ICO AI Recruitment Report: A Checklist for UK Recruitment Agencies

The ICO expects an agency using AI for sourcing, screening or scoring to assess privacy risks before deployment, identify a lawful basis, explain the use to candidates, minimise and protect their data, test for unfair outcomes, and keep meaningful human oversight. Obtain written supplier assurances and record how each safeguard works in your actual recruitment workflow.

The ICO's November 2024 report came from consensual audits of tool developers and providers, run between August 2023 and May 2024, yet the six procurement questions the ICO published alongside it are directly useful to agencies. A policy saying “a consultant checks the result” is weak evidence if the consultant never sees rejected candidates. Treat each item below as a control with an owner, an artefact and a test.

Scope the exact candidate decision first

List every AI-supported step separately: finding profiles, parsing CVs, ranking matches, recommending interview invitations and drafting candidate messages. A tool can perform several of these under one product name. For each step, record the input, output, who uses the output, and whether anyone is removed from consideration. This is the first agency checklist item because a broad description such as “AI helps with hiring” hides the point where a recommendation becomes an exclusion.

For an illustrative eight-consultant agency, consider a role that receives 120 applications. Its applicant-tracking system parses all 120 CVs, assigns scores and presents the top 25. A consultant opens only those 25 profiles. The agency may call this human review, but the other 95 candidates have effectively been excluded by the ranking. The agency should either review a defensible sample and the decision rules for the lower-ranked group, or change the workflow so a person can assess borderline and rejected records before the final shortlist.

Write the processing map in plain language: “Applicant uploads CV; parser extracts work history and skills; matching tool compares those fields with the role brief; consultant reviews recommendations; consultant records shortlist reason.” Include data obtained from professional profiles or candidate databases, because a candidate may never have submitted the information for this particular role. If several clients have different instructions, map their uses separately. A single generic control may conceal different retention and disclosure arrangements.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Complete a DPIA at procurement, then keep it current

A data protection impact assessment, or DPIA, is a structured record of likely effects on people and the measures that reduce risk. The ICO advises carrying it out before using a recruitment AI tool, ideally during procurement, and revisiting it as the processing changes. Put the actual product, dataset, users and decisions into it. A supplier's generic DPIA can inform yours; it cannot describe your agency's own role settings, shortlist practice or candidate communications.

• Record the intended benefit and why the agency needs candidate data for that specific step.
• List the input fields, including any inferred or sensitive attributes, and state which fields the tool should never use.
• Describe plausible harm: missing a qualified applicant, exposing a confidential CV, misidentifying a career break or creating a biased shortlist.
• State the safeguard, its owner and a repeatable test. “Consultant review” needs a review queue, sampling rule and overturn record.
• Give the decision to proceed, change or stop, with a date for review after a material product or workflow change.

In the 120-application illustration, the agency might record an initial pilot with 20 previously assessed, appropriately handled CVs. Two consultants compare the AI rankings with the recorded role criteria and note disagreements. Those figures are an illustrative test plan, not an ICO threshold. If the tool repeatedly scores a required certification below a vague “culture fit” signal, amend the configuration or stop the pilot. An assessment signed once and filed away will miss a later model update, a new client sector or a wider candidate dataset.

Ask a privacy specialist to review the DPIA when the agency cannot explain a lawful basis, the tool infers sensitive attributes, or a score determines who is considered. Keep the final decision with the agency rather than accepting a vendor's sales assurance. The broader question of evaluating AI features in recruitment software belongs in a separate product assessment; this checklist is about the evidence you need for accountable use.

Record the lawful basis and the parties' instructions

For each use of personal information, note the agency's lawful basis before processing begins. Do not treat a candidate's submission of a CV as a blanket permission for every subsequent AI use. If the agency relies on legitimate interests, the ICO says to complete a legitimate interests assessment. Sensitive special category information requires an additional condition. The choice depends on the specific activity and relationship, so obtain qualified advice rather than copying another agency's wording.

Check who determines the purpose and means of processing. An agency may decide how it screens its applicants while a vendor processes data on written instructions. A client may set its own decision criteria. Document those roles and instructions in the contract. Ask whether the supplier can use uploaded CVs to train a general model, build a shared candidate database, or pass records to another provider. If the answer is vague, pause the upload. The ICO's 2024 outcomes report explicitly warns against providers retaining or repurposing candidate information beyond recruiter instructions when acting as processors.

Make a procurement evidence request that can be answered with documents: “Show the data-flow diagram, processing terms, sub-processor list, deletion process, incident contact, model-change notice and your test evidence for fairness and accuracy.” Mark each response as supplied, incomplete or absent. A polished security page alone does not demonstrate that the supplier follows your specific instructions. The ATS add-on versus system-switch comparison helps when you are deciding which integration exposes fewer candidate records.

One concrete failure to watch for: a trial account imports the entire database of 8,000 candidates so the supplier can show match rates for one vacancy. That is disproportionate to the initial test. Use a small, appropriately selected sample with a defined deletion date. Ask the supplier to confirm deletion in writing after the test, and check that exported results and support tickets have not kept the same CVs elsewhere.

Give candidates a usable explanation

Transparency means a candidate can understand how their information is used without reading a model architecture document. The ICO asks recruiters to explain why the tool is used, the logic involved in predictions or outputs that may affect people, and how to challenge automated decisions. Update the candidate notice and the actual point of collection. A notice buried in a footer cannot do all the work when the agency begins scoring applicants for the first time.

A short illustrative notice could say: “We use a screening tool to extract skills and experience from your CV and suggest matches to role criteria. A consultant checks the suggestions before deciding who to contact. You can ask us to correct information or review a decision by replying to your consultant.” This wording is only suitable if it accurately describes the live process. If only the top-ranked applicants are seen by a person, the second sentence is misleading and must be changed alongside the workflow.

Make the notice specific about what enters the tool and who receives the output. If it uses profile data gathered from elsewhere, explain that route. Give a retention period or the rule used to set it. State whether the agency shares scores with a client or only sends an independently reviewed shortlist. Test the wording with a colleague who has not seen the system: can they tell what information is used, how it affects them and whom to contact? If not, rewrite it.

Another illustrative case is a candidate with two versions of a CV. The agency imports an older version from its database while the candidate submits a newer one for the role. A score based on the older CV is inaccurate even if the model works as designed. The candidate needs a visible way to correct the source record; the agency needs a way to re-run or disregard the score and record the correction.

Make human review visible in the record

Decide exactly which decisions a person makes and when. A meaningful review allows the consultant to see relevant source information, disagree with the output and change the result. Merely clicking “approve” on a recommendation is weak. The ICO's March 2026 report, Recruitment rewired, says that where an employer relies on meaningful human involvement, it must be applied consistently to all candidates within a hiring stage, and that candidates must be told how to challenge a decision and ask for a human review. Check current ICO guidance and seek legal advice for the precise rules applicable to your decision process.

Use an audit field for the review: role criterion checked, AI suggestion, consultant decision, reason for any change and date. Do not force consultants to write essays for every routine match; a short structured reason is enough to show they considered the right evidence. For the 120 applicants, the agency could review all 25 suggested names and a selected set of lower-ranked candidates, with special attention to scores close to the shortlist boundary. Define that sample rule before seeing the results, then track whether the review finds missed qualified people.

A realistic borderline example: the role needs a current licence and two years of client-facing experience. The tool gives a high score to a CV with the right job title but an expired licence; it gives a low score to a candidate whose licence is current but recorded in an attachment the parser missed. A consultant can identify both only if the source documents and extraction fields are visible. Fix the parser or intake field, not merely those two records.

If the agency sends a rejection automatically on the basis of a score, document that as a different process. It may require stronger safeguards than using a score as a lead for human assessment. Do not assume a nominal human somewhere later in the hiring process makes an earlier automatic rejection harmless. For operational review points beyond this compliance checklist, see recruitment AI agents and their review points.

Test for unfair screening and wrong data

The ICO's audits found examples of tools allowing filters based on protected characteristics and inferring characteristics from names. Ask the supplier what fields it extracts or infers, which filters are available to an agency user and what evidence supports its bias claims. Disable inappropriate filters and test the actual configuration your consultants will use. A generic benchmark from the vendor may not represent your roles, CV formats or candidates.

For each pilot role, compare what the AI extracted with the CV itself for a sample that includes high, middle and low scores. Record false positives and false negatives, then inspect whether they cluster around a CV style, career path or other characteristic. The aim is to find a correctable failure, not to manufacture a reassuring percentage. If you have a lawful reason to use sensitive information for fairness testing, ask a specialist how to handle it separately and securely. Never infer protected characteristics casually from names to create your own test labels.

Suppose an agency checks 30 illustrative CVs and finds that four candidates with relevant experience were ranked low because the tool treated volunteer work as no experience. Correct the criterion and recheck the full affected pool before sending the shortlist. If the supplier cannot explain how to correct the issue or show it was resolved, stop using the ranking for that role. For a deeper operational routine, the small-agency bias test checklist covers the recurring measurement rather than the initial ICO evidence file.

Keep accuracy separate from fairness. A parser that consistently drops a licence field is inaccurate. A scoring rule that penalises a type of career break may create unfair outcomes even if it reads every CV correctly. Both need tests. Record who accepts a change in scoring criteria, how a client request is checked and how earlier candidates are reconsidered when the rule changes mid-campaign.

Reduce collection, retention and access

Specify the minimum data needed for each AI function. Sourcing may need a contact route and a few role-relevant fields; it does not automatically need every historical CV, interview note or identity document. Screening may need a redacted CV rather than sensitive notes from prior placements. Set access so consultants see only the records required for their roles, and ensure supplier support access is controlled and logged.

Set a deletion and export procedure before the pilot. Ask which records are deleted when a candidate withdraws, which logs remain, how long backups persist and how a data request can be fulfilled. The 2024 ICO report described tools retaining excessive information indefinitely to build databases without candidates' knowledge. An agency should be able to state a retention rule and demonstrate the tool follows it. A “delete” button in the interface may remove a visible record while an export or training copy remains; ask the supplier to explain every copy.

A further example: a consultant pastes a CV into a general AI chat to draft a summary, then stores the output in the ATS. The agency may now have two extra copies outside its mapped workflow. Use an approved tool with known data terms or remove identifying information where appropriate. Record the resulting summary's source and review status. These checks apply even if the chat never produces a candidate score.

Run a release gate and a monthly evidence review

Use a simple release gate for each tool and use case. A named owner signs the DPIA, lawful-basis decision, supplier terms, candidate notice, retention rule, fairness test and human-review procedure. A missing item means the tool remains in a sandbox without live candidate records. This is an internal management rule, not a regulatory safe harbour. Keep copies of the artefacts in one controlled folder so a colleague can reproduce the decision.

Review the evidence on a calendar and after any material change: a new supplier model, additional data source, client requirement, automated rejection rule or retention setting. A monthly check can be short: inspect a sample of reviewed decisions, outstanding candidate corrections, supplier incidents and any drift in extraction errors. Assign an owner for each action and a deadline. If your agency cannot find the evidence during that review, the control probably does not operate reliably.

Add the 2026 automated-decision expectations to the file

The 2024 audit report is a starting point, not a freeze-frame of the rules. The Data (Use and Access) Act 2025 rewrote the automated decision-making provisions, and on 31 March 2026 the ICO published Recruitment rewired, built from conversations with more than 30 employers, together with a consultation on draft guidance about automated decision-making. Its expectations for anyone automating hiring decisions are short: monitor proactively for bias, tell candidates when automated decisions are used and how they work, and explain how to challenge a decision and request a human review. The ICO's general guidance on AI and data protection now carries a notice that it is under review because of the Act, so date-stamp any guidance you rely on.

Two details from that work belong in an agency checklist. First, the ICO describes asking developers about their own bias testing during procurement, and considering monthly bias reviews, as good practice; that matches the monthly evidence review above, so record the bias check in the same sitting. Second, the report found employers who believed their human involvement was meaningful and therefore put no safeguards in place at all. Run the 120-application illustration through that lens: if consultants only ever open the top 25 profiles, the 95 lower-ranked candidates may be on the receiving end of an automated decision, and the agency should take advice before assuming otherwise.

Turn those points into three extra lines on the release gate: a dated note of which ICO guidance version you checked, a candidate-facing route to request human review that someone actually monitors, and a named owner for the monthly bias review. Verify current regulatory guidance whenever you change a consequential screening process, and ask a qualified data-protection adviser where the agency's legal position is uncertain. That final check protects candidates and gives consultants a process they can actually follow.

Further reads

• AI CV Screening for Recruitment Agencies: Setup and Safeguards — Set up a practical CV-screening workflow.
• Questions to Ask a Recruitment Agency About Its AI Screening — See the questions a client may ask your agency.
• Recruitment CRM AI: Candidate Records, Duplicates and Follow-Up — Check candidate-record and follow-up controls.
• AI for Recruitment Agencies: Re-Engaging Your Candidate Database — Assess an older database before re-engagement.
• Recruitment Agency KPIs: A Weekly Scorecard From Your CRM Export — Build a small operational scorecard.
• Job Intake Form for Recruiters: Fill It From the Call With AI — A call-to-form method with a filled example, extraction prompt and confirmation email for ambiguous role requirements.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ICO, AI tools in recruitment: audit outcomes report (November 2024); ICO, Thinking of using AI to assist recruitment? Our key data protection considerations (6 November 2024); ICO, Recruitment rewired and its news release on automated decisions in hiring (31 March 2026); ICO, Guidance on AI and data protection (review notice, October 2026).