Reformation reported third-quarter net revenue of US$173.6 million this week, up about 19% on the year before. On the earnings call, CEO Hali Borenstein said the apparel brand had implemented a new AI-backed planning tool that improves how it evaluates product performance and forecasts demand, and that it bolsters in-stock rates and full-price selling. The company began trading as a public company in July, and William Blair analysts said AI-enabled merchandising should further help it localise assortments and chase demand as it scales.

What is actually being claimed

Strip out the earnings-call language and the claim is small and specific. A planning tool helps the merchandising team decide what to stock and how much. The result shows up in two numbers the company tracked long before the tool existed: how often a product is in stock, and how much of it sells at full price.

That is the part I like. Nobody said the tool changes the business. Nobody called it an agent. The number came first and the tool came second.

What the article does not give is the size of the move. There is no before-and-after in-stock rate and no percentage. So I read this as a direction, not a proof. Reformation says the number moved, and the next few quarters will show by how much.

The analysts’ comment matters too. Reformation already read and reacted to demand faster than its peers, with weekly drops. The AI was added to a process that already worked and already had a scoreboard, so it sped up something that was already being measured.

The test: can you say it as a move in a number?

One of my ten AI principles is “Two readers, and measure outcomes.” The second half is the one most AI projects skip.

Here is the test I apply to anything AI-related in my own businesses. Before it is built, can you state the project in a sentence of this shape: “We expect X to move from A to B”, where X is a number you already look at every week?

If you can, you have a deployment, and it can go in the budget. If you can’t, you have a demo. Demos are useful, because they are how you learn what a tool can do. But a demo doesn’t belong on an earnings call, and it shouldn’t be counted as something that pays for itself.

Notice what the test rules out. “We’re exploring AI” fails. “We’ve rolled out an AI assistant across the team” fails, because adoption was never a number the business watched. “Time saved” usually fails too, because nobody tracked the hours before and nobody will check them after.

Reformation’s version passes. In-stock rate existed before the tool. Full-price selling existed before the tool. The tool either moves them or it doesn’t, and someone outside the project can check.

What this looks like for a smaller operator

You don’t need to be a public company for this. Take a shop with one person doing the buying.

The numbers already exist. How often did a customer ask for something that was out of stock? How much of last season went to markdown? Both sit in the order system and the clearance report. Pick one, write down where it stands today, then put the AI on that one decision.

A clinic has no-shows, or the share of booking requests answered the same day. A service business has quotes sent within a day. A marketing team has enquiries that arrive with the brand already named, or the share of commercial queries where the brand appears in the AI answer.

At AeroChat, where most of my time goes, the number a merchant watches is how many conversations get handled with no person in them. For one bakery merchant it handles 87 to 94% of conversations with no person, through festive peaks. That figure belongs to that merchant alone, and it is what tells the owner whether the thing works. The demo never did.

Why most AI projects fail this test

The tool gets chosen first. Someone sees a capability, buys it, then looks for somewhere to use it. The number arrives at the end as a justification, and by then nobody wrote down the starting point, so the before-and-after can’t be done.

The other failure is measuring the AI instead of the business. Accuracy scores, answer quality and hours saved are all inputs. The business number is the output. What Reformation chose to report was in-stock rate and full-price selling, and those are outputs.

What to do

  1. List the five numbers you already look at every week. Only ones that are already on a report.
  2. Write each AI idea on the table as “X from A to B”. If it won’t fit that sentence, park it as a demo.
  3. Record the starting value and the date before you switch anything on. AI output changes week to week, so the date is part of the result.
  4. Put the AI on one decision, the way Reformation put it on what to stock. Leave the rest of the department as it is.
  5. Check the number at a fixed interval and be ready to say it didn’t move. The one that moves gets more. The one that doesn’t gets switched off.

Source: Retail Dive: Reformation implements new AI planning tool, helping in-stock rates