You post a job on Tuesday. By Friday, eighty CVs have come in. Some look strong on a quick skim; others clearly don’t fit. Many look roughly similar. Someone suggests an AI tool could sort the pile in a few minutes, and they’re probably right. But three of the most instructive things to know about AI in recruitment come from Amazon, Uber, and a video-interview company called HireVue. None of them had an easy time.
What are UK businesses actually using AI for in recruitment?
A 2023 Recruitment and Employment Confederation survey found that 48% of UK recruitment agencies had adopted some form of AI technology, up from 32% two years earlier. The practical applications are narrower than the marketing suggests: CV parsing, sourcing from job boards, chatbots handling candidate FAQs, and interview scheduling are the common use cases in both dedicated recruitment firms and in-house teams.
Koru Kids, a UK childcare platform, uses AI to screen applications and generate interview questions, shortening manual review time per application. Platform-level case studies across the sector report time-to-shortlist reductions of 30 to 70% when AI pre-screens CVs before recruiter review. These figures come from the platforms themselves rather than independent research, so they should be treated with appropriate caution. An OECD review of AI implementation across multiple sectors found that real productivity gains depended on process redesign, not just tool adoption. Firms that added an AI layer to unchanged workflows got far less back than firms that rebuilt the process around the capability.
Why do Amazon’s and HireVue’s mistakes matter to your business?
In 2018, Amazon scrapped an AI hiring tool after discovering it downgraded CVs mentioning women’s organisations, because the model had learned from ten years of male-dominated hiring data. HireVue, a video-interview platform, dropped its facial expression analysis in 2021 following regulatory pressure and scientific criticism of its predictive claims. Both failures share the same structural cause.
Amazon’s failure was self-created: the model learned from decisions that were already biased, then applied that bias at scale. An owner-managed business making twelve hires a year from a historically similar applicant pool faces the same risk in miniature. If you point an AI screening tool at your past hiring decisions and those decisions skewed toward a particular profile, the tool will reproduce that pattern faithfully.
HireVue’s failure was a different shape. The claim that facial expression analysis could predict job performance had no peer-reviewed evidence behind it. When researchers challenged the methodology, the feature could not withstand scrutiny. UK owner-managers looking at AI tools that claim to infer personality, cultural fit, or potential from voice or facial cues should ask the vendor for the evidence before signing up. That evidence is rarely there.
Where will you actually meet the legal risk?
The regulatory exposure sits at a specific moment: when an AI tool makes a rejection decision and no human reviews it. Under UK GDPR Article 22, individuals have rights where decisions based solely on automated processing produce legal or similarly significant effects. The ICO has confirmed that rejecting a job applicant solely by algorithm falls in scope. The risk is not theoretical; it has already reached the courts.
In 2021, a Dutch court ruled that Uber had to provide more transparency about its algorithmic systems and could not rely solely on automated processing to make decisions that significantly affected workers. The case involved delivery riders rather than traditional employment applicants, but the legal principle applies equally to hiring. The moment an algorithm makes a binding decision about a person without human review, the rules apply.
The Equality Act 2010 adds a second layer. Employers remain liable for discriminatory outcomes even when an automated tool produced them. The Equality and Human Rights Commission has been explicit: responsibility does not transfer to the vendor. If your AI screening process disproportionately filters out applicants from a protected group, the legal exposure is yours. A practical test is straightforward: if AI rejects a candidate and no human ever reviews that decision, you are in solely-automated territory. If AI shortlists and a person then approves or overrides, you are not.
When does AI help in hiring and when should you hold back?
AI earns its place in recruitment when it handles administrative work before a human makes the meaningful decisions. CV parsing, first-stage message drafting, interview scheduling, and screening against factual eligibility criteria are low-risk starting points. These free recruiter time for the parts that require judgement: the telephone screen, the reference check, the offer negotiation.
The CIPD’s 2023 People Profession survey found that HR leaders are most comfortable using AI for scheduling, basic eligibility screening, and drafting, and least comfortable with AI making final hiring decisions. The UK Civil Service, through the Government Digital Service and Civil Service Fast Stream, has piloted AI-supported sifting for graduate recruitment: AI scores work-sample submissions, a human reviews the ranked output, and the human makes the shortlist decision. There is no fully automated rejection in this approach, and documentation exists at every stage.
For an owner-managed business, the equivalent is simpler. Give the AI a clear set of criteria: qualifications required, specific experience needed, practical requirements that relate directly to the role. Instruct it explicitly not to consider personal characteristics. A person then reviews the AI’s ranked list before any candidate is rejected. The ISE’s UK case compilation recommends exactly this: start with narrow deployments and ensure a human validates AI-shortlisted candidates before any rejection is issued.
Hold back when you are making a final decision, when the criteria involve soft judgement, or when you cannot explain how the tool reached its output. If a candidate asks how they were assessed and you cannot give a clear answer, the tool is too opaque to rely on.
What do you need in place before you start?
The ICO and the Centre for Data Ethics and Innovation both stress documentation as the core accountability requirement for algorithmic decision-making. For an owner-managed business, this means a one-page description of how AI is used in your recruitment process, the criteria and prompts you give it, and a log of periodic spot-checks comparing AI recommendations against your own judgement.
If you are processing CVs, covering letters, or any personal information through an AI tool, the NCSC recommends treating it with the same security discipline as any other system handling sensitive data. Feeding full CV documents or right-to-work evidence into an unmanaged consumer AI tool risks both a data breach and a UK GDPR violation of the data minimisation principle.
A data protection impact assessment is recommended by the ICO, and often mandatory, when using AI for systematic evaluation of personal characteristics. For a small owner-managed business running a single role through an AI screening tool, a brief written assessment of the risks and your safeguards will usually satisfy the requirement.
One further consideration: the EU AI Act classifies AI systems used for recruitment or selection as high-risk, with obligations including risk management, transparency, and human oversight. UK businesses recruiting EU candidates using AI screening tools may face these requirements as the Act phases in from 2026. If your hiring crosses borders, it is worth tracking.
The practical start is a short document. Record what AI you are using, why, what criteria it applies, and who reviews its output. That record protects you legally, keeps the process auditable, and means you can explain your approach clearly to any candidate who asks.
If you would like help thinking through where AI fits in your hiring process, or whether your current recruitment setup is carrying risks you haven’t priced in, Book a conversation.
