AI Recruiting Tools: What AI Can and Cannot Do at Each Hiring Stage

AI recruiting tools can save time on reading, summarising, drafting and organising, but they should not make hiring decisions. At each stage, from the job description to the offer, AI works best as a first pass that a person checks, with its reasons visible and a human owning every yes and no.

What people mean by AI recruiting tools

The phrase covers very different things. Some tools use AI to write or improve a job description. Some read CVs and summarise them or score them against a job. Some draft emails, suggest interview questions, or turn an interview transcript into notes. Others claim to rank people, predict who will succeed, or screen video answers. These are not equal in risk. Drafting an email carries little. Ranking people carries more, because mistakes shape who gets seen. Predicting success from a face or a voice is the part to avoid, because it is hard to explain and hard to defend. This guide goes stage by stage, saying what AI can reasonably help with, where it fails, and what a person must do. It is written for small and mid-size hiring teams who will use AI without a data science team behind them. We describe what Rezumz does only as far as our own documentation says, and we do not make claims about other products. Rules on AI in hiring differ by country and some places require notices or audits, so check local law. For a related checklist see /blog/ai-resume-screening-compliance-checklist-for-small-employers and the overview at /ai-resume-screening.

Stage one: writing the job and its requirements

AI is useful here as an editor. It can tidy a messy description, suggest a clearer structure, point out vague phrases and propose a first list of must-have and nice-to-have skills. What it cannot do is know your business. It will happily list ten must-haves where three are real, or copy requirements that exclude good people, such as degree demands for a role where skills matter more. It may also reproduce the biased wording that sits in the examples it learned from. The human rule is simple: you decide what the job truly needs, and you cut the list. A short list of real must-haves is what makes every later step fairer, because everything downstream is measured against it. Ask yourself for each requirement whether a person could do the job without it and learn it in a month. If yes, it is a nice-to-have. Our guide at /blog/how-to-write-must-have-criteria-an-ai-or-a-human-can-screen-against shows how to phrase them so a person or a tool can check them against a CV.

Stage two: sourcing and attracting applicants

AI can draft outreach messages, suggest search terms and help you rewrite a post for a particular audience. It cannot tell you whether someone wants to move, and it cannot fix a weak offer. Mass AI-written messages that look generic tend to be ignored, and sending them in bulk can breach the rules of the platforms you use or local privacy law. If you contact people who did not apply, you have collected their data, and in many countries you need a lawful reason and must tell them. Keep a person in charge of who gets contacted and what is said. Also be careful with tools that scrape profiles. Check the terms of the site and local law before using any. For a small team, the honest value of AI here is speed on drafts, not the discovery of hidden talent. A good job post and a simple apply link still do more than a clever tool.

Stage three: reading and screening CVs

This is where AI is most used and most risky. Done well, it reads many CVs quickly, pulls out the evidence that matches each requirement, and shows the lines it relied on so a person can check. Done badly, it reduces a person to a number with no reasons, rewards keyword stuffing, or learns from past hiring patterns that were themselves biased. It also cannot read what is not written: a career gap, an unusual route into a field, or a CV in a different format may be misread. The rules are firm. Never let a tool reject anyone on its own. Always show the requirement, the evidence and the missing items. Read every shortlisted CV and a sample of the ones below the line. Record the reason when you reject. Our comparison of methods at /blog/ats-keyword-matching-vs-evidence-based-resume-screening explains why showing evidence beats a bare score, and /resume-parser covers how CV text is read.

Stages four to six: interviews, feedback and offers

For interviews, AI can suggest questions based on the job and the CV, and it can help draft a structured scorecard. It can also turn a transcript into notes mapped to requirements, with quotes, which saves typing. It cannot judge character, honesty or fit, and tools that claim to read emotion or truthfulness from voice or video are unproven and risky. Use AI notes as a draft, let the interviewer correct them, and keep the interviewer's own rating as the record. For feedback and decisions, AI can summarise where interviewers agree or differ. The decision itself, advance, hold or reject, stays with a person who writes the reason. For offers, AI can help draft wording, but pay, terms and legal text must be set and checked by people, with approval and a record. Our guide at /blog/using-ai-to-analyse-interviews-fairly goes into transcripts, and /interview-feedback-form and /offer-letter-template give starting points.

A stage-by-stage risk table

Use this table to decide how much checking each use of AI needs. The risk levels are our judgement for planning, not a legal ruling.

The pattern is clear: the closer a step is to a decision about a person, the more you need a human and a record.

Human-review rules to adopt this week

Write these down and share them with everyone who hires. One: no candidate is rejected by a tool without a person reading the evidence. Two: any score is shown with its reasons, and if a tool cannot show reasons, do not rely on it. Three: every rejection records a reason in plain words. Four: check a sample of low-scored candidates each month to see whether good people are being missed. Five: do not feed protected or sensitive details into a tool unless you have a lawful reason, and strip names and photos where you can for first-pass reading. Six: tell candidates when AI is used, if local law or good practice calls for it. Seven: keep a record of which tool and which version was used on which job. Eight: have one named person responsible for AI use, even if that is the owner. Nine: review the rules after any complaint or surprise. These rules cost little and protect both candidates and your business. Laws on automated decisions, notices and audits differ by country, so check local law and update the rules to match.

A worked example with example numbers

Consider an example team hiring a customer support lead, with 90 applications. These numbers are examples. Reading each CV for four minutes would take six hours. A tool scores all 90 against five must-haves in minutes and shows the quoted lines. The recruiter reviews the top 20 in full, about 80 minutes, and checks a sample of ten from below the line, another 40 minutes. In the sample, two people score low only because their CVs describe the skill in different words, so the recruiter adds them to the shortlist. That catch is the point of the sampling rule. Total time is under two and a half hours instead of six, and two people were rescued from a wrong cut. Without the sample, the tool would have looked accurate while quietly missing them. The lesson: AI saved time, a person protected fairness, and the written reasons made the final shortlist easy to explain to the hiring manager. If the tool had given only a ranked list with no reasons, the same checking would have been impossible.

Common mistakes with AI in hiring

Treating a score as a verdict is the biggest. Others follow. Buying on a demo built from clean sample CVs. Assuming a vendor has tested for bias without asking to see how. Pasting full candidate records into a general chatbot, which may send personal data to a third party you have not vetted. Letting AI write rejection emails that sound cold or hold wrong details. Using the same prompt for every role, so requirements stay generic. Forgetting that AI-written CVs now exist, so polished wording proves less than it used to, which our guide at /blog/how-to-screen-ai-written-resumes covers. Failing to keep records, so that when a candidate asks why they were not shortlisted, nobody can say. And assuming that because a tool is popular it is lawful where you operate. Each of these is avoidable with the review rules above and a habit of asking where the data goes and how a decision can be explained.

Questions to put to any AI vendor

Before you trust a tool, ask six plain questions and write down the answers. What exactly does the AI read, and what does it produce? Can I see the reasons behind every score or summary? Where does the candidate text go, and which outside providers process it? Is my candidate data used to train any model? Can I switch the AI features off and still use the product? What happens when the AI is wrong, and how do I correct it? A vendor who answers these clearly is easier to trust than one who talks about accuracy percentages without saying how they were measured. Be careful with any figure you cannot check, and with words like unbiased or fully compliant. Test the tool on ten CVs you have already judged and see whether it agrees with you for reasons you accept, not only in the final ranking. Disagreements are useful: they show you what the tool misses.

Where Rezumz fits

Rezumz uses AI in limited, stated ways. Scoring a CV compares it with the job's must-have and nice-to-have skills, and that scoring does not need AI. Uploaded CVs also get an AI summary of strengths, and the job description enhancer, interview questions, transcript analysis and the hiring assistant use AI. Applications through the apply page are scored without AI. Rezumz shows the CV lines it relied on and lets you add notes against them. It never rejects a candidate for you, and AI output is decision support, not a verdict. It does not detect lies, read emotion or predict success. Pricing: a free plan with 2 free candidate analyses, $29 a month, $199 a year, or a limited $499 lifetime plan; Team is $79 a month or $699 a year for up to 10 teammates, suited to teams with 4 or more recruiters. It is not the right choice if you need a large enterprise suite with a deep HR system or payroll, or one-click posting to many job boards. See /pricing or try /demo.

When not to use AI in hiring

Do not use AI where you cannot explain the result to a candidate. Do not use it for the final decision, for judging character, or for any step where local law restricts automated decisions. Do not use it if the vendor cannot tell you where candidate data is stored and who processes it. Do not use it on very small volumes where reading ten CVs yourself is faster and fairer. Do not use it to compensate for an unclear job description. And do not use it in secret: if your candidates would be surprised to learn AI read their CV, you probably need to tell them. This is a place where caution costs little. Start with low-risk uses such as drafting and summarising, add screening support once you have review rules in place, and keep a person accountable for every outcome. Rules differ by country, so check local law before relying on any of this.