Tuesday. 09:06. Manolo shows me a screen with the satisfaction of a man who has just cleared out a storage unit.
—Four hundred and twelve candidates, Jose. Reviewed in eight minutes.
—Who have you hired?
—No one yet. But we've rejected three hundred and ninety-eight.
He says it as if they were termites.
—Were they no good?
—They weren't a fit.
—Did you look at them?
—That's exactly why we brought in AI: so we wouldn't have to.
The poster on the wall still says “PEOPLE AT THE CENTRE”. They have put the printer underneath it.
The vacancy is for someone to coordinate orders and deliveries. They need a person who can spot when a shipment is about to go wrong, talk to the warehouse and call the customer before the customer calls shouting. Last month they sent the same order to Murcia twice. The one for Zaragoza is still waiting for theirs.
—What did you ask the filter for?
—To find the ideal candidate.
He opens the job advert. It was written with AI. Manolo added “results-oriented” because he thought it lacked personality.
They want a university degree, SAP experience, advanced English, five years in a similar role and the ability to work under pressure.
—You've got the working under pressure part covered —I tell him—. You supply that yourselves.
Marta from recruitment walks in carrying a folder.
—Ana called. The one the warehouse manager recommended. She's received the rejection email.
—Then she must not be a fit —says Manolo.
—She has spent eleven years coordinating shipments. When we spoke, she asked who confirms address changes here after an order has been closed.
—Why did she ask that?
—She says that's usually where the trouble starts.
Manolo stops smiling. The Zaragoza order changed address on a Thursday.
—Let's see her CV.
Marta puts it on the table. Two pages. Eleven years at a distributor. Three years in customer service before that. Vocational training. A management system that is not SAP.
—There you are —says Manolo—. She doesn't meet the requirements.
—What does she need to know how to do in her first month?
—Organise deliveries. Resolve incidents. Coordinate with the warehouse.
—And do we know whether she can?
—She didn't pass the filter, Jose.
Marta sits down. She recognises the kind of meeting in which one sentence starts going round in circles because nobody wants to open the door for it.
—The advert says “stakeholder management” —she reads—. Ana wrote that she talks to hauliers, salespeople and warehouse managers every day.
—That's the same thing —says Manolo.
—We understand that. Have you checked how the tool scored it?
—We gave it the job description.
—The one the other AI wrote.
—And I reviewed it.
—You added results-oriented —says Marta.
Silence.
He opens the settings. The tool ranks applications by how closely they match the advert. Manolo set the cut-off: anyone below it gets an automatic email. They had to save time. And eighty sounded like a serious number.
—Why eighty?
—You have to draw the line somewhere.
—Of course. What we're trying to find out is what you're drawing the line through.
Marta suggests a test. In a working copy, without sending anything to any candidate, she describes Ana's experience using the language from the advert. She keeps the facts. She does not invent a degree, English or SAP.
They run it again.
Her score rises sharply.
—Now she's getting close to the profile —says Manolo.
—Ana is exactly the same person, Manolo. She hasn't learned anything since nine o'clock. We're the ones who have learned how to write for the filter.
He looks at the two versions. Puts them side by side.
—Then she should have explained herself better.
—Maybe. But we're hiring someone to make sure the lorry reaches Zaragoza. Not someone whose biography your software likes.
The problem is timely. On September 4, WIRED reported that Doist tested automated ranking on applications from hiring rounds that had already finished. According to its head of people, two trials excluded from the shortlist employees who had been hired and were performing well around six months later. These are two cases described by one company, not a general error rate. But they justify a question: would your filter have accepted someone you now know does the job well?
—Are you telling me to get rid of the AI?
—I'm asking how you know it selects well.
—It gives us the best candidates.
—It gives you fourteen. You still have to establish that they're the best.
It is worth separating tools from decisions. An applicant tracking system does not necessarily mean AI rejection. The Greenhouse Talent Matching documentation, for example, says its feature does not automatically reject or advance candidates. At Manolo's company, the email goes out because he decided to connect a score to a rejection.
—We've got four hundred CVs. We can't interview everyone.
—No. And Marta shouldn't spend all day copying phone numbers out of PDFs either. The tool can take that work off her plate. But someone has to define what evidence matters for this job and check whether the tool is finding it.
Marta points to SAP.
—Is it essential to know it on day one, or can we teach the order workflow? We've been using a different system until now anyway.
—It can be learned —Manolo admits—. The hard part is knowing what's going on in the warehouse.
He crosses SAP off as a deal-breaker.
Then they look at the English requirement. Almost all the hauliers and customers for this role operate in Spanish. There is one foreign supplier, and Procurement deals with them.
—I put it in case we grow.
—Then leave it as something we value. Not as a reason to go without someone who can coordinate deliveries today.
They separate what is needed to start from what can be learned. The advert loses its shine. It begins to resemble the actual job.
—And how do we test the rest? —asks Manolo.
—With a situation from the role —says Marta—. An order, an address change and a haulier who has already set off. Fictional data. Twenty minutes. Ask them to explain what they would check and whom they would call first.
—What if they use AI?
—We can allow it and ask them to explain what they accepted, what they corrected and why. If they're going to use tools here, it makes sense to see how. What I want to know is whether they notice the address has changed, not whether they can write “synergies” unaided.
Manolo nods. This time he is looking at the Zaragoza order.
—And we review a sample of the rejected applications —I add—. Using the same job criteria. Without looking at the filter score first. If we find people who deserved an interview, we investigate why they were excluded and adjust the cut-off before we continue.
—That's going to take time.
—Yes. How long has that desk been empty?
We can see it from the meeting room. There is a plant Marta waters and a phone four people take turns answering.
—Three months.
Manolo opens the candidate list again. For the first time, he clicks the rejected tab.
—Call Ana.
—I'll offer her the same exercise as everyone else —says Marta—. The recommendation doesn't give her the job. But the name of the software she used shouldn't take it away either.
She leaves with the folder.
We do not know whether Ana will be right for the role. The conversation might uncover an important gap. There might be someone better. She might no longer want anything to do with a company that rejected her before asking how she works.
We will have to find out.
The only thing we knew this morning was that her CV did not look enough like the text another machine had written.
The phone on the empty desk rings again.
Manolo looks at it. Then he looks at the report: four hundred and twelve candidates processed. Eight minutes. Everything green.
—Who's going to answer that? —he asks.
—One of the fourteen, Manolo. Try calling their PDF.
Thanks for reading.
We all know a Manolo. Sometimes we are the Manolo. A new one every Tuesday.