Most AI hiring projects fail for the same reason: everything gets switched on at once, on top of data nobody checked. Here is a three-month plan that avoids that.
Days 1 to 30 — measure, switch nothing on
This month feels like doing nothing. It is the month that decides whether the rest works.
- Clean the stage data. Every candidate in the right stage, every rejection with a real reason. Everything later depends on this.
- Run your funnel by hand once. You need a picture of before, and you cannot go back and create one later.
- List your tools. Everything that touches candidate data. Most teams find something nobody expected to be on the list.
- Check how well CVs are read. The 25-CV check from Module 4, on whichever tool you plan to use.
- Pick one type of job to start with. Something you hire repeatedly, with enough volume to learn from, where a mistake is recoverable.
The temptation is to skip to switching things on, because that is the part that feels like progress. Teams that skip this month spend months three and four working out why their numbers are wrong.
Days 31 to 60 — one type of job, done properly
- Run the intake meeting and build the scorecard for that type of job.
- Screen with written rules, and do the disagreement check every single run.
- Run a structured interview process with written descriptions of what each score means.
- Publish your candidate notice before the first automated step runs, not after.
- Do a first bias check. Even a rough one gives you the baseline you will compare against.
One type of job, completely, beats every job partially. You get a real answer about whether this works for you, and a small enough scope that you can undo it.
Days 61 to 90 — connect and extend
- Connect screening to interview booking, with one human check.
- Get the weekly dashboard running.
- Add a second type of job.
- Review: what changed in the funnel, what broke, and what the disagreement checks caught.
What not to touch in month one
| Leave alone | Why |
|---|---|
| Automatic rejection | Cannot be undone, candidates see it, and you have no calibration yet |
| Automated messages to candidates | Mistakes are public and cannot be recalled |
| Tools that learn from your past hires | You have not checked what they learned |
| Recorded video assessment | The highest legal exposure in the whole toolkit |
| All your jobs at once | Nothing to compare against, and nowhere to retreat to |
What to measure as you go
Track these from before you start, not from when you deploy:
- What share of applications actually get reviewed. This is where the honest gain shows up.
- Days between screening and first interview.
- How often the disagreement check finds something.
- Share of candidates getting through screening.
- The bias check against your baseline.
Watch the fourth one carefully. A sharp drop in how many people get through usually means over-filtering, not better filtering — and it is easiest to catch in the first month, while you still remember what you changed.
Decide when you would stop
Write it down now, before anyone is invested. What would make you turn this off or roll it back? A bias check that moves in the wrong direction? A hiring manager disagreeing with the shortlist three times running? Candidate complaints?
A stopping rule written in advance is easy. The same decision made under pressure, with a project everyone has defended in meetings, is very hard.
Download the rollout plan · all templates
The order that matters
Fix the data, then automate. Doing it the other way round automates your existing mess and makes it faster, more consistent, and much harder to see.