The AI Agent Cost Problem Nobody's Talking About

I had a conversation on the Reboot podcast this week that I think every recruitment agency owner needs to hear.

Kat and I were talking about the rising costs associated with AI agents, and halfway through I realised something: most people buying these tools have absolutely no idea how the pricing actually works. They're signing up for things they fundamentally don't understand.

And that's not their fault. The vendors aren't exactly rushing to explain it.

So let me break it down properly.

How AI Agents Actually Work (The Simple Version)

Forget everything you've read on LinkedIn about "autonomous AI" for a second.

You set a timer. At 9am, the agent wakes up. It gets told "go check Nick's LinkedIn messages." It picks the right tool (a LinkedIn connector), grabs the data, processes it through an LLM like GPT or Claude, and gives you an output.

Same as a human recruiter who logs in at 9, checks messages, and updates their CRM. Except the agent doesn't get distracted, doesn't call in sick, and doesn't spend 20 minutes making a coffee first.

The difference? A human costs the same whether they process 5 messages or 500. An agent does not.

The Bit They Don't Put in the Brochure

LLMs charge based on tokens. Tokens are roughly equivalent to words. You pay for input tokens (what the agent reads) and output tokens (what it generates).

Here's where it gets ugly.

Let's say your agent checks LinkedIn messages every morning. Monday, 4 pages of messages come back. The LLM processes all 4 pages. Costs a few pence. No drama.

Tuesday morning. The agent needs context from Monday to understand what's new. So it re-reads Monday's 4 pages, plus today's 3 new pages. That's 7 pages through the LLM.

Wednesday. 10 pages. Thursday. 13. Friday. 16.

Day	Pages Processed	Cumulative Context
Monday	4	4 pages
Tuesday	7	Mon + Tue
Wednesday	10	Mon + Tue + Wed
Thursday	13	Mon – Thu
Friday	16	Full week context

One week. One tool. One simple task. And the token cost has quadrupled from Monday to Friday.

Now here's the question nobody seems to be asking: what happens after a month?

And that's just LinkedIn messages. What about when your agent is connected to Apollo, your CRM, job boards, email, and an outreach platform? Each one pulling data. Each one feeding it all to the LLM. Every day. Compounding.

There's a new tool called OpenClaw that's getting a lot of attention right now. And I'll be honest, it's cool. Genuinely cool. You set timers, connect all your tools, give an LLM access to everything and let it run autonomously. It uses the exact principles I just described — alarm clock, tools, LLM processing.

I looked at it properly. Played around with it. You do need to be a bit technical to set it all up, but the concept is solid.

And I decided not to use it.

Not because it's bad. Because the underlying cost structure is the same problem I've been talking about. The LLM processes everything. Every time. Every tool. Every day. At scale, with multiple recruiters connected to multiple tools, your token bill will be eye-watering within a few months. And most people getting excited about it right now haven't done that maths yet.

The agencies I'm speaking to who are buzzing about this stuff haven't asked the cost question yet. When they do, the excitement tends to cool off rapidly.

The Fix (And Why Most Current Tools Don't Have It)

The way to solve this is embarrassingly simple in concept: stop feeding the LLM everything.

Instead of sending 16 pages of LinkedIn messages directly to the LLM, you use a database layer. The database stores all the raw data. It processes it. It extracts only what's relevant. Then it sends half a page of specific, targeted information to the LLM.

Think of it like training a new recruiter. You wouldn't hand them 5 textbooks on day one and say "memorise all of this." You'd give them one chapter at a time. Less leakage. Better retention. Better results.

The LLM is the same. Give it less, get more.

The problem? Most AI agent platforms on the market don't have this database layer. They're designed to throw everything at the LLM because, well, LLM providers charge per token. More tokens = more revenue for someone in that chain. I'll let you figure out who benefits from that setup.

What About Your CRM?

I get asked this a lot. "Can't my existing CRM be the database layer?"

Honestly? Not really. Not yet, anyway.

Most recruitment CRMs weren't built for this. Their databases aren't structured to feed specific, small packets of data to an LLM. They'll send you 2 pages of candidate records when you only need 3 lines.

Atlas is probably the closest thing I've seen to getting it right. But the traditional players — their data structures will actually cost you more in LLM fees because they over-deliver data that the agent then has to process.

The GDPR Angle Nobody Wants to Discuss

Here's the one that should really worry you.

If your AI agent is processing all your candidate data, all your client data, all your communications through an LLM... where is that data going?

Can your vendor tell you specifically where that data is processed? Who has access? How it's stored? How long it's retained?

If the answer is "um, well, it goes to OpenAI's servers..." then you've got a compliance problem. Because you've just sent personal data to a third party that's processing it in ways you can't fully audit or control.

What I'd Actually Recommend

Whether you work with me or not, here's what I'd tell any recruitment agency looking at AI agents right now:

1. Ask about token costs BEFORE you sign. Not monthly fees. Actual token consumption projections. If they can't give you a straight answer, walk.
2. Ask about context accumulation. Does your agent need more and more data over time? If yes, your costs will compound. Ask how they manage this.
3. Ask about the database layer. Is there one? Where does raw data get stored? What goes to the LLM? If the answer is "everything goes to the LLM" — be very, very careful.
4. Ask to see real-time cost tracking. If you can't see exactly what you're spending on tokens in real-time, you're flying blind. And flying blind with AI costs is like giving your trainee an unlimited company card.
5. Ask about GDPR. Where is candidate data processed? By which third parties? Can you audit it?

These are not unreasonable questions. Any vendor worth working with should be able to answer them clearly and confidently. If they can't, or if they deflect, that tells you everything you need to know.

It's exactly why I built the Recruitment Agent Hub with real-time cost tracking and a database-first architecture. Not because it's fancy. Because recruiters deserve to know what they're paying for.