Salesforce reported $800 million of Agentforce ARR for the fiscal year ended 31 January 2026, announced on 25 February. Its results for the quarter ended 31 July 2026 reported over $1.5 billion and disclosed that the definition now included Slackbot and Headless 360. These are company-reported figures, reviewed on 22 October 2026.
The disclosure identifies a scope change. It does not quantify how much of the increase came from that change, so the two headlines alone cannot establish growth for a constant set of products. Ask for a reconciliation before drawing that conclusion.
Apply the same check to an internal dashboard whenever a team adds a product, channel or event to a metric.
How definitions drift in AI products
AI metrics are unusually easy to redefine, because the underlying events are fuzzy. Is a conversation resolved when the customer stops replying, or only when they do not come back within a week? Is a code suggestion accepted when the developer presses tab, or only when the code survives review? Does a generated summary count as used when it is displayed, opened or copied?
Each of those choices can move a headline number by tens of percentage points. And each tends to change for an innocent reason. A new data source arrives and the old event is deprecated. A team adds a channel and starts counting it. Someone fixes a logging bug and the fix changes the count. Six months later the trend line on the leadership dashboard joins two different measurements. Nobody remembers which month the second one began.
Three habits that keep a metric honest
Write the definition down where people will see it. Not in an analytics wiki that nobody opens, but next to the number on the dashboard: what counts, what does not, the time window, and the date the definition took effect. For an AI feature, include which model and prompt version the number was measured on, because a provider update can change behaviour without a single line of your code changing.
When the definition changes, restate the history. If you start counting re-contact on email as a failed resolution, recompute the last few quarters the new way so the trend compares like with like. Show both definitions side by side for at least one period, so people can see how much of the movement came from the change.
Keep a short change log. Date, old definition, new definition, reason, approver. Bureaucratic? Only until a board member asks why a number jumped and the answer takes thirty seconds instead of a week of archaeology through dashboards, notebooks and old Slack threads.
Reading other people's numbers
The same discipline applies to numbers you cite. AI markets generate a lot of headline figures: revenue run-rates, resolution rates, accuracy claims, adoption counts. Before using one in a strategy document or an interview answer, ask what exactly was counted, over what period, and by whom. Vendor resolution rates, for example, are often measured on a vendor's own definition of resolution, which may not match the one your operation would use.
Forecasts need a different label from observed results. A projection of future project cancellations does not measure cancellations that have already happened. Before using such a number, find the forecast date, target period and method; if those are unavailable, leave it out of the decision.
In the interview
Metric questions in AI PM loops reward candidates who define before they measure. When asked how you would measure a writing assistant, a support agent or a coding tool, the strong answer opens with the definition of success for one event, then names the edge cases that definition has to handle, then picks the primary metric and its guardrails.
For the practice prompt "Your metric jumped 30% this month; what do you check?", start with changes to the definition, logging, traffic mix and model version. Then test whether the product improved within comparable cohorts.
That reflex, checking the ruler before reading the length, is most of what metric literacy means in this job.
Questions about this topic
Salesforce reported $800 million of Agentforce ARR for the fiscal year ended 31 January 2026 and over $1.5 billion for the quarter ended 31 July 2026. The later definition includes Slackbot and Headless 360. These are company-reported figures with different scopes.