Screening is the part of recruiting that scales worst. You spend 90 seconds on a resume, 80% of which is deciding whether it is even in the right bucket. AI can do the 80% in bulk: but only if you tell it what "right bucket" actually means for this role this month.
The AI for HR tool and the Resume Builder (used in reverse) are built for this. You feed the real criteria once, then run batches.
The three inputs that turn AI into a useful first pass
- The must-haves that are non-negotiable this month (exact years, specific tech or domain, location rule, visa status, etc.). List them as bullet criteria.
- The red flags that make you reject in the first 15 seconds (career gaps you actually care about, title inflation patterns, job hop that is too fast for this level, etc.).
- The positive signals you actually reward when you scan quickly (recent scope increase, quantifiable impact in the same domain, promotion velocity, specific tool you use internally).
Everything else ("strong communication," "team player") is noise that makes the model say yes to everyone.
Table: criteria that work vs. criteria that produce false negatives
| Criteria type | Example | What the model does |
|---|---|---|
| Vague | "Strong background in marketing" | Accepts almost everyone with "marketing" anywhere |
| Specific must-have | "Ran paid acquisition for a B2B SaaS company, $50k+ monthly spend, last 18 months" | Filters correctly on the JD language + numbers |
| Red flag you actually use | "Three or more roles in 24 months at the senior level" | Rejects the hoppers you would have rejected |
| Red flag you don't actually use | "Any gap longer than 6 months" | Eliminates parents who took time off and strong candidates who were laid off in 2023 |
| Positive signal | "Increased pipeline by 3x in a measured channel" | Surfaces the people who already speak your language |
A real batch workflow that keeps you in the loop
- Export the new resumes (or LinkedIn exports) as text or PDF text.
- Paste the three inputs above + the batch into the AI for HR tool. Ask for: "Return only a table: Name | Match score 0-100 | One-line reason | Link or filename"
- Sort by score. Open only the top 15-20% yourself.
- For the ones you open, spend your human 90 seconds on the actual signal: does the story in the bullets match the claim in the summary?
You still read the promising ones. You just stopped reading the 140 that had no chance.
Worked numbers from teams that tried this
Teams that gave the tool the real must-haves and red flags report:
- 70-85% reduction in resumes they personally opened
- No increase in "we should have interviewed that person" misses when they spot-check the rejected pile weekly for the first month
- The false negatives that do happen almost always trace back to a must-have they wrote that was actually flexible ("must have 5 years" when 4 + right domain was fine)
The teams that gave vague criteria saw the model surface everyone and the time saving disappeared.
Guardrails you should add on day one
- "If a candidate is missing one must-have but has an unusually strong adjacent signal (exact domain + 2x the impact), flag them with 'borderline: strong story' instead of rejecting."
- "Never reject solely on employment gap length. Only reject on gap if the gap is combined with title regression or no explanation in the resume."
- "When in doubt, include. Your job is to reduce the pile, not to make the final yes/no decision."
These three lines prevent the over-filtering that makes recruiters stop trusting the tool.
How the Resume Builder helps on the candidate side
When candidates use the Resume Builder with the actual JD language, their bullets already contain the numbers and the exact terms you are screening for. That makes the AI screen more accurate on both sides.
You are not "fighting AI resumes." You are using the same structured input on your side that good candidates are already using on theirs.
The one metric that tells you the filter is working
After two weeks, calculate: of the resumes the model scored >80 and you opened, what % turned into a human screen that you would have done anyway? If it is above 60-70%, your criteria are good. If it is below 40%, you gave the model criteria that are too loose or too strict: go back and tighten the must-haves and red flags with the actual rejections you made by hand.
Use the AI for HR tool with your real three inputs for every batch. You will still do the final human read on the shortlist. You will just stop wasting the first pass on the long list.