What the wider data says happens next

What the wider data says happens next

You don't have to weigh the cost of building and running an AI recruitment agent on faith. Plenty of agencies have already run this experiment for you, and the results aren't kind.

Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, and finds that at least half of GenAI projects run over budget. MIT's NANDA study found that initiatives built with a delivery partner succeed about 67% of the time, against roughly 33% for the ones built and maintained in-house: a two-to-one gap that points squarely at who runs it, not who can code it.

And the deepest source isn't a vendor survey at all. A decade ago, Google engineers laid out in a peer-reviewed paper (Sculley et al., Hidden Technical Debt in Machine Learning Systems, NeurIPS 2015) that the model code is the small part of a real ML system, and the ongoing maintenance is the massive, recurring cost. That was true before agents existed. Agents, with their moving APIs and deprecating models, only made it truer.

Two-thirds of in-house builds don't make it. The technology isn't what fails them.

SLAs, and who holds the leash

There's one more thing money buys that a solo build can't, and it's the reason any of this matters to a client.

When something goes wrong with a managed service, there's a service-level agreement behind it: a defined response time, monitoring that runs whether or not you're awake, a named team whose job is to fix the problem before you see it. When something goes wrong with a DIY build at 2am, the SLA is "hopefully the one engineer is awake, and hasn't quit." Those are not the same promise. And your client can tell the difference the first time a CV with a date of birth on it lands in their inbox and nobody caught it.

But here is the line we've held from the very start: managed does not mean handing over the decisions. The leash stays with you. The agent does the tireless 90%; you own who moves forward, which placement closes, what goes to the client. Managed automation moves the operational burden off your desk: the pager, the dashboards, the API drift, the on-call rota you were never going to staff. It does not move the accountability for hiring decisions, and it shouldn't. Automate the work. Don't automate the accountability.

Hand off the running. Keep the leash. They're different things, and conflating them is how good tools become liabilities.

The decision, stated plainly

Lay the three doors down one last time.

Buy gives you something generic that scores your candidates the way it scores everyone's: wrong for you, by design. Build and run it yourself gives you exactly the tool you want and a recurring bill dominated by one fully-loaded engineer at 4–6× a managed retainer, plus an on-call problem a single person genuinely cannot solve, which is precisely the gap two-thirds of in-house builds fall into. Managed gives you the tailored tool and the team that keeps it alive, at the midpoint of a market range that sits well below what one engineer costs you loaded.

We're not going to tell you which door to walk through. The numbers already did. The only honest question left is the one this whole exercise was meant to answer: not can this be built (you've seen that it can) but who should be the one keeping it alive.

More placements. Same team. No new tools. And now you know who runs them.

Next: exactly how this gets designed, deployed and run for you, and what "live in 30 days" looks like in practice.