Twelve Things AI Still Cannot Do in Recruitment

Almost every claim made about AI in recruitment is about what it can do. The more useful list is the other one, because it tells you where your recruiters still have to be good and where buying software will not help. Some of these are temporary limitations and some are structural, and the difference matters when you are planning headcount.

Last reviewed July 2026. Limitations drawn from Sortinghat delivery experience.

What it cannot do about the brief

1. Extract a requirement the client has not articulated. The information that decides most rejections was never written down. Getting it requires asking a hiring manager an uncomfortable question, and nothing automates that conversation.

2. Know that a requisition is not real. Budget unapproved, role internally contested, or a manager gathering market information. A system will search diligently against a role that was never going to be filled.

3. Tell you the client is wrong. An unrealistic combination of skill, salary and location is a conversation, not an output. The system will simply return nothing and let you conclude the market is thin.

What it cannot do about candidates

4. Persuade a reluctant senior candidate. Someone with three options is deciding on trust and judgement. This never got automated and it is the highest-value thing a recruiter now does.

5. Read a hesitation. A candidate who says yes without meaning it sounds like a candidate who means it, on paper and often on a call. Experienced recruiters catch this and cannot fully explain how.

6. Manage the offer-to-joining gap. Counter-offers, family opinions and cold feet. This is where placements die, and it is entirely relationship work.

7. Repair a bad candidate experience. Once somebody feels mishandled, an automated apology makes it worse.

What it cannot do about your firm

8. Make your team use it. Industry data consistently shows a wide gap between firms claiming AI usage and firms with it genuinely embedded. Adoption is a management problem and software does not have the conversation for you.

9. Fix a database nobody maintained. It can find what is there. An audit of one firm's 3 million records found 10 percent duplicates and 10 percent unusable, and no search recovers information that was never captured.

10. Compensate for a weak fee structure. If the margin was wrong at contract signature, efficiency delivers the loss on time rather than preventing it.

What it cannot do about accountability

11. Take responsibility for a decision. Under the EU AI Act's high-risk classification for recruitment, human oversight of automated decisions is an obligation. Beyond compliance, a client rejecting a shortlist wants a person to explain it.

12. Guarantee it is not wrong. A score is a judgement expressed as a number. If a system cannot show the reasoning underneath, recruiters ignore it, and they are right to. This is the argument for evidence being visible on every score rather than a confidence figure standing alone.

Which of these will change

Being honest about direction matters as much as being honest about the present.

Likely to improve: reading hesitation from conversation patterns, and identifying unrealistic requirements by comparing a brief to market data. Both are pattern problems with enough signal to learn from.

Unlikely to change: anything requiring accountability, anything where the value is that a person chose to spend time on you, and the offer-to-joining gap. These are not technical limits. They are what the relationship is for.

What this changes on a staffing desk

Market analysis is only useful if it changes something on Monday. Four practical consequences that follow from everything above.

Your prospect list is probably built from history

Most staffing firms sell to the accounts they already know, which means the client list reflects where demand was five years ago rather than where it is now. Rebuilding a prospect list against current hiring activity rather than past relationships is unglamorous and it is usually the highest-return week a founder can spend.

Speed matters more than it used to

In contingency and RPO, the firm that submits first usually gets paid. A recruiter working manually manages around 40 calls a day and finds three or four people worth submitting. Filtering before dialling changes that arithmetic: a pool search on a volume role surfaces around 1,200 relevant profiles, narrows to a ranked 300 contacted across call, WhatsApp and email in parallel, and produces roughly 10 submission-ready profiles in hours.

Your own database is the cheapest source you have

An audit of one firm's 3 million records found roughly 10 percent duplicates, 10 percent unusable and 80 percent real people whose records were simply out of date. Making that 80 percent findable took job portal spend from about $100 a role to $60. Most firms are paying to source strangers while sitting on people who already know them.

Definition beats effort

Most wasted submissions are a briefing failure that surfaced three weeks later, not a sourcing failure. Generating weighted evaluation criteria from the job description, and agreeing them with the client before the search starts, removes more waste than any increase in activity.

Frequently asked questions

The useful way to read this list

Go through the twelve and mark which ones your firm is currently failing at. Most firms find the failures cluster in the first three, which are all about the brief.

That is worth knowing before buying anything, because a firm whose problem is briefing will not be helped by better sourcing, and will conclude the technology does not work when the actual problem was upstream of it.