The AI Agents Table:  what recruitment leaders actually think about AI agents and where we are today.

Where are you with AI agents today?

Everyone's talking about AI agents. The technology is moving fast, the possibilities seem endless, and the advice out there often feels contradictory. Ask five people what “agentic AI” means and you'll get five different answers, most of them confident, few of them the same.

So instead of adding another opinion to the pile, Kyloe brought together a group of recruitment leaders in London for a frank, closed-door conversation to look at where people actually are with AI agents, where opinions genuinely split, and where there's more agreement than anyone expected.

What came out of it wasn't a tidy list of answers. It was a series of live debates - the kind you have with your own team, if you're brave enough to have them out loud. Here's what came up, and where we landed.

Where the room actually started

Nobody in the room was standing still. Just over 40% described themselves as still experimenting, a third had built something in prototype, and a quarter were already live, with agents doing real work day to day.

But here's the number that mattered more: only one in six said they'd properly locked down “shadow IT” risk - the personal, unsecured AI accounts that staff use without anyone signing off on it. Anything typed into one of those accounts could end up training that model. If candidate data is in there, that's a business risk hiding in plain sight, not a hypothetical one.

Interestingly, nobody in the room was chasing AI to replace people. The ambition was capacity and competitiveness; doing more with the same headcount, staying competitive in the market, and improving the day-to-day experience for consultants. Cost and time, not appetite, was the thing holding people back.

This sparked 6 questions or debates.


One AI tool for everyone, or let people choose?

The case for one tool: it's simpler to manage, easier to secure, and easier to support. Some businesses have picked a single platform, built it into onboarding, and quietly switched off everything else.

The case for choice: the market is moving too fast to bet on one winner. Different tools are genuinely better at different jobs, and forcing everyone onto one option usually means picking someone's least-favourite tool, whatever you choose.

Kyloe's take: there's no universally right answer here, but there is a wrong one - never actively deciding. Whichever way you lean, make it deliberate. If you do standardise, build it into onboarding from day one, so new starters never learn the wrong habit in the first place.

Lock everything down, or let people experiment?

This one turned out to be less of a two-sided argument and more of a sequencing problem.

Anything typed into a personal, unsecured AI account may be used to train that model, and if it includes candidate data, that's on you. If someone leaves the business, their chat history can effectively walk out the door with them. But locking things down without a workable alternative doesn't remove that risk - it just pushes people toward free, unmonitored tools instead, which is worse.

Kyloe's take: give people a secure, paid option before you restrict anything. Update your handbook and contracts so AI use is explicitly covered. Put a plain-language AI policy on your website so clients and candidates know what you do and why. And decide your risk tolerance on purpose, in writing, rather than letting it happen by accident.


Is it agentic AI, or just clever automation?

A recent industry claim that 30% of recruiters use agentic AI was met with real scepticism in the room. True agentic AI decides and acts on its own. Most of what gets called agentic today is closer to a scheduled prompt. Useful, but pre-programmed rather than genuinely independent.

Does the label matter? We think so, because it changes how much oversight something needs.

A scheduled task following fixed rules needs a different level of checking than something making live decisions on its own. Before anyone reports their own AI adoption number, it's worth agreeing internally what that number actually means - otherwise it's meaningless, even to the person quoting it.


Variable costs. Real fear, or overblown concern?

Pricing has shifted from predictable subscriptions to token-based billing, and nobody in the room could give a straight answer to “how much will this actually cost us?”

Two stories stood out. A small business ran up a very large bill in a single weekend on an unsupervised routine. A much larger company burned through a full year's AI budget by April.

Kyloe's take: the fear is reasonable, and it doesn't have to stop you. The businesses managing this well aren't the ones spending the least, they're the ones with visibility.

Budget for spend to roughly double every six months, put hard alerts on every credit pool you use, and never let an unsupervised process run without a time or spend limit attached.


AI screening. Efficiency win, or candidate risk?

In high-volume, skills-scarce recruitment, response times matter enormously. One example quoted a 45-second gap between a candidate applying and being engaged, a better experience than most human teams could deliver at that scale.

But candidates are showing real resentment about being screened by AI before speaking to a person, and there's a growing arms race: candidates using AI to write applications, agencies using AI to screen them. That risks burying a genuinely well-suited candidate who simply didn't optimise their CV.

Kyloe's take: this is a fit question, not a for-or-against one. High-volume, quick-turnaround roles suit AI screening. Specialist and senior roles usually don't, and forcing it there does more brand damage than the time it saves. If you use it, build in a human verification step, and be honest with candidates about when they're talking to a bot.


Two things the room actually agreed on

Not everything was a split decision. Two ideas had real, broad support.

Treat agents like a new hire, not a piece of software. Give every agent a clear, plain-language description of what it can and can't touch. Default to read-only access. Write access should be a deliberate decision. Delete access should be rare to non-existent, and “never delete” instructions should be stated clearly at both the start and end of any prompt or brief.

Change how people are paid, to change how they work. There was strong agreement that comp plans built entirely around signed deals and cash collected give consultants no reason to change how they work, even with good tools sitting right in front of them. Nobody had a finished answer on the right model, but the direction (reward AI-assisted, efficient work, not just outcomes) had real support.


What to do next

If you take one thing from this, take this: certainty, not appetite, is the thing most businesses are missing. Everyone wants to move forward. Few have actually decided how.

A good starting point is small and low-risk: write a one-line job description for your single most active AI agent right now. If you can't do it, that's the guardrail you're missing, and it's the cheapest, fastest fix on this whole list.

This piece is drawn from Kyloe's AI Agents Table roundtable, held in London with a group of recruitment leaders talking candidly about where they actually stand with AI. We'll be running this again - if you'd like to be in the room next time, or want a copy of the full checklist covering security, cost, tooling, agents and incentives, get in touch.

 

Get in touch