Paid, enabled, used: the three statuses of an AI

Ask three people in your company whether “the AI in the office suite” is in place. The IT manager will say yes: he bought the licenses. The department head will say no: nobody on his team uses it. The administrator will say the feature has been enabled since the September update. All three are right, because they are not talking about the same status.

Three statuses, never merged #

The method for the AI register, public and under a CC BY 4.0 license, defines three statuses for each AI feature.

• Paid: at least one seat or subscription that includes it is billed to the company.
• Enabled: the feature is available to at least one user, through an administrator setting or by the vendor's default.
• Used: at least one user has used it over the last twenty-eight days, according to the vendor's aggregated usage reports.

The rule that matters fits in one sentence: the three statuses are independent and are never merged. We tend to think that a paid feature is necessarily enabled, and that an enabled feature is necessarily used. The gaps between the three are precisely what needs to be seen.

The three gaps #

Paid without being enabled. You are paying for a closed door. These are the seats bought for a pilot and never deployed.

Enabled without being used. The door is open, nobody comes in. These are the features switched on by a vendor update that nobody announced to the teams.

Used without having been decided. People come in through the window. These are the tools your employees use every day and whose existence management discovers on an invoice.

Each gap calls for a different decision. But to decide, you first need to know which one you are in, feature by feature.

One line per feature, read in the tools #

Two rules make the statuses readable.

One line per feature, not per vendor. A feature is a unit named by the vendor, attached to a product and an edition. “Our office suite” is not a line. “Assisted writing in the mailbox, Enterprise edition” is one.

The status is read in the tools, never in the content. Paid: the invoice. Enabled: the setting in the admin center. Used: the vendor's usage report, per tool and over twenty-eight days, never per person, and never for a group of fewer than ten people.

What follows from the statuses: four measures and a total #

The total is the first figure to look at: what AI costs you per year, across all vendors. In most companies, nobody has ever seen it, because it is spread across ten invoices and three budgets.

A fictional example to make it concrete. A company with three hundred accounts pays for 120 seats of an AI assistant, 41 of which were used over twenty-eight days: 79 dormant seats. Its marketing department pays for twelve subscriptions to a general-purpose assistant that covers the same use: a duplicate. Its recruitment tool has been sorting applications since an update nobody read: a feature enabled without a decision, which touches candidate data. None of these figures is a verdict. Each one puts a line in front of the person who has to decide.

Where to start on Monday #

1. Open a spreadsheet. One line per feature, one column per status, the date at the top.
2. Fill in ten lines: five from the licenses, three from the invoices, two from the list of applications authorized in your directory.
3. Set the three statuses separately, each from its own source.
4. Calculate the total spend, even roughly.
5. Have the page signed by the AI lead.

You will have a register that is wrong and incomplete, and that is fine. It is dated, it has an author, and the next version will correct it. That is the difference between a register and an inventory.