Whose number moves

A specialist building products distributor. Around £180m of revenue, six hundred staff, family owned with a minority institutional stake. Two AI pilots inside fourteen months. Same company, same appetite, same head of data running both, and in the second case the same vendor.

One of them is still causing problems. The other one is in the budget.

Part one. The pilot nobody stopped

The first one started the way most of them start, which is to say it started well.

Steve, the operations director, had sat through a demo of a document assistant that could answer technical product questions from the manufacturer spec sheets. The branches field those questions constantly. Adhesive coverage rates, substrate compatibility, curing times. The technical helpdesk was two people and a phone queue, and everyone in the business had a story about waiting on hold to find out whether a product could go on a damp screed.

Steve was not doing anything cynical. He watched a demo, recognised a genuine irritation, and moved. He got £120,000 from the transformation budget, which existed and had money in it, and he asked Marcus, the head of data, to run it.

Nobody asked which line on the plan would move. If they had, the honest answer was that the technical helpdesk did not appear on the plan. It was a cost of about £70,000 a year sitting inside a branch operations line that nobody reviewed, and no one had ever set a target against it.

Nobody asked who had been trying to fix it before. The answer was nobody, in the sense that mattered. Everyone complained about the helpdesk. Complaining is not the same as spending. There was no contractor, no overtime, no spreadsheet somebody maintained at home. The business had already priced the problem and the price was zero.

The build took eleven weeks and it worked. That is the part people find hardest to believe. It ingested four hundred spec sheets, it answered questions correctly most of the time, and the two branch managers who trialled it liked it.

Then the pilot period ended.

There was no decision. There was no meeting where anybody said this has earned its place, here is the budget line, here is who owns it now. There was also no meeting where anybody killed it. Steve had moved on to a depot consolidation by then, which was on the plan and did have a number against it, so his attention went where his objectives were. That was rational.

The two branch managers kept using it, because it was genuinely useful to them. Somebody put a link to it on the intranet. Within six months a third of the branch network was using it to answer customer questions, and nobody had approved that, because nobody had approved anything.

Eleven months after the demo, a customer was quoted a coverage rate for an adhesive that had been reformulated. The spec sheet in the system was the old one. The installation failed, the customer claimed, and it cost about £40,000 to settle plus a relationship that had taken nine years to build.

Marcus found out from the warranty claims report.

The pilot eventually did land on a number that somebody owned. It landed on warranty claims, in the wrong direction, eleven months late, and the first person to see it was not on the programme.

What the six would have caught

Run backwards, the pilot fails at five of them and it fails at the first one.

• Which line moves, and by how much. No line. The helpdesk cost sat inside an unmanaged operations line with no target. That alone was a stop.
• Who was already trying. Nobody. Complaint without spend. The problem was real as an irritation and unfunded as a problem, and the £120,000 figure was built backwards from the demo.
• Value or learning. Nobody asked. If asked, the answer would have been "both", which means neither.
• Whose year is worse. Steve's name was on the funding and no name was on the outcome. When his objectives moved, the pilot had nothing holding it.
• Who signs it into production. Nobody, which is why it went into production anyway. This is the question that catches the pilot that works, and this pilot worked.
• Whose budget picks it up. The transformation line, which has no year two. Everybody knew that except the people running it.

Any one of those, asked out loud in the first fortnight, ends this differently. Not one of them needed a consultant, a workshop or a single line of code.

Part two. The pilot that shipped

Fourteen months later Marcus had a second proposal on his desk, and this time he asked the six before he asked for anything else.

The proposal was credit control. Chase overdue invoices, prioritise the queue, draft the follow-ups.

Which line on this year's plan moves if this works, and by how much? Days sales outstanding. Fifty-eight days against a target of fifty-four. The target already existed, it was already in Jonathan's objectives as finance director, and it had been missed two years running. Marcus did not invent the figure. He got it from Jonathan's objectives, which is where figures should come from.

Who was already trying to move that line, and what were they using? Three people in credit control, a shared inbox, and a spreadsheet one of them rebuilt every Monday morning. One of them had been doing it on Sunday evenings for two years. That was the answer that made Marcus confident. He was not proposing a capability. He was offering to make a fight somebody was already losing cheaper.

Value or learning? Value, unambiguously, and it went in the funding paper in those words before the money moved.

If this doesn't work, whose year is measurably worse? This is where it nearly came apart.

Marcus went to Jonathan and did not ask him to sponsor anything. He asked which of Jonathan's numbers this should move and whether he would want it if it did.

Jonathan said days sales outstanding, obviously, and then said something more useful. The reason DSO was stuck was not chasing speed. It was that a third of the overdue balance sat behind disputed delivery notes, and no amount of faster chasing moves a disputed invoice. If the thing could read the proof of delivery images and match them to the disputed lines, that would move his number. The chasing would not.

The pilot changed shape in the first conversation, before anybody built anything, because the person whose year it was said the original version would not have worked. He was right.

It would have been a well-run pilot pointed at the wrong half of the problem, and it would have produced a technically successful result that moved nothing.

Same name on question one and question four. Jonathan.

Who signs this into production, and what do they need to see to sign? Claire, the CFO. When Marcus first asked what she needed to see, she said she would know it when she saw it, which is not a bar. So he asked the other half. What would you need to see to shut this down?

That she could answer immediately. If it has not held four days for two consecutive months by the end of March, stop it, because by then we will be into the year end and I am not carrying an unproven process through audit.

Both endings, in a diary, before the build.

When the pilot money runs out, whose budget picks it up? The proposal had it coming from the transformation line again, which is how the first one ended. Marcus took it out and asked Jonathan directly. Jonathan agreed to carry the run cost in the finance function from April, on the condition that the March bar was met, and he agreed it in writing before the build started rather than after.

That is the answer that only happens when questions one and four are genuinely true. Nobody agrees to carry a cost for something that does not move their own number.

The build took nine weeks. It was less technically impressive than the first pilot.

DSO came down to fifty-four days by the end of February and held through March. On £180m of revenue that is a little under two million pounds of working capital that stopped sitting in other people's bank accounts. Claire signed it in on the second of April. The run cost sits in the finance function budget where Jonathan put it.

Nobody wrote it up. There was no case study and no internal announcement, because from the outside it looked like the finance director hitting a target he had been chasing for two years, which is exactly what it was.

What was actually different

Same company. Same head of data. Same board, same year, same risk appetite, same vendor on the second one as the first. The second pilot was less technically ambitious than the first and the first one worked better on the day it was demonstrated.

The difference was six questions, asked out loud, in order, before the money moved. They took one afternoon and two conversations, one of which was uncomfortable for about ninety seconds.

And notice what the questions actually did. They did not judge the technology. They did not produce a score, a maturity rating or a governance framework. They changed the shape of the pilot in the first week, on the word of the one person whose year depended on it, and they put a name against both endings before anybody could get attached to the thing.

The first pilot is still running, incidentally. Nobody has ever ended it.

If the first half of that is recognisable

The questions above are free and you can ask them yourself. What makes them hard is not asking. It is that Finance calls it a cost problem, Operations calls it a process problem and the AI team calls it a data problem, everybody is partly right, and the meeting ends with an action to align offline.

Three accounts of one problem do not resolve by meeting again. What resolves them is somebody who can hold all three in the room without belonging to any of them, and get to the one sentence all three will sign. That is most of what I am brought in to do, and section 07 sets out the other two cases where it applies.

You do not need to finish this paper first. Send me a few lines about your pilot and the point at which it stopped moving. One conversation, no charge.

troy@serteam.co.uk

The legitimate exit from a blank

A blank does not have to mean stop. It means stop pretending. Go to question three and answer it honestly. If this is learning, say so in the funding paper, name the person who is measured on capability this year, and set the bar as capability rather than impact. That is a real purchase and it survives a review, because when somebody asks what impact it made you have the receipt showing that impact was never what was bought.

What does not survive a review is a blank dressed as a value case.

Sometimes the value is real and it is genuinely not isolable. The work makes twenty people's days better in ways that show up nowhere, or it removes a risk that was never quantified, or it improves something that moves for nine other reasons at the same time. Demanding a line and a figure for that is not rigour. It is a way of never doing anything that is not already counted. This is where a lot of mid-market companies actually are.

Two things are true at once. First, a smaller number you can defend beats a larger one you have to argue for every quarter. If the full benefit is eight things and only one of them is checkable, build the case on the one and let the other seven be upside you mention and do not claim. Attribution arguments are lost slowly, by attrition, in years two and three.

Second, the isolability problem is real but it is much rarer than it is claimed. Most of the time "you can't really measure it" means nobody has tried for four weeks. The genuine cases have a particular shape: the benefit is distributed across many people, each individual effect is small, and the aggregate is real. The false cases have a different shape: one team, one process, and a reluctance to write the current state down.

The test is cheap. Ask what it would take to measure it. If the answer is a spreadsheet and a month, it is not hard, it is uncomfortable.