Holding a Release: A Q&A on the Hiring Model We Chose Not to Ship

Archi3 min read

aipolicyhiringsoftware

Contents
  1. A release held for two quarters
  2. Q: What did the evaluation measure?
  3. Q: Why not ship with a disclaimer to recruiters?
  4. Q: Who actually made the call?
  5. Q: What did the two quarters buy?
  6. Q: How did customers take it?
  7. Q: What changed in how you decide?
  8. Q: What would you tell a team facing this call next month?
  9. Q: What is still unresolved?

A release held for two quarters#

In March, a 240-person HR software company had a hiring assistant ready for general availability. The tool drafted interview feedback and rejection notes for recruiters across roughly 1,900 customer accounts. Two weeks before launch, an internal evaluation found the drafts scored worse for candidates who wrote in short fragments or mixed languages.

The company held the release. I spoke with its director of AI governance, who asked that her name and the company's name be withheld because the evaluation is still under review. The transcript below is condensed.

Q: What did the evaluation measure?#

A: About 14,000 anonymized recruiter notes, scored on a 1 to 5 rubric for factual accuracy, tone, and whether the draft kept the candidate's own wording. Overall, the assistant averaged 4.2. Then we split the results by how the applicant had written. Standard business English came in at 4.4. Applicants who wrote in fragments, or mixed two languages in one paragraph, came in at 3.1.

That 1.3 point gap was the entire problem. The aggregate number looked fine. The segment table did not.

Q: Why not ship with a disclaimer to recruiters?#

A: Because the disclaimer never reaches the person who gets hurt. A recruiter opens the draft, reads three lines, edits a word, hits send. Median time from open to send in our pilot was 11 seconds. No notice survives that. The text a candidate receives is the product.

Q: Who actually made the call?#

A: We had written a rule in January: any feature that generates text a candidate will read must pass a segment threshold, not an aggregate one. The threshold was no segment more than 0.5 points below the overall score. We failed it by 0.8.

So the decision was already made by the policy, which is the point of having one. The argument was about whether to waive it. Engineering wanted to ship and patch. Legal was neutral. I recommended holding. The CEO took about 40 minutes and agreed. We pushed general availability by two quarters and booked the delay against projected first-quarter revenue of roughly $700,000 that moved out of the period.

Q: What did the two quarters buy?#

A: Three things. We rebuilt the eval set with 3,200 additional examples from applicants who write in fragments, and hired two contract reviewers who read those languages natively. We changed the training mix. By August, the gap was 0.3 points, inside our threshold.

The third change was a refusal path. When the model's confidence is low, or the applicant wrote in a language the model handles poorly, the tool produces a structured template with no generated prose. A recruiter fills in the blanks. It is less impressive in a demo and it is the part I would keep if I could keep only one.

Q: How did customers take it?#

A: Sales had already given launch dates to 40 accounts. We told them the reason. Six asked to take it anyway with an opt-out for candidates, which we declined. We lost two accounts over it, worth about $95,000 in annual contract value combined. That is less than the cost of one bad rejection letter quoted in a trade publication, and everyone in this building knew it.

Q: What changed in how you decide?#

A: Three habits. First, the eval budget is now a line item at about 6 percent of a feature's engineering hours, approved before the build starts. Second, segment thresholds get written down before anyone sees results, so the number is not negotiated against a launch date. Third, we set kill criteria: if a fix needs more than two quarters, we cancel the feature rather than extend it. Nobody has hit that yet.

Q: What would you tell a team facing this call next month?#

A: Write the threshold down before the eval returns. Know who signs the waiver, because a decision with no named owner becomes a decision by meeting drift. And expect the aggregate score to look good. It almost always does.

Q: What is still unresolved?#

A: Procurement. Buyers send questionnaires that ask for one accuracy number, and we now send a nine-page segment table instead. Two prospects said that was more than their review process could absorb, so we are packaging a two-page summary with the full table as an appendix.

We also have not mapped our documentation to the phased obligations for hiring tools under the EU AI Act. That work starts next quarter, and I expect it to change our templates more than our model.

More articles