25 Trillion or 40 Trillion Tokens? The Adoption Claim Needs Clarification

An InfoQ presentation's title refers to 25 trillion tokens, while its summary says Brendan O'Leary draws from 40 trillion tokens of usage data. The supplied description does not explain how those figures relate ¹. That discrepancy does not establish which figure, if either, is incorrect; it does mean readers should ask for clarification before citing a single total.

A token total alone does not show who generated the tokens, what tasks they were doing, or whether those tasks succeeded. The supplied description gives no such breakdown for either figure ¹. For a platform leader weighing whether to cite this AI agent adoption data, that is a reason to seek more detail, not to infer adoption or outcomes from volume alone.

What the presentation says about enterprise AI adoption

The summary says real-world AI adoption depends on building trust rather than chasing model benchmarks. It also says the talk explains a four-rung AI adoption ladder and analyzes failure points in context construction, multi-model routing, and feedback loops ¹.

That framing can be useful as a prompt for discussion: trust and benchmark performance are distinct considerations. But the supplied description does not define the four rungs, explain how a team or workflow would be assigned to one, or describe the evidence behind the trust claim ¹. Without those details, other organizations cannot tell whether they are applying the same framework or using the same terms differently.

Three areas to investigate: context construction, multi-model routing, feedback loops

The summary identifies context construction, multi-model routing, and feedback loops as key failure points ¹. Treat them as questions for your own instrumentation, not as measured findings:

• Context construction: What information reaches the agent, and where might relevant context be lost?
• Multi-model routing: How are requests assigned across models, and does the selected route correlate with good or bad outcomes?
• Feedback loops: Is user feedback captured, and does it lead to changes in the agent or its workflow?

The supplied description does not provide event definitions, denominators, or a method for identifying these failure points ¹. Token volume alone is consistent with different patterns of use; the description does not establish which pattern the data reflects ¹.

If you are scoping a pilot around these questions, the pre-pilot checklist for evaluating Amp's AI agents beyond coding covers a related set of evaluation questions.

What to ask before citing the presentation

Before using the figures or framework in a strategy document, ask the presenter for:

1. Dataset scope: Which users, organizations, and products are included, and what are the inclusion rules? The supplied description does not specify this ¹.
2. Token accounting: How were tokens counted, over what period, and how do the 25-trillion and 40-trillion figures relate? The supplied description does not say ¹.
3. Rung definitions: What are the four adoption stages, and what evidence moves a workflow from one to the next? The supplied description does not define them ¹.
4. Method and denominators: What population underlies each reported failure pattern? The supplied description does not provide this information ¹.

Until those details are available, avoid presenting the four-rung framework as a defined measurement system. If your team creates its own stages, label them as an internal framework.

Build your own measurement of developer trust in AI agents

The three areas in the summary can serve as a starting checklist. For each, define an observable event and connect it to a business outcome before collecting data:

• Context construction: Did the agent receive the intended context, and did the task complete as intended?
• Routing: Does the selected model correlate with task success, rather than only with a completed call? Comparing routes across a model catalog is easier when each request records which model handled it.
• Feedback: Is feedback captured, and does it lead to a change in the agent or workflow?

The summary puts trust at the center of adoption but does not explain how trust was measured ¹. Define what evidence would show that a team is willing to let an agent perform a specified task unsupervised, then keep that definition consistent when comparing results over time. For a comparison of how demonstrations differ from reliability evidence, see why a long-running agent demo is not a reliability case.

One practical option is to route traffic through an enterprise AI gateway, so that model choice and outcomes can be logged in one place. That is a design choice for your team, not something the presentation describes.

Next step: Ask the presenter for the dataset scope, token-counting method, and rung definitions before citing the presentation internally. In parallel, consider measuring context, routing, feedback, and trust against workflow outcomes in your own environment. That would give your team evidence tailored to its own use cases rather than relying on an unexplained token total.