Edition
Download PDF
KITH White paper · Going aloneRunning an agency · Version 1, November 2026
Recruitment and AI

Agents search. You make the call.

Where AI pays in recruitment, and what should stay human.

Recruiters lose hours every week just finding people, while clients drown in applications. Much of the AI sold to fix it stalls. Here is where AI pays, and what should stay with people.

Every figure is published research, credited and dated; our own arithmetic is marked. Written for UK permanent and search recruitment; most data on candidates’ attitudes is from US or global surveys, and we say which. Kith sells the service this paper describes.

Where the week goes

Fourteen hours a week, just finding people.

Recruiters say they spend 14.6 hours a week searching for candidates:1 more than a third of a 40-hour week (our arithmetic). Meanwhile LinkedIn applications rose 45% in a year, to 11,000 a minute.3 That’s the client’s problem, and it makes a short, checked list worth more: a sourced candidate is five times as likely to be hired as one who applied, 2.0% against 0.4%, at in-house teams on Gem, a vendor.4

A recruiter’s week 14.6 h the rest of the week Searching for candidates 0 10 20 30 40 hours
Weighted average of self-estimates, 1,500+ recruitment professionals worldwide (Bullhorn GRID 2025, a vendor survey).1 The 40-hour week is our assumption.
Where AI stalls

Point AI at the right work.

Many AI projects stall, and the reasons firms give are cost, unclear value, risk and data. We read that as AI bolted onto work nobody redesigned: agents aren’t simply plug-and-play: companies urgently need to redesign how work gets done.7 The work to hand over is search, the first sift, scheduling and notes.

40%+

of agentic AI projects will be cancelled by the end of 2027, Gartner predicts, due to escalating costs, unclear business value or inadequate risk controls. Gartner adds: Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied.5

42%

of companies were abandoning most of their AI initiatives, up from 17% a year before. All sectors, North America and Europe; reported March 2025.6

10%

of staffing firms run agentic AI across their whole workflow. The barriers they name: data readiness, security, unclear plans. Nearly 2,300 surveyed; a vendor.8

None of this says people judge CVs better than software. In one study of 15 firms, managers who appeared to hire against test recommendations ended up with worse average hires, mainly shorter-staying.9 The case for people is the rest of the job.

What stays human

Keep the call.

Keep the call, the relationship, persuading and closing, and accountability for who goes forward. The hardest part of today’s market is getting people to move, and that’s a human job.

Prefer people to AI 58% 68% Two years earlier Q3 2025
Candidates, Gartner, 2,901 surveyed; geography not stated.10
71%

of US adults oppose AI making the final hiring decision; 7% favour it. December 2022.11

permanent candidates remained reluctant to move jobs … whilst technology and AI are powerful tools, human interaction is vital to deliver the most successful recruitment outcomes for both clients and candidatesPageGroup, half-year results, August 202512
Trust. Taste. The call.

That stays yours; agents do the searching. Taste is knowing what the client never wrote down, and hearing on a call whether someone will really move.

Kith · Agents search. You make the call.1 / 3
KITH Edition: Going alone
Your own desk

You can bill alone. Running the business is the hard part.

At an agency, most of what you bill pays for the building. In one benchmark of UK agencies, wages took 67% of net fee income and overheads 25%, leaving 7.5% profit against the 20% it recommends.13 One accountant finds high-performing firms average a 3x return on consultant cost in fees billed.14 Alone, you keep more of it. You also take on everything else, and many new firms don’t last.

Where an agency’s fee income goes Wages 67.0% Overheads Profit 7.5% 25.4% Share of net fee income, UK, H1 2026
Recruitment Accountants with APSCo; the firm’s own clients, sample not published. Profit is EBITDA.13
New UK placement firms still trading two years on 48%58%56%76% 2017201820192020
Firms born each year, SIC 78109, VAT- or PAYE-registered; may include payroll vehicles. Our arithmetic from ONS births and survivors (2017: 2,550 of 5,265; 2020: 1,736 of 2,285).15

What going alone costs, before you resign

  • Insurance: professional indemnity and liability.
  • Tools. An agency ATS lists at $95–$215 a user a month;16 Bullhorn’s Starter at $99.17 List prices, excluding tax.
  • An accountant, and your own National Insurance and pension.
  • The ramp-up before your first fee lands.
  • Cash flow on 30–60 day payment terms.
  • Rebates when a placement leaves early.

Check your contract’s restrictive covenants. Never take your employer’s candidate or client data.

The usual fixes

Why the usual fixes only close part of the gap.

OptionFits how you workWho keeps it runningWhat you keep if you leaveCost shapeWhere it’s weaker
Buy a stack of toolsYou fit the toolsEach vendorYour data, if it exportsPer seat, per monthLogins that don’t talk to each other
A virtual assistant or offshore sourcerWith trainingYou manage themTheir notesPer hour or monthYour time managing; knowledge leaves with them
Build it yourself with AI coding toolsYes, at firstYou, between callsEverything you builtYour time, then upkeepSecurity and upkeep; it stops when you’re busy
A partner who builds and runs it (Kith)Built around youKithYour data, exported any timeA build, then monthlyYou rely on a small supplier; there’s a build first

Kith’s view, by category; no vendor is named or rated.

Build the agency around you.

You run the desk. You close. Kith builds and runs the search, the first grade and the tools around you.

Kith · Agents search. You make the call.2 / 3
KITH Edition: Running an agency
A flat market

Do more with the team you have.

Margins are thin and permanent volumes are down. In one benchmark of UK agencies, profit ran at 7.5% of net fee income, against the 20% it recommends,13 and six in ten firms said their margins were squeezed in 2023.18 The big firms already run fewer consultants, each doing more: Hays reports group-wide consultant productivity up 5% in a year.20 In a flat market, the gain from AI is the same output with fewer hours off the phone, and better shortlists. It isn’t “more calls, more fees”.

Fewer placements. Fewer consultants. Hays UK & Ireland consultants 1,629June 2024 1,285June 2025 UK permanent placements (REC estimate) 806,4002023 536,4002024
Each measure on its own scale from zero. Hays FY25 investor deck;20 REC estimate, modelled by Cebr from 186 recruiters.19
£2.1m

PageGroup’s operating profit in the first half of 2025, against £28.4m a year before.12

37%

of organisations are integrating or experimenting with generative AI in hiring, up from 27% a year before. Mostly in-house teams; a vendor survey.2

10 × 15 h × 46 weeks = up to 6,900

non-billable hours a year, for ten recruiters. The 15 hours is our assumption, informed by Bullhorn’s 14.6;1 46 weeks is 52 less the 5.6-week statutory minimum leave, rounded down; our arithmetic.21

The hours come back. Decide with your team what they’re for: more searches each, better shortlists, or time on the phone.

The stack

The ATS isn’t the problem. Everything bolted round it is.

66%

of agency recruiters rate their ATS 4 or 5 out of 5. 57% name too many tools that don’t integrate; 67% have delayed switching software because of tech fatigue. US, 239 recruiters, a vendor survey.22

45%

of tests found AI-written code introduced risky security flaws.25 A scan of over 5,600 public vibe-coded apps found 2,000+ vulnerabilities, 400+ exposed secrets and 175 instances of personal data.24 Both vendors.

67% vs 33%

Tools bought from partners who customise them reached deployment about twice as often as internal builds. Self-reported by 52 organisations; it doesn’t prove cause, and it’s contested.26

Vibe coding closes the fee gap, not the upkeep gap: 66% of developers say AI’s answers are almost right, but not quite.23 That leaves AI-built apps nobody has reviewed, holding candidates’ personal data. We build with AI coding tools too, with code review and tests.

OptionFits how you workWho keeps it runningWhat you keep if you leaveCost shapeWhere it’s weaker
Buy another toolYou fit the toolThe vendorYour data, if it exportsPer seat, per monthAnother login that doesn’t talk to the rest
AI added to your current toolsPartlyThe vendorYour data, if it exportsAn add-on per seatBuilt for every customer, not you
AI-native point toolsOne job, wellThe vendorYour data, if it exportsPer seat or usageAnother silo to join up
An offshore sourcerWith trainingYou manage themTheir notesPer hour or monthManagement time; turnover
Vibe-code it or build in-houseYes, at firstYou, or nobodyEverything you builtYour time, then upkeepSecurity, upkeep, one person’s knowledge
A partner who builds and runs it (Kith)Built around youKithYour data, exported any timeA build, then monthlyYou rely on a small supplier; there’s a build first

Kith’s view, by category; no vendor is named or rated.

Kith · Agents search. You make the call.2 / 3
KITH The rules, the model, the sources
The rules

The law expects real oversight.

The law doesn’t say a person must make every hiring decision. It does say oversight has to be real: a rubber stamp isn’t oversight. UK and EU law here; US rules, such as New York City’s Local Law 144, differ and aren’t covered here.

EU AI ActAI systems for recruiting and selecting people are high-risk, and always so when they profile people. Providers must build them so overseers can disregard, override or reverse the output; deployers must give oversight to people with the necessary competence, training and authority, and tell people when it’s used on them. From 2 December 2027.27
GDPRArticle 22 limits decisions based solely on automated processing; Recital 71 names e-recruiting practices without any human intervention.28
SCHUFA rulingA score that another party relies on heavily can itself be an automated decision. Court of Justice of the EU, December 2023.29
UK, since 5 February 2026A decision is solely automated if there is no meaningful human involvement. The UK relaxed the rule, with safeguards; decisions based on special-category data are still restricted.30
Now May 2018GDPR applies Dec 2023SCHUFA ruling Feb 2026UK safeguards Dec 2027EU high-risk duties
The model we’re betting on

AI built around how you work, run by a partner.

People keep the calls, the relationships, and the decisions on who to contact and put forward. Here is what “AI-native” means at Kith, in things you can check:

  • Software Kith builds and runs does the search and the first grade.
  • A person decides who to contact and who to put forward.
  • Every recommendation carries its evidence, caveats and source.
  • Every decision is kept, and you can say why. Every brief is versioned.
  • Your data exports any time.
Going alone? Book your audit before you hit send. Count your hours off the phone. Then book your audit.

kith.contact

Sources and method

Version 1, November 2026. Figures as published, quotes as written, our arithmetic marked. “Vendor”: the publisher sells to the people it surveyed. Figures we couldn’t check against the primary source or a verbatim reprint aren’t used; where a primary page was blocked, we say so. Contains public sector information licensed under the Open Government Licence v3.0. Stack Overflow data under the Open Database License. Pew Research Center bears no responsibility for the analyses or interpretations of the data presented here.

  1. Bullhorn, GRID 2025 Industry Trends Report, 2025. 1,500+ recruitment professionals worldwide. Vendor. bullhorn.com
  2. LinkedIn, Future of Recruiting 2025. Survey September 2024; 1,271 recruiting leaders, mostly in-house, 23 countries. Vendor. business.linkedin.com
  3. Semafor, reporting LinkedIn figures from the New York Times, 24 June 2025. semafor.com
  4. Gem, 2025 Recruiting Benchmarks, 16 January 2025. In-house teams on Gem, 2021–2024. Vendor. gem.com
  5. Gartner, press release, 25 June 2025, as reprinted verbatim by HPCwire. A prediction. hpcwire.com
  6. S&P Global Market Intelligence, via CIO Dive, 14 March 2025. 1,000+ respondents, North America and Europe. ciodive.com
  7. BCG, press release, 30 September 2025. 1,250 executives. Vendor. bcg.com
  8. Bullhorn, GRID 2026, press release, 25 February 2026. Nearly 2,300 respondents, November–December 2025. Vendor. bullhorn.com
  9. Hoffman, Kahn & Li, “Discretion in Hiring”, Quarterly Journal of Economics 133(2), 2018. 15 firms, low-skill service jobs; worse means mainly shorter tenure. doi:10.1093/qje/qjx042
  10. Jamie Kohn (Gartner), HR Dive, 20 January 2026. Gartner Q3 2025 survey of 2,901 candidates. hrdive.com
  11. Pew Research Center, “AI in Hiring and Evaluating Workers: What Americans Think”, 20 April 2023. 11,004 US adults, 12–18 December 2022. pewresearch.org
  12. PageGroup, Half Year Results, 12 August 2025. investors.michaelpage.com
  13. Recruitment Accountants with APSCo, benchmark for H1 2026, 3 September 2026. The firm’s own clients; sample not published; the 20% is its own model. uk.apsco.org
  14. Recruitment Accountants, Q1 2025 benchmark. A rule of thumb. recruitmentaccountants.com
  15. Office for National Statistics, ad hoc AH1574, 10 October 2024. Our arithmetic. ons.gov.uk
  16. Recruit CRM, pricing, Pro to Enterprise, checked 8 November 2026. recruitcrm.io
  17. Bullhorn, pricing, Starter, checked 8 November 2026. bullhorn.com
  18. REC, Industry Status Report 2022/23. Survey September–October 2023, 274 members. rec.uk.com
  19. REC, press release, 8 December 2025. Estimates modelled by Cebr from 186 recruiters. rec.uk.com
  20. Hays, FY25 investor deck. Productivity is group-wide. haysplc.com
  21. GOV.UK, Holiday entitlement. gov.uk
  22. Top Echelon, 1 August 2025. 239 agency recruiters, US. Vendor. topechelon.com
  23. Stack Overflow, 2025 Developer Survey, AI section. Base 31,476. survey.stackoverflow.co
  24. Escape.tech, 29 October 2025. A passive scan; the authors call it a lower bound. Vendor. escape.tech
  25. Veracode, 2025 GenAI Code Security Report, October 2025. 100+ models; Java, JavaScript, Python and C#. Vendor. veracode.com
  26. MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025. 52 organisations; hybrid builds not measured. mirror at mlq.ai
  27. EU AI Act, Regulation (EU) 2024/1689 as amended by 2026/1744: Annex III 4(a), Articles 6(3), 14(4)(d), 26(2), 26(11). eur-lex.europa.eu
  28. GDPR, Article 22 and Recital 71. gdpr-info.eu
  29. Court of Justice of the EU, C-634/21 (SCHUFA), 7 December 2023. eur-lex.europa.eu
  30. Data (Use and Access) Act 2025, section 80; in force 5 February 2026. legislation.gov.uk