AI recruiting tools not increasing placements

The conversation goes the same way every time. Sourcing is up, sometimes by a lot, screening is faster, and outreach volume is a multiple of what it was. The dashboard has never looked better. Then I ask how many placements the desk made last quarter compared with the quarter before the tool went in, and there is a pause, and the answer is roughly the same number, sometimes a little lower.

Owners tend to blame the tool, or the market, or the consultants. I think the explanation is simpler and less comfortable. The tools do exactly what they say. They were pointed at a section of the process that does not set the output, and the maths of that has been known since 1961.

Little's law, which no vendor will put on a slide

In 1961 John Little proved a result about queues that applies to any system where things arrive, wait, and leave. The number of items inside the system equals the rate they leave multiplied by the time each one spends inside. Written the way it is usually taught, L = λW. Items in the system, throughput, time in the system. It holds for a bank queue, a factory line, a hospital ward, and a recruitment desk.

Map it onto an agency. The items are candidates in your pipeline. The throughput is the rate at which they leave the pipeline, which for a candidate means an offer, a rejection or a withdrawal. The time is how long each one sits in there. The point that matters is which of those three you actually control.

Throughput is set at the exit, and the exit is a hiring manager deciding. On the desks I have looked at closely, that runs at two or three decisions a week per role and has for as long as anyone has been measuring, because it depends on a diary and a meeting room. Nothing you buy changes it. So λ is fixed by someone who is not your customer's software budget.

Now look at what the AI tools do. AI sourcing raises the number of candidates entering the pipeline. AI screening lets you put more of them through to the next stage. AI outreach gets more of them to reply. All three raise L, and with λ fixed, the equation has exactly one free variable left, and it is W. Every candidate waits longer, and that is the whole effect, and it is arithmetic, and it is why the placement number did not move.

Where the extra waiting goes

A longer W would be harmless if candidates waited politely, and they do not. A candidate in your pipeline is also in three others, and waiting time is the single best predictor I know of that they will accept one of those instead. So the longer W does not just sit there. It leaks, and it leaks the good ones first, because the good ones have the other offers.

This is the part that turns "no gain" into "a loss". You did not just fail to add placements. You put the strongest candidates in a longer queue, and a share of them left while they waited. The consultant then sources replacements, which raises L again. From the outside this looks like a busy, productive desk. From inside the equation it is a queue eating its own best inputs.

Why the tools are all built at the front

Having built in this space for a few years, I can tell you why the vendors point at the front of the pipe. It is where software can count. Sourced, screened, scored, messaged: every one is an event a product can log and put on a chart that goes up. Decisions per week per hiring manager is a number the vendor does not own and cannot improve, so it never appears in a demo.

There is a second reason, and it is the uncomfortable one. Nobody sells a product whose honest instruction is "source fewer people". Yet that is what the equation says. If λ is fixed and W is what you want to bring down, the lever is L, and you bring L down by putting fewer candidates into the pipeline than you could, sized to what the client will decide on this week. Every activity target in the industry points the other way, and so does every tool built to hit one.

The two numbers that tell you whether a tool is working

Ignore activity. Sourced, screened, contacted and replied are measures of L, and L going up is the problem you are trying to fix. There are two numbers worth reading after an AI tool goes in, and both are the kind of thing a desk dashboard should put in front of you without being asked.

The first is placements per consultant per month, which is λ measured where it matters. If it did not move, the tool did not touch the exit, whatever else it did. The second is time from first contact to client decision per candidate, which is W. If that went up after the tool, you have measured the leak. A tool that moves neither is a tool that raised L and nothing else, and you can say that without knowing anything else about it.

I will say one thing about how we handle this in Recruitly, because it is a decision we argued over. The relevancy score on an application is informational, it orders the pile, and it never rejects or advances anyone, because the moment a score is allowed to raise L on its own, the equation above starts running against you without anyone having decided it should.

What to do with the tool you already bought

None of this is an argument for switching the tool off. It is an argument for pointing it at the right variable, and for measuring the thing it was supposed to change instead of the thing it is good at counting.

1. Measure λ, which means for each live role, how many decisions did the client make last week. That number is your throughput and no tool moves it.
2. Cap L to λ, and put no more candidates into a role's pipeline than the client will decide on in a sensible wait, which on most desks is one or two weeks' worth. That means sourcing fewer people than you could, on purpose.
3. Point the tool at W. Scheduling, chasing, confirmations and feedback are the parts of waiting time that are yours rather than the client's. Automate those and W falls without touching L.
4. Read placements per consultant and days-to-decision monthly, and read nothing else about the tool until those move.

The agencies I watch stall after an AI purchase are running the equation backwards, and the tool cannot tell them. The ones that get placements out of the same tools are the ones that read Little's law first, and then bought.

Written for agency owners who were promised more placements and got a busier dashboard.