Skip to content
Sign in
Artwork for DEV
DEV · August 26 · 7 min

The Eval Gap: How to Know if Your Internal AI Tool Actually Works

Shipping an internal AI tool is the easy part. Knowing whether it's still working six months later — that's where most teams go silent. This episode of Development tackles the "eval gap": the absence of any repeatable system to detect when an AI tool's output quality has silently drifted, and what it actually takes to close it. The episode walks through a concrete, step-by-step framework for building an evaluation practice around custom internal tools — from structured data extraction to open-ended response drafting. Key topics covered include: Why silent degradation happens — model provider updates, shifting inputs, and accidental prompt edits can all erode output quality without triggering any error or alert. Building a golden set — how to curate 20–30 representative historical examples with verified correct outputs, and why this investment pays off more than any other part of the process. Defining a rubric — the difference between binary scoring for structured tasks and dimension-based scoring for open-ended outputs, and how to make either one fast enough to actually run. Using a model as a critic — why prompting an AI to answer specific yes-or-no questions about another AI's output is a reliable evaluation method, and what makes it work. Setting thresholds with consequences — how to turn evaluation scores into operational decisions, so results on a reporting dashboard drive real action rather than sitting unread. The three failure modes — stale golden sets, evaluations too slow to run consistently, and waiting until something breaks before building the practice at all. The episode makes a strong case that evaluation isn't a one-time audit or a technical luxury — it's the operational layer that separates a tool that holds up over time from one that quietly becomes a liability. The earlier it's built into the workflow, the more useful it becomes as part of a broader business operating system. If you enjoyed this one, the episode How to Run a Debrief So It Actually Feeds Your Next Bid applies a similar operational lens to a different part of the business cycle. VB RFP

0:00-7:37
Play in

transcript

No transcript — this publisher did not publish one.

show notes

Shipping an internal AI tool is the easy part. Knowing whether it's still working six months later — that's where most teams go silent. This episode of Development tackles the "eval gap": the absence of any repeatable system to detect when an AI tool's output quality has silently drifted, and what it actually takes to close it.

The episode walks through a concrete, step-by-step framework for building an evaluation practice around custom internal tools — from structured data extraction to open-ended response drafting. Key topics covered include:

  • Why silent degradation happens — model provider updates, shifting inputs, and accidental prompt edits can all erode output quality without triggering any error or alert.
  • Building a golden set — how to curate 20–30 representative historical examples with verified correct outputs, and why this investment pays off more than any other part of the process.
  • Defining a rubric — the difference between binary scoring for structured tasks and dimension-based scoring for open-ended outputs, and how to make either one fast enough to actually run.
  • Using a model as a critic — why prompting an AI to answer specific yes-or-no questions about another AI's output is a reliable evaluation method, and what makes it work.
  • Setting thresholds with consequences — how to turn evaluation scores into operational decisions, so results on a reporting dashboard drive real action rather than sitting unread.
  • The three failure modes — stale golden sets, evaluations too slow to run consistently, and waiting until something breaks before building the practice at all.

The episode makes a strong case that evaluation isn't a one-time audit or a technical luxury — it's the operational layer that separates a tool that holds up over time from one that quietly becomes a liability. The earlier it's built into the workflow, the more useful it becomes as part of a broader business operating system. If you enjoyed this one, the episode How to Run a Debrief So It Actually Feeds Your Next Bid applies a similar operational lens to a different part of the business cycle.

VB

RFP

links6

more episodes

All episodes