Rehearsal Space, Not Stage: Auditing What Your AI Pilot Actually Built
Photo: Taiwan Presidential Office, CC BY 4.0, via Wikimedia Commons
If a pilot cannot survive an ordinary Tuesday afternoon, it was never a pilot. It was a rehearsal.
Enterprise AI programs have grown remarkably good at rehearsals. A team assembles a working prototype against a curated corpus, stages a polished walkthrough for an executive audience, and watches the room nod. The model answers fluently, the interface is clean, and the business case writes itself. Weeks later the same system meets real traffic, real edge cases, and real users, and the distance between what was demonstrated and what was actually engineered becomes visible in a single sprint.
That distance deserves a term: demo debt. It is the sum of operational work that a demonstration never required and a production system cannot avoid. Like technical debt, it is not inherently a failure. Incurred deliberately, written down, and scheduled for repayment, it is an inexpensive way to buy information. The problem is that demo debt is almost never written down, because the people who incur it are rarely the people who inherit it.
The Recurring Artifacts
Four kinds of debt appear in nearly every pilot that reaches production and stumbles.
The first is hardcoded behavior. Prompts tuned by hand against a dozen favorite examples. Thresholds selected because they looked right on a Friday. A routing rule that sends document type A to model B because, during development, that was the only pairing that worked. None of it is exactly wrong. All of it is invisible to the next engineer and immovable without breaking the demonstration.
The second is curated or synthetic data. Pilots run on clean corpora, and clean corpora are a courtesy the business rarely extends in production. A pilot that has never encountered a scanned PDF, a duplicate record, a missing field, or a document in the wrong language has not been tested. It has been sheltered.
The third is the manual handoff — the colleague who reformats the input, spot-checks the output, or restarts the pipeline when it stalls. In a demonstration, that person is helpful. In production, that person is an unstaffed role nobody budgeted for.
The fourth is the undocumented integration. The webhook that retries forever. The service account that will outlive its creator. The connection to a legacy system that works only because someone still runs a scheduled job from a laptop.
Four Ledgers Worth Opening
An audit before scaling is less a code review than a bookkeeping exercise. Four ledgers, opened in this order, surface most of what a polished demonstration is hiding.
The data ledger. Where did the pilot's inputs originate, who approved their use, and how far do they diverge from what production will deliver? Record the sampling method, not merely the source system.
The configuration ledger. Every prompt, threshold, model version, and feature flag, each with an owner and a date. A setting that cannot be traced back to a recorded decision is a liability wearing a default value.
The human ledger. Every point where a person intervenes, how long the intervention takes, and whether the intervention was designed or improvised under deadline.
The integration ledger. Every external call, its authentication scheme, its failure mode, and whether that failure is handled or merely logged.
Warning Signs in a Polished Demonstration
- Only one person can run the demo end to end. That is not a skills gap; it is an undocumented system.
- The walkthrough begins with a reset step. Production offers no reset.
- The evaluation set was drawn from the same pool as the demonstration examples.
- Nobody can state cost per completed task, only cost per month.
- Error paths have been described but never exercised.
- The team discusses the pilot's limitations exclusively in the future tense.
Each item individually is survivable. Taken together, they suggest that the impressive artifact in the room is a stage set rather than a structure.
Running the Audit
The audit itself can be finished in two weeks for most pilots, provided the sponsoring executive agrees in advance that a negative finding counts as a useful result rather than a political embarrassment.
Select the three workflows the pilot is meant to perform and replay each with production-shaped data, including the ugly records. Remove the person who normally performs the manual handoff and observe what breaks. Run a hostile-user session with someone who has no incentive to be gentle. Measure latency and cost per completed task rather than per request. Then ask the team what fraction of the current implementation they would retain if the demonstration requirement vanished tomorrow. The answer tends to be clarifying.
Discharge, Extend, or Write Off
The audit should produce a decision, not a scorecard. Some pilots will graduate with a defined repayment plan: replace hand-tuned prompts with versioned configuration, formalize the manual step as a staffed function or automate it, harden the integrations and assign named owners. Others should be extended as pilots, explicitly, with a new date and a narrower question. A few should be written off, with findings preserved so the next team does not rebuild the same rehearsal.
None of this argues against ambitious pilots. It argues against confusing a demonstration with a system. The most impressive pilot in the portfolio is frequently the one least prepared for production — not because the team was careless, but because persuading an audience and surviving a Tuesday afternoon are different engineering problems, and only one of them photographs well.