AI Guides › Playbooks
The Three-Test Hiring Match Card
By Nigel Guy · 6 min read
Most people judge an "AI recruiter" by how clever the demo feels. The tool returns a tidy shortlist, the summary reads fluently, and nobody asks what the shortlist was built from. A keyword filter with a chatty front end looks identical to a system that reads a career in context, right up to the moment it rejects the person you most wanted to meet, or surfaces someone only because their CV repeats your job advert back at you.
The rule: a match is only as good as the evidence you can read behind it, so test any AI recruiter on what it can show, not on what it says.
This guide gives you a named mechanism, the Three-Test Match Card, for hirers and candidates alike. One note on scope: we have not tied this to any single vendor, because claims about specific products change quickly and we could not verify them. What follows rests on UK regulator and government guidance, which is durable.
The Match Card: three tests
A tool that claims to match on context rather than keywords should pass all three. Fail one and you are looking at keyword search with better manners.
| Test |
Question |
A pass looks like |
A fail looks like |
| 1. Evidence |
Can it show why this person matched? |
Quotes or points the specific work (a project, a result, a scope of responsibility) that supports each match reason |
A score, a percentage, or a paragraph of praise with nothing to check |
| 2. Fairness |
Has it been checked across groups, and does it avoid guessing at protected traits? |
The supplier can produce bias audits and impact assessments, and does not infer things like gender or ethnicity from a name |
"Our model is unbiased" with no documents |
| 3. Recourse |
Can a human override it and can a candidate challenge it? |
A named person reviews outputs, and applicants are told AI is used and how to ask for a person to look again |
The tool's ranking is the decision |
These map onto what UK bodies actually say. The Information Commissioner's Office audited AI recruitment providers and reported in November 2024 that some tools let recruiters filter out candidates with certain protected characteristics, some inferred characteristics such as gender and ethnicity from a candidate's name, and some collected far more personal information than necessary and kept it indefinitely. The government's "Responsible AI in Recruitment" guide (published 25 March 2024) tells buyers to ask suppliers for training-data details, performance across protected characteristics, bias audits and model documentation, and to keep effective human oversight.
If you are hiring: run the card before you buy
- Write the role as outcomes, not a keyword list. "Has reduced onboarding time for a team of ten or more" gives a context-reading tool something to match. "Python, SQL, stakeholder management" gives it nothing a keyword filter could not do.
- Run Test 1 on five people you already know. Feed in two strong past hires, two weak ones and one person with an unusual route in. Ask the tool to explain each result. If it cannot point to evidence, or it ranks the unusual route last with no reason, stop.
- Ask for the Test 2 documents in writing. Per the government guide: bias audits, performance across protected characteristics, known limitations, intended use. Also ask what personal data it keeps, for how long, and whether it infers anything about candidates. Complete a data protection impact assessment before you start.
- Settle Test 3 before launch. Name who reviews every rejection the tool influences, and how a candidate asks for a person to look again. The Data (Use and Access) Act 2025, section 80, defines a decision as solely automated where there is no meaningful human involvement, and for significant decisions requires safeguards: information about the decision, a chance to make representations, human intervention and a way to contest. Per legislation.gov.uk, that section was fully in force from 5 February 2026. A shortlist someone skims for ten seconds is not meaningful involvement. Take legal advice on where your process sits.
- Pilot, then monitor. The government guide recommends piloting with diverse users, planning reasonable adjustments for disabled applicants, and repeating bias audits at intervals.
If you are the candidate: get found on evidence
You cannot see the tool, so make your material easy to match and hard to misread.
- Lead each role with outcomes. Scope, what changed, and by how much, in your own words and only where true. Context-reading systems have more to work with than a list of tool names.
- Keep the vocabulary of the advert where it is honest. Dropping the employer's exact terms for your real experience helps any system, keyword or otherwise.
- Describe unusual routes plainly. A career break, a switch of field or part-time work needs one clear sentence of context, not silence.
- Ask. Under UK data protection law you can ask an employer how your data is used, and where a significant decision is automated the safeguards above should let you ask for human review. Ask politely and in writing.
Where both sides get it wrong
| Mistake |
Who makes it |
Fix |
| Treating a fluent explanation as evidence |
Hirers |
Check each stated reason against the CV yourself on a sample |
| Stuffing a CV with invented keywords or hidden text |
Candidates |
Never claim what you cannot defend at interview; it fails at the first human |
| Assuming "context" means unbiased |
Both |
Context can encode the same bias as keywords; demand the audit |
| Letting the tool filter on proxies for protected traits |
Hirers |
Remove such filters; ask the supplier how they are prevented |
| Keeping rejected candidates' data indefinitely |
Hirers |
Set and publish a retention period |
How do you know it's working?
Judge it against your own baseline, not the vendor's dashboard.
- Reviewers agree. Sample shortlisted and rejected candidates each month. If a human reading the evidence disagrees often, the match is not reading context.
- Funnel by group. Compare progression rates across the groups you lawfully monitor. A gap is a prompt to investigate, not proof either way.
- Later performance. After six to twelve months, check whether people the tool favoured did better than those chosen another way. Small samples mislead, so treat this as a signal.
- Complaint and override logs. If nobody ever overrides or challenges the tool, check whether the route exists, rather than assuming it is perfect.
What to skip
- Vendor "accuracy" figures with no description of how they were measured.
- Tools that will not explain a single match.
- Personality or "culture fit" inference from names, photos or video. The ICO findings above are precisely about guessing traits.
- Candidate-side "beat the algorithm" tricks. They age fast and risk your credibility.
Guardrails
- This is general information, not legal advice. Equality Act 2010 and UK GDPR duties sit with the employer; get advice for your own process.
- We could not verify any specific vendor's claims, so none are made here. Apply the card to whichever product you are shown.
- Law and guidance move. Check the ICO and GOV.UK pages below for the current position before relying on this.
- Never feed a candidate's sensitive data into a general-purpose AI tool without a lawful basis and an agreement covering that use.
- Keep a human accountable for every hiring decision.
Sources
All 751 AI guides · JulieMango plans from £17/mo