How much one initiative should cover

An initiative is one outcome, with one owner, one baseline and one decision ahead of it. If two parts of the work would be judged by different numbers, owned by different people, or kept or cut separately, they're two initiatives.

Split where a decision would split

If you could see yourself expanding the support team's use and stopping finance's, measure them apart. One initiative can only be kept or cut whole, so a decision that would treat two parts differently needs two numbers.

Don't split below what has its own number

Two support teams answering the same tickets, measured by the same handle time, are one initiative. Splitting them gives two numbers that move together and a baseline cut in half.

A company-wide assistant: measure each use

A license bought for everyone, such as Microsoft 365 Copilot, Claude Enterprise, Gemini, ChatGPT Enterprise, Glean or Amazon Quick, is used by many teams for many things. Measure each use as its own initiative, such as the support team researching tickets or finance drafting variance commentary. The templates that start Assistant for each time one task, so each use gets its own baseline and value.

Split the license's bill across those initiatives by seats, with an allocation rule on the assistant's vendor. A rule splits bills and charges from a connected QuickBooks, Xero, NetSuite or American Express account. Open Finance → Allocations → Rules, choose New Rule, pick Fixed shares, and give each initiative the seats its team holds divided by the seats licensed: 120 of 800 seats is 15%. Every split needs a cost type too; a license is software. Count a person once, under the use they make most of. When seats change, revise the rule. See Allocation rules and accounting periods.

The shares don't have to reach 100%. What no initiative takes books to no initiative. That part of the license may well be returning value; nothing measures it. At renewal, it's the figure to ask about. For a vendor AI Spend counts as AI, such as OpenAI for ChatGPT Enterprise, or Glean, it shows as Unbooked on Finance → AI Spend. A Microsoft or Google invoice isn't counted there, because it carries other licenses too; an Owner or Admin can mark the vendor as AI, but then the whole invoice counts.

Two cases need care:

A bill that carries other licenses too, such as Gemini inside a Google Workspace invoice, or Copilot on a Microsoft invoice with your other licenses. Take each share of the assistant's part only: if the assistant is a fifth of the bill, 120 of 800 seats is 15% of a fifth, or 3%.
Amazon Quick, on your AWS bill. Link the AWS connection to each initiative as a cost source and map the service Quick is billed under, named as it is on your bill. Where the connection counts Quick's seats, which takes three list-only permissions in AWS (the AWS setup guide gives the command), pick the Quick groups each initiative covers beside its mapping. Its share then follows those groups' seats each month, at Quick's list prices, and a person in groups of two initiatives is split evenly between them. Seats in no group, Quick's account fee and the rest of the line book to no initiative. Without groups, type each mapping's share instead: 120 of 800 seats is 15%. Roiva won't save shares that add up to more than 100%. See Linking connections to initiatives.
An assistant billed somewhere Roiva doesn't read, such as a card your books don't carry. Record each initiative's share of it as a recurring cost instead. See Vendors, shares and recurring costs.

Each initiative counts toward the free account's measured initiatives the way any initiative does (paid plans measure all of them). A share of the license doesn't make an idea count; value recorded on it, or moving it into Implementing or Live, does.

Or measure it as one

The AI Assistant Rollout template measures the whole assistant as one initiative: it values the minutes saved on the tasks you timed, for each person who used the assistant that month. A Microsoft 365 Copilot connection counts those people from Microsoft's usage report, a ChatGPT Enterprise or Claude Enterprise connection from OpenAI's or Anthropic's analytics, a Gemini in Google Workspace connection from Google's Gemini audit log, and a Glean connection from Glean's Insights; for another assistant, you record them by hand. That's a fair way to start. What you give up is knowing which teams' use pays for the licenses. See Measuring a workplace AI assistant.

Examples

Too broad	About right	Too narrow
"GenAI program"	"Copilot for support macro drafting"	"Copilot for tier-1 macros, EMEA, weekdays"
"Amazon Quick"	"Quick for month-end variance commentary"	One initiative per finance analyst
"AI in Sales"	"AI call summaries for account executives"	One per sales region, all with the same metric