Skip to content Skip to footer

AOV: three letters and two ways of running a company

In brief. Salesforce improved Tableau’s answers by making its business definitions explicit. The case shows how much work on metrics comes before a good automated answer. In pre-launch tests accuracy went from 30–45% to over 85% (Salesforce, October 2026).

AOV. For many retailers it means Average Order Value. At Salesforce it means Annual Operating Volume. Give the acronym to an assistant and you can get a flawless calculation on the wrong number.

The trouble is that the answer can look excellent. A clean chart, a comparison with the previous quarter, a credible explanation of the increase. No division errors. It is enough to have picked the wrong meaning of the three letters at the start, and everything else, however well executed, becomes of little use to whoever has to decide.

In the case published on October 12, 2026, Salesforce reports that its analytics assistant had an accuracy of 30–45% without shared business definitions. After adding context, rules and verified answers, with several evaluation cycles on more than 200 real questions, the pre-launch result exceeded 85%. These are measurements of the internal project as stated by the company, not a performance guarantee for any business. Salesforce’s original account, October 12, 2026

That work on definitions interests me because I find it, in a far less spectacular form, in my bootcamps. There are people who spend part of the day moving numbers from one Excel file to another. It is one of the activities I notice most often. You can see straight away how much time could be recovered. What the person knows while copying is less visible: which column must be excluded, which customer is the same one under two names, which row is not yet final.

Part of that knowledge is not in the file. It is in the head of the person who prepares it. If we automate the step before bringing it to the surface, we can transfer the number much faster and lose exactly the information that made it possible to interpret it.

Three active customers

Take a question from a sales meeting: how many active customers do we have? Sales may count open contracts. Finance, those who bought in the period. Marketing, those who interacted recently. They are three potentially sensible measures. The conflict comes when they are presented as one.

A customer with an open contract who has not ordered for months can matter a great deal to sales. For a reactivation campaign it is almost the definition of the right recipient. If instead you need to estimate the warehouse workload, counting them the same way as someone who orders every week can produce a misleading indication.

The definition of a metric is a business decision. It establishes which behaviors we make visible and which we leave out. Asking the head of data to settle this choice alone means handing them a piece of commercial strategy as well, often without telling them.

The so-called semantic model serves to connect data to the meanings approved by the company. Put that way it sounds like a technical object. In practice, someone has to be able to complete a sentence: when we say active customer, for this decision we mean this. They also have to state the period, the exceptions and who can change the definition.

The disagreement worth keeping

There is a good reason to resist the idea of a single definition: different departments have different questions. Making everything uniform can erase useful distinctions. And waiting for the whole organization to agree on every word risks postponing the project with no end date.

The answer is to bound the decision. For the sales budget you need a definition shared by those who approve it. To assess a campaign you may need another, with a name clear enough not to be confused with the first. The assistant has to recognize the difference and, if the question does not make it clear, ask which measure to use.

A system that asks one more question can look less brilliant in a demo. In a meeting it avoids building a discussion on two incompatible meanings. Speed has value when it reduces the time needed to reach a usable decision, not just the time needed to make a number appear.

To start, I would choose a question that comes back every week and that today requires a manual reconciliation. The people who prepare the answers can bring their respective sheets and show where they diverge. The outcome of that meeting should be a definition, a few examples already checked and a person to turn to when an exception arrives.

The definition has to keep its own history too. If the criterion for an active customer changes, the comparison with the previous quarter has to be read knowing which version was applied. Otherwise an increase can depend on the new rule while the chart presents it as a commercial success. The change deserves a date and an explanation.

Then you test the assistant on the same questions in different wordings, including close to a period end. The quarter that has just closed, a customer acquired mid-month, a cancelled order: these are details that bring an ambiguity to light long before a big presentation. This is where the work of those who know the data becomes part of the product.

That colleague who moves numbers between two Excel files, then, should not simply be freed from copy and paste. They should be listened to first. They may be holding a set of instructions nobody has written down yet. The next company assistant starts there too, from the reason why one row is skipped and another is corrected by hand.

Read next Scenari · Decision Intelligence (podcast, in Italian) · One department’s productivity, another’s queue

Image: Salesforce