how you watch them anyway. And the cheapest time to build it is before the feature has any users at all, because retrofitting instrumentation onto a live system means you are already blind for the window it takes to add.

## What to actually watch

Observability for an AI feature is not a single dashboard, it is a set of signals you decide on before launch. The ones worth instrumenting from day one:

- **Latency.** Slow is a failure mode. Users abandon before they complain.
- **Token cost.** The bill is a live metric, not a monthly surprise. Watch it per feature.
- **Hallucination rate.** How often the output is confidently wrong, measured against your evaluation set, not vibes.
- **Output drift.** Whether behavior today still matches behavior at launch. This is the one intuition never catches.
- **User satisfaction.** The signal that ties all the rest to whether the feature is actually earning its place.

Latency, token cost, hallucination rate, output drift, and user satisfaction are the five you want wired up before the first real user arrives, not the week after.

## How we approach it at Density Labs

When we scope an AI feature in an AI Readiness Assessment ($2,500), evaluation and observability are part of the design, not a follow-on project. We help teams define what good behavior looks like concretely enough to build a golden dataset, decide which of the five signals matter most for this feature, and figure out where human supervision replaces human-in-the-loop safely.

The teams that skip this ship blind, and blind is fine right up until the day the model shifts under them. Then it is a customer discovering the regression before they do.

You do not get to observe an AI feature retroactively. Either you built the instruments before launch, or you are guessing. Build the instruments.