AI for recruitment agencies: what actually works, from people who run it.

The Access Group’s AI in Recruitment 2026 Outlook, a survey of 159 agency leaders across the UK and Australasia published in July 2026, found that 78 per cent of agencies had adopted AI in some form and only 6 per cent had embedded it in their core operating model. Seven in ten used it for writing or optimising job adverts; a quarter used it for reporting. The blocker agencies named most, at 74 per cent, was poor data and disconnected CRMs. The same survey found agencies that had invested significantly in training reported far stronger returns than those that had not.

That matches what we see. Writing adverts is where every agency starts because it is the easiest thing to do with a chat window. It is also the lowest-value thing, because the advert was never the bottleneck. The bottleneck is what happens in the hours after someone applies, and that is where the evidence of AI paying back is strongest.

Candidates go cold fast. Greenhouse research reported in the UK trade press in 2024 found that 45 per cent of candidates had been ghosted after an initial recruiter contact and 42 per cent after an interview, with candidates from under-represented backgrounds more likely to be ghosted. The public does not trust AI to decide: the ICO’s 2025 research found 64 per cent of people expected over-reliance on AI in recruitment and 61 per cent expected worse decisions than a human would make, and Totaljobs research collated by StandOut CV found three-quarters of Britons think AI use in hiring should be disclosed and 80 per cent oppose AI making the final decision.

Put those together and the design brief writes itself. Use AI to respond in minutes rather than days, to keep every applicant informed, and to get the good ones in front of a recruiter quickly. Do not use it to decide, and say what it is doing. Our guide on why applicants go cold sets out the timing data in detail.

Screening at the top of the funnel. Every application read, scored against the brief and the client’s must-haves, and sorted, as it arrives, at any hour. On our own platform a CV is read and scored in under two seconds in the typical case, which means the recruiter starts the morning with a ranked list rather than an inbox. The score is advisory: a recruiter decides who is called, and the reasons for the score are shown so the recruiter can disagree.

Chasing by message. The applicant who has not completed their profile, not confirmed availability or not replied to a question gets a message in the window when they are still engaged, over WhatsApp, SMS or email, and the conversation is handled until it needs a person. This is the single change that moves the numbers most, because it turns a pile of half-finished applications into a shortlist without a recruiter spending the day on it.

Booking. Interviews with the client booked from the candidate’s side, with reminders and rescheduling handled, without a recruiter awake. The reminder matters as much as the booking; a booked interview nobody attends costs more than an unbooked one.

The phone. An AI voice agent that answers inbound calls when the desk is busy and makes outbound screening calls. This is the job with the most hype and the most caveats, which the next section covers.

Voice is where agencies want to start and where they should be most careful. From running outbound screening calls at scale, roughly one call in eight reaches a human on the first attempt; the rest go to voicemail, ring out or are declined. That single fact should shape the design. The system has to be built around the seven that do not answer, with a message follow-up, a retry schedule and a route to a recruiter, rather than around the one that does. We design the fallback before we design the voice.

The calls that do connect work best when the agent is honest about being an AI, keeps the call short, asks the three or four questions that actually gate the role, and books the next step with a human. Our guide on whether AI calling works publishes the outcome data we are prepared to stand behind.

Across the platform we run, the large frontier model handles roughly one call in eighty; everything else, including reading and scoring applications and handling messages, runs on the small model, which is several times cheaper and faster with no loss in outcome once the workflow is designed properly. The cost that agencies underestimate is the channel: a text message or a minute of telephony costs more than the model behind it. So a realistic budget starts with volume of messages and calls, not with a per-token price.

For an agency buying rather than building, the published UK market for a first AI implementation is £4,000 to £15,000 at fixed-price suppliers, with running costs that scale with candidate volume rather than headcount. Our running-cost and agent-cost guides give worked examples; the pillar page sets out our own prices against the market.

Recruitment is the use case regulators are watching most closely, because decisions about individuals are involved. The ICO’s AI guidance expects a lawful basis for processing candidate data, transparency about automated processing, and a DPIA where profiling is systematic, which candidate scoring is. The rules on significant automated decisions, Articles 22A to 22D of UK GDPR as amended by the Data (Use and Access) Act 2025, require safeguards where a decision about a person is made by automated means: the candidate is told, can contest it and can obtain human review. In practice a human makes the shortlisting decision and the candidate can ask for review. CIPD’s employment-law guidance adds discrimination risk: a model trained on past placements can reproduce past bias, and the agency, not the vendor, answers for it.

In practice this means four things: a privacy notice for candidates that says AI is used and for what; a DPIA for the screening workflow; scores that a recruiter can see the reasons for and overrule; and monitoring of outcomes by protected characteristic, so that bias is found by you before it is found by a tribunal. The AI acceptable-use policy template in our guides has a recruitment clause for exactly this.

Fix the data first, or at least know how bad it is. Three-quarters of agencies name disconnected CRMs as the blocker, and an AI that cannot write back to the CRM creates a second system nobody trusts. Pick one desk and one job type with real volume, and put screening and chasing on it for six weeks with the recruiter still deciding. Measure time to first contact, completion rate of applications, and interviews booked per hundred applicants, before and after. Then decide whether to buy a platform, extend the one you have, or build. Only after that, add voice.

The agencies that reported the strongest returns in the 2026 survey were the ones that trained people. The tool does not change what a recruiter does with a shortlist; it changes how fast the shortlist appears. If the desk does not change its rhythm to match, the gain disappears into the same afternoon it always did.