- Increased usage does not establish a return on investment.
- Total cost includes integration, review, support and maintenance.
- Time savings become economic value according to how they are actually used.
What usage signals reveal, and their limits
OpenAI’s Enterprise Signals, updated on 12 August 2026, describe observed usage among its enterprise customers. They help examine changes within that base. They do not represent every company, and token consumption alone is not a measure of return on investment.
For your organisation, measurement should start with a task and a decision. Do you want to expand a pilot, change a workflow or stop a solution? Relevant indicators are those that support that decision, with visible assumptions and data of known origin.
Define a baseline
Describe how work is performed before the change. Record steps, duration, rework and expected quality. Use several representative cases rather than one unusually easy or difficult example. Specify the period and people involved.
A before-and-after comparison can be affected by experience, volume, seasonality or another process change. Note these differences. Where possible, compare similar tasks performed under similar conditions. A precise-looking table does not fix a poorly defined comparison.
Add the costs actually required
Include licences, model calls, infrastructure, data preparation, integrations and support. Add review time and maintenance. Distinguish initial from recurring costs and specify the volume to which they apply.
Technical consumption may vary with document length, retries or architecture. Measure it on representative cases and build several volume scenarios. An average unit cost should be accompanied by its spread and the situations that increase it.
Connect outcomes to concrete value
Reduced preparation time may enable more dossiers, higher quality or shorter waiting periods. These are different effects. Do not automatically convert every saved minute into financial savings. Ask how the released capacity is actually used.
Also track errors, corrections and approval delays. Faster production can create more rework. Conversely, slightly longer preparation may make a decision more reliable. Criteria should reflect the business objective and its constraints.
Build a decision-ready dashboard
We suggest four groups: usage, quality, operations and value. Define each metric’s formula, source, period and owner. A metric without a stable method can change meaning from month to month.
Show assumptions beside results. If the return depends on uncertain volume, present several scenarios. If quality is assessed using a small sample, say so. Leadership can then distinguish observations from projections.
Set the conditions for expanding a pilot
Before launch, determine what would justify expansion: sufficient quality on selected cases, acceptable cost, sustainable support effort and actual team usage. Also define reasons to stop or revise the approach.
The review should lead to a decision and assigned responsibilities. A useful dashboard shows what changed, what remains uncertain and what work is needed next. To define measurement, bring a process, several recent cases and the data already available in your tools.
| Group | Example metric | Interpretation |
|---|---|---|
| Usage | Tasks actually completed | Workflow adoption, not merely account access |
| Quality | Dossiers accepted without major correction | Outcome evaluated against business criteria |
| Operations | Cost per verified task | Calls, infrastructure, review and support |
| Value | Released capacity actually reused | Documented operational or economic effect |
Sources & methodology
This insight combines cited publications with editorial analysis. Illustrative examples are not client results. Vendor features and terms may change.