A finance team can approve ten AI seats and still have no idea what next month’s invoice will be. The seat count is stable. The usage behind it is not. One team summarizes short tickets. Another sends large document sets through a reasoning model. Procurement sees one vendor while the meter sees two businesses.
Freshworks put that mismatch in plain view on October 12 when it announced an agreement to acquire Torii. The planned combination extends Freshservice from software inventory into AI spend governance. Torii tracks AI consumption down to tokens, attributes usage to people and teams, forecasts spend, and supports limits and chargebacks.
That is useful plumbing. It is not yet a product decision.
A token is a meter, not a unit of value
Consumption data answers how much. Product teams still need to answer what for.
A million tokens spent drafting contracts may replace days of review. The same volume spent rewriting internal notes may save minutes. A cost dashboard that treats them as equivalent will identify the larger bill, then quietly miss the larger return.
The governing record needs four fields: owner, workflow, intended outcome, and cost. The first three turn usage into context. Cost makes the tradeoff visible. Without all four, the organization gets a detailed receipt with no reason to keep or cancel the purchase.
This is the same discipline that makes an API billing unit consequential. As the API pricing article explains, the unit on the invoice shapes customer behavior. Internally, the measurement unit shapes management behavior too.
Put the ledger beside the workflow
Traditional software asset management is built around discovery, licenses, renewals, and reclamation. Freshservice already frames that job as a continuous view of software ownership and usage. Variable AI consumption adds a faster clock and a less obvious asset boundary.
The model may be sanctioned while the workflow is wasteful. A cheap model may create expensive rework. An employee may use an approved assistant for data that should never leave a controlled system. None of those cases appears in a license count.
Product instrumentation should therefore attach spend to the work event. Log which feature invoked the model, which model tier ran, whether a human corrected the result, and whether the task reached its target state. Sampled quality review can cover cases where direct outcome measurement is impractical.
Now the ledger can distinguish growth from leakage.
Every threshold needs an action
Visibility products often stop at the alert. A red chart reaches an operations channel, several people acknowledge it, and the invoice continues its brisk little walk uphill.
Each threshold should trigger a named product response. A sudden token spike might switch a workflow to a cheaper model, reduce retrieved context, require approval, or pause the feature. A sustained rise paired with better resolution time may justify a larger budget instead. The action depends on value, not color.
Set those rules before the alert fires. Assign an owner who can change the product, not merely explain the bill. Keep an exception path for high-value work. Cost controls that ignore urgency eventually become buttons users learn to route around.
Govern the decision, not only the tool
Freshworks describes a future system of record across applications, licenses, identities, and AI spend. The stronger product opportunity is a system of decisions: what work deserves premium inference, where a limit should bend, and when increased consumption proves adoption rather than waste.
AI spend governance will earn its place when teams can trace a dollar from model call to workflow result and then act. Tokens are excellent meter readings. They are terrible strategy.
Comments
Comments are powered by GitHub Discussions through Giscus.