Awesome Hires ← Back to home

AI in hiring in 2026: what it actually does, and what it doesn't

May 16, 2026 · by Vinay Devaraja · 8 min read

We have spent the last year talking to recruiters about AI. The conversations have shifted.

Two years ago the question was "do we need this?". Last year it was "which vendor?". This year it is "we bought one of these, the panel still hates it, why?". That is the question worth writing about.

So this post is not another listicle of AI hiring tools. It is what we would tell a head of talent over coffee if they asked us, honestly, what works and what does not.

Where AI actually helps

Four places. That is the whole list.

Sourcing. Pulling a longlist from LinkedIn, GitHub, an internal CRM, and the careers page, and ranking them against a JD. The work is genuinely tedious, the output is genuinely useful, and if the model gets a candidate wrong, the cost is one recruiter glance. Low stakes, high volume, perfect fit.

Screening. A 30 minute conversational screen with the candidate, transcribed, summarised, scored. Done well, this gives every candidate the same fair shot, every recruiter the same five sentences to read, and every hiring manager the same evidence to argue from. Done badly, it is a chatbot that asks "tell me about a time" and a model that hallucinates scores. The difference is whether the score cites the transcript. If it does not, do not buy it.

Scheduling. Five panellists, three time zones, a candidate, a calendar. This is solved tooling now. Any vendor worth their fee can stand up a panel inside a day. If yours cannot, that is the easiest switch you will ever make.

Signals. The part our team has spent the most time on. The first conversation a candidate has with you contains, in their own words, almost everything you need to predict whether they will accept the offer. Notice period. Competing processes. Salary. How they talk about your company in particular. Structuring that into a defensible score is the highest leverage AI work in the funnel. We call ours join likelihood and it is the thing that changed how our customers run pipelines.

Where AI quietly fails

Three places.

Borderline candidates. Models are great at the obvious yes and the obvious no. The 30 percent in the middle, the ones whose CV is light but whose conversation was excellent, or the senior with the perfect CV who phoned it in, that is where a real recruiter earns their salary. If your AI tool is making the borderline calls for you, you are quietly leaving good hires on the table.

Culture and team fit. This one is contentious. We have seen vendors claim culture-fit scores. We have not seen one we would let near our customers' hires. Culture is what your team does on a bad Tuesday. A model cannot see that, and it cannot ask the question the way a human can.

Edge cases the model has not seen. A returning parent after eight years out. A career switcher whose first role looks junior on paper. A candidate whose strongest signal is the question they asked at the end of the call. Models are average machines. They optimise for the median candidate. The brilliant non-median candidate is the one a good recruiter notices and a model rejects.

The split most tools miss: qualification versus intent

The deepest mistake in the category right now is conflating two different questions into one score.

The first question is, can this candidate do the job. Call it qualification. CV match, technical signal, behavioural evidence. Every AI hiring tool on the market scores this.

The second question is, will this candidate take the offer if we make one. Call it intent. Notice period, competing offers, salary alignment, how they talk about your company specifically. Almost no tool scores this.

The cost of conflating them is the candidate who aces every round and accepts the other offer on Friday. We wrote a whole post on this because it is the single biggest hiring waste we see, and it is structurally invisible to the tools most teams already pay for.

If you remember one thing from this post, remember this. Qualification and intent are different scores. Ask your vendor which one they predict. If they say "both", ask to see the citations.

What to ask a vendor before you sign

Five questions. Print them out. Take them to your next demo.

  1. Show me the citation. If your tool scores a candidate, can you click into the exact transcript line that drove the score. If not, your hiring managers will not trust the tool, and they should not.
  2. What is the false reject rate. Every vendor quotes precision. Recall is the number that costs you hires. Ask for both.
  3. How do you handle bias audits. Specifically, who runs them, how often, and can we see the last report. "We have a policy" is not an answer.
  4. What is the integration with my ATS. Real bidirectional sync, or a one-way push. Most "integrations" are the latter.
  5. What happens to my data when I leave. Is the model trained on my pipeline, and if I churn, what does the vendor retain. Read the DPA before the marketing deck.

If the vendor stumbles on more than one of these, walk.

What changes for the recruiter

We want to close on this because it is the part that gets lost in the noise.

The recruiters we see thriving in 2026 are the ones who treat AI as a research assistant. The model does the inbox, the scheduling, the first-pass summary, the candidate brief for the panel. The recruiter does the judgement, the relationship, the negotiation, the close.

The recruiters we see struggling are the ones who either ignore AI entirely, and watch their colleagues hire twice as fast, or who hand the judgement work over to the model and stop reading the transcripts. Both extremes lose. The middle is the only place that works.

AI in hiring is not a product category any more. It is a working assumption. The question is which parts of your funnel you let it touch, and which parts you keep for yourself. Get that split right and the rest is execution.

Sources

Frequently asked

  • What is AI in hiring?

    AI in hiring is the use of machine learning and large language models across the recruitment funnel. The common applications are sourcing (finding candidates), screening (filtering and ranking), scheduling (booking panel interviews), and signals (predicting outcomes like offer acceptance from conversation transcripts). In 2026 it sits alongside human recruiters rather than replacing them.

  • Is AI in hiring biased or illegal?

    It can be both, depending on what the system was trained on and how it is used. The EU AI Act classifies hiring AI as high risk, and several US states require bias audits and candidate disclosure. The safest setups keep a human in the loop for any rejection, audit model decisions against demographic data, and document why a candidate was advanced or declined.

  • Does AI replace recruiters?

    No. AI removes the rote work, sorting inboxes, drafting first-pass replies, scheduling interviews, summarising calls, and gives recruiters back the judgement work. The recruiters we work with hire more in less time, not the other way round.

  • What is a realistic ROI from AI in hiring?

    The honest answer is that it depends on funnel volume. Teams hiring 50 plus roles a year typically reclaim 30 to 50 percent of recruiter time on screening and scheduling, and 10 to 20 points of offer acceptance rate once intent signals are used. Below 50 roles a year the savings are real but smaller, and the case is more about consistency than headcount.