Skip to content

UK AI Risk Management Toolkit: a worked recruitment example

A useful risk review asks what could go wrong, how you would detect it and what would make you stop the pilot.

· 11 min read

Blue shield with a white check mark

Researched 30 October 2026 · By Screened

“How would we find someone the tool wrongly rejected?”

A strong shortlist shows who made it through, not whether everyone else was screened out correctly.

Plan how to catch a wrong recommendation and correct it. This recruitment example shows how to use the UK AI Risk Management Toolkit to record a risk, test the safeguards and decide whether a pilot can proceed.

The government published the toolkit on 8 September 2026. It is guidance for assessing risks, not approval of a product or a replacement for the privacy review your project needs. Official toolkit

What does the UK AI Risk Management Toolkit help you do?

The toolkit helps teams identify risks, assess their likelihood and impact, and decide who will address them. It includes Excel and OpenDocument workbooks. The guidance also allows you to use an existing risk register that captures the relevant information. Read the official guidance

The guidance is built around UK public-sector risk management. Private recruiting teams can adapt its method, but it is not a new legal requirement for every UK employer or a Singapore hiring team.

It does not certify products or list approved suppliers. Use it alongside the privacy, security and legal reviews your proposed use needs. Practise with test CVs before sharing real candidate information.

How do you start an AI recruitment risk assessment?

Describe the task before rating the risks. What will the tool do, which candidate information will it receive and what decisions or messages could follow its result?

For a support-operations vacancy, recruiters write CV summaries, but the manager still rereads every application to understand the recommendation. The pilot tests whether clearer summaries reduce that repeated work.

Define the tool's role:

We want the tool to locate relevant experience and explain its assessment. A recruiter will compare the report with the CV and recommend the next step. The initial pilot will not automatically reject applicants, move applications or send messages.

Ask the proposed reviewer to try the process on test material. Can they open the CV, check the assessment and decide the next action? Measure the effort, then make the pilot small enough for that review to happen.

Compare it with a clearer manual brief too. Reduce repeated work without assuming another tool is the only way to do it.

Check what happens after the report arrives

Check what the report can trigger. A recommendation for a recruiter to review is different from a result that immediately sends a rejection. The risk assessment needs to include those actions, not just whether the summary reads well.

Have the administrator follow a test application through both systems, using contact details your team controls. Check shared settings before changing them. When the report returns, inspect the application stage and pending messages as well as the inbox.

Record which actions are permitted in this trial. Testing summaries is not approval to start contacting or rejecting candidates automatically.

Test whether the tool recognises relevant experience

The support-operations job requires experience investigating billing complaints and recommending a response. Finance approves refunds. The team uses Zendesk and will teach the new hire how to use it.

Compare these three CV passages with the manager before looking at the AI scores:

What the CV says What it tells you
“Investigated hotel billing complaints, checked receipts and recommended refunds. My supervisor approved the refunds.” Describes relevant investigation in a different setting. The applicant did not give final approval.
“Used Zendesk daily to log billing complaints and pass them to specialists.” Describes software familiarity. The passage does not establish that the applicant investigated the complaints.
“Responsible for customer billing issues.” Leaves the person's contribution unclear. You may need an example.

The hotel passage describes the relevant investigation. The Zendesk passage describes using the familiar system, but not doing that work. The third leaves the responsibility open. Agree those distinctions with the hiring manager before reading a score.

Investigate an unexpected result before changing the risk rating. The tool may have misread the CV or followed an incorrect instruction. These passages help you distinguish those problems; a complete hiring decision still needs the rest of the application.

What should you write in the recruitment risk register?

The report recommends rejecting the hotel applicant for lacking Zendesk experience. The saved instruction requires that software even though the manager agreed to teach it. Correct the instruction, then retest all three passages. Run this test on sample data rather than a live hiring decision.

Add the finding to the risk register, the list of risks, safeguards and people responsible for them. This sample entry describes a successful retest while making clear that another review check remains unfinished:

Question Recruitment use
What could happen? Applicants with relevant investigation experience are excluded because they used different software.
Why did this test fail? The original instruction required Zendesk experience, although the job brief said the company would teach it.
What has changed? The instruction now accepts relevant investigation experience in other settings. A recruiter will still check each assessment before a hiring action.
What did the test show? All three test passages now receive the expected treatment. A second recruiter has not yet tried the review process.
What still needs testing? Whether another recruiter can open the CV with their usual account, recognise a mistake and complete the review.
Who acts next? The pilot lead arranges the second recruiter's review. The administrator confirms that automatic actions remain off.

Test the proposed safeguard next. Give another recruiter the practice CV and deliberately wrong assessment without the answer guide, using their normal account. Ask them to review it, explain the issue and follow the route to correction.

Watch where the reviewer gets stuck. Access to the CV, a way to report the mistake and authority to stop the next action are separate needs. Fix the missing step, then repeat it.

One colleague catching a planted error does not establish an error-detection rate. It shows whether the review can work in this setup and what support it still requires.

Test errors, access and candidate messages

A correct answer can still reach the wrong person or trigger an unwanted message. Continue with the same vacancy:

Problem to test What a useful result would show
A CV cannot be read. The application remains visible, with someone assigned to obtain a usable file or review it another way.
The report says the applicant approved refunds, but the CV says their supervisor did. The reviewer can report the error and make sure the manager sees the corrected description. Check whether an automatic action still uses the old result.
A result triggers an invitation. Your administrator can identify and stop the rule, check queued messages and explain what was already sent.
An interviewer has access to unnecessary candidate information. The permissions can be corrected and checked using that interviewer's account.
Reading takes less time, but copying and correcting the report adds work. The pilot measures that extra work instead of reporting only generation speed.

Add checks for who receives candidate information, how long it is kept and who can open it. For voice interviews, include interrupted calls, requests for another format and the person responsible for helping candidates.

Test how the team would stop further submissions and handle assessments already running or messages already queued. Pausing one step may leave the others active.

Keep candidate follow-up assigned while the supplier investigates. Your team still needs to correct misleading briefs, explain mistaken messages and send promised updates.

Review exclusions as well as recommendations

A manager liking the shortlist tells you little about the applicants left out. The Information Commissioner's Office (ICO) describes this gap in its findings on human involvement in automated recruitment. ICO: meaningful human involvement

For this initial pilot, we recommend checking proposed rejections before they take effect. Include unreadable CVs and missing reports rather than reviewing only completed scores.

Have the reviewer point to the relevant evidence and explain the next step. Resolve disagreement against the source and job requirement; the tool, reviewer or brief may be wrong. Count that review time in the business case.

How do you turn the test results into a risk assessment?

The test log records a fault and its repair. It does not yet explain how serious the risk is or whether the pilot should start.

The toolkit's guidance uses likelihood and impact scores from 1 to 5, multiplied to produce a risk score. Put the reasoning beside the numbers. A successful retest is not, by itself, a reason to mark the remaining risk low.

For the same pilot, the assessment could record:

Part of the assessment What this example supports
Who could be harmed? A qualified applicant could lose the chance of an interview because their experience was gained with different software. The employer could miss relevant experience.
How likely is that? The test demonstrated one way it could happen: an incorrect requirement. The revised instruction fixed those examples. That does not tell us how often other mistakes will occur in real applications.
How serious could it be? An incorrect draft recommendation caught before action is different from a rejection already sent. Assess the risk of the final hiring decision, including whether the team can find and reconsider an affected application.
What reduces the risk? The instruction has been corrected. Recruiter review and manual decisions are planned, but the second recruiter's access and ability to catch the error still need testing.
What remains unresolved? We have not yet shown that another recruiter can complete the review. We also have no estimate of mistakes across the full range of applications.
Decision and responsibility The pilot lead completes the outstanding review test before requesting approval. The person authorised to approve the pilot records the risk ratings, reasons, remaining limits and next review date. The live pilot has not started.

Choose the likelihood and impact ratings with the colleagues responsible for this assessment, and explain the evidence behind them. The three CVs expose a possible failure. They do not tell you how often it will occur.

Use that uncertainty to limit the trial. Finish the second recruiter's review test before starting, and keep the first group small enough for the team to check every result. Removing those checks to improve the time-saving calculation would mean assessing a different, less supervised process.

Does the toolkit replace a recruitment DPIA?

No. A data protection impact assessment (DPIA) examines risks to people from the proposed use of personal data. Your privacy colleague should determine whether it is required and complete it before the relevant processing begins. The general AI workbook serves a different purpose. ICO: DPIA guidance

Both assessments can use the same explanation of where CVs go, who can read them and how candidates can question a decision. Keep that explanation consistent, while tracking the checks and approvals each assessment needs.

The ICO's 31 March 2026 recruitment update addresses the automated-decision rules under the Data (Use and Access) Act. It calls for transparency, bias monitoring and a way for candidates to challenge decisions and request human review. Ask your legal reviewer to examine the current requirements and each decision your hiring process makes. A human final interview does not explain what happens to applicants rejected earlier. ICO's recruitment update

Test the contact route described in your privacy notice. A recruiter should be able to receive the question, find the application and arrange the appropriate review.

End the assessment with a decision about the trial

Approve the next step supported by the evidence, not the product for every possible use. You may need more testing or a narrower pilot. For this vacancy, the recommendation could read:

I recommend a limited pilot for the support-operations vacancy once the privacy and access checks are complete. A recruiter will check every report and handle applications without a usable result.

Messages, stage moves and rejections will remain manual. We will pause affected actions for an unauthorised disclosure, a material assessment error or an action taken without approval.

We will compare completed reviews, errors and total staff effort with our current process before adding another role.

List the start conditions and who will confirm each one. A conditional approval does not complete the privacy review or the access test. The person starting the pilot needs confirmation that those conditions have actually been met.

Match the first group to the review time available and arrange cover. Keep the tests for later changes. Adding another role, more candidate information or automatic decisions changes the proposed use; the original approval should not silently expand to cover it.

Ask recruiters whether they still check the source after repeated use. The Cabinet Office's human-centred AI toolkit examines risks in reliance on tools as well as technical failure. Human-centred AI toolkit

Apply the same tests to Screened. Our voice interviews provide transcripts, topic summaries and assessments. Have the reviewer check an important claim against the original answer and explain their next action.

Before starting, make sure the person approving the pilot and the recruiter understand the same plan: who checks proposed exclusions, who resolves disagreements and who reconsiders affected applications. Name any step not yet tested.


Use this example to prepare your own assessment with your hiring, privacy and security colleagues. They will need to confirm the requirements and approvals for your proposed use.