Problem breakdown · No. 036 · November 21, 2026
Late interview scorecards stall offers at small companies
At small companies, hiring often slows down after the interviews, not during them. Interviewers leave scorecards blank, the debrief runs on memory, and the candidate waits; the fix belongs mostly to a clock, with a model kept away from the verdict.
The problem
At a company with a few open roles and no recruiting operations team, interviewers are engineers and managers with full calendars. They finish an interview, go to their next meeting, and the scorecard waits until someone asks for it. Sometimes that is days later.
The costs build on each other. The debrief happens with half the feedback missing, so the loudest recollection wins. Notes written days later are vaguer and lean more on overall impressions. Meanwhile the strongest candidates are in other processes, and every idle day is a chance for a competing offer.
Applicant tracking systems already send reminder emails. Those emails land in the same inbox the interviewer is ignoring. They also do nothing about the real friction: turning scribbled notes into a structured scorecard takes effort.
The problem is not that nobody knows feedback is due. Writing it is tedious, and nothing escalates when it is late.
Constraints
- Budget: a small team will not buy another seat-priced platform. This has to sit on top of the existing ATS through its API.
- Skills: nobody will maintain it day to day. It has to run unattended and fail quietly.
- Data: interview notes are sensitive personal data about candidates. Retention has to be short, and the ATS stays the system of record.
- Regulation: tools that score or rank candidates fall under rules such as New York City's law on automated employment decision tools and the EU AI Act's high-risk category. The design should stay clearly outside evaluation.
- Trust: interviewers must recognise the scorecard as their own judgment, or they will stop using the tool.
System design
Four components, and only one of them needs a model.
- Event intake: a webhook or polling worker reads interview schedules and completions from the ATS and writes one row per interview, with interviewer, end time and status.
- Deadline engine: plain code. Each interview gets a due time. A Slack DM goes out when the interview ends, a nudge follows if feedback is still missing, and the hiring manager is told when the deadline passes. The debrief is held until every scorecard is in or explicitly waived.
- Drafting step: the interviewer replies to the DM with rough notes or a voice memo. A model maps that text onto the role's competency fields and quotes the interviewer's own words as evidence. It leaves the overall rating blank. The interviewer edits, picks the rating, and submits.
- Writeback: deterministic code posts the confirmed scorecard to the ATS, then deletes the raw notes and the draft.
The model is limited to restructuring. It never sees other interviewers' feedback and never proposes hire or no-hire. Deadlines, escalation and the debrief gate are rules, so they belong in code where they can be tested and explained.
Trade-offs
Keeping the model out of the rating gives up the one feature that looks most impressive in a demo. That is acceptable: the bottleneck was getting feedback written at all, and a suggested rating would anchor the interviewer and invite regulatory scrutiny.
Deleting raw notes means there is no audit trail outside the ATS. I prefer that to holding a second copy of candidate data that nobody governs.
Slack-only delivery leaves out teams on other chat tools. One surface done well is enough to prove the idea.
What I would build first
I would start with the deadline engine alone, connected to one ATS and Slack: the DM at interview end, a nudge, and the escalation to the hiring manager. No model yet.
Then I would measure the time from interview end to scorecard submission for two weeks against the prior month's history from the ATS. If that gap shrinks, the drafting step is next. If it does not, the problem is cultural, and no amount of drafting will fix it.
When a workflow stalls on human follow-through, put the clock and the escalation in deterministic code, and use the model only to reduce the effort of the step people are avoiding.