Accenture’s AI tokenomics report gives health systems five rules for controlling token costs

Key takeaways

Token use is growing faster than prices are falling, and only about 20 cents of every token dollar can be tied to a financial outcome. The report’s fix is five disciplines: log every workload at a gateway, bill the teams that use it, route each task to the cheapest capable model, require a value case before production and set the right default by policy. For health systems, the caveats matter as much as the rules, since clinical work that needs a signature sits outside Accenture’s routing math and most of the $5.6 million headline payoff is cost growth avoided, not new money.

Most technology costs are known before the purchase. AI costs are not. Every model interaction burns tokens, and a single request to an agent can consume millions of them before anyone sees a bill.

In health systems, that meter is already running. An AJMC study found 62.6 percent of hospitals on Epic had adopted ambient AI documentation as of June 2025, and an Eliciting Insights survey of 120 health systems put adoption of clinical note-taking and ambient listening at 68 percent.

Boards want to know what is driving AI costs, how fast they will grow, and what the organization gets for the money. With hospitals’ year-to-date operating margin index at 2.5 percent in Kaufman Hall’s June report, there is little room for a bill nobody forecast.

Accenture’s new report, The CIO’s guide to AI tokenomics, puts numbers on that conversation by drawing on a July 2026 survey of 750 senior executives. Respondents expect token consumption to rise 78 percent over 24 months while per-token prices fall 19 percent, which grows an unmanaged bill by about 44 percent. Only about 20 cents of every token dollar can be tied to a quantified financial outcome.

Accenture prescribes five disciplines to make token spend predictable and provable:

See every workload before it scales

The first discipline is a gateway or observability layer that logs every interaction by workload, team, model, and cost. Organizations that can trace consumption this way score nearly twice as high at tying spend to a financial outcome.

In a health system, that means knowing which service line, clinic, or revenue cycle team is behind the usage. Where AI arrives inside the EHR or another vendor’s product, there is no gateway to install, and visibility has to be negotiated into the contract as usage reporting.

Bill the teams that use it

Showback lets a team see its consumption. Chargeback makes the team pay for it. The share of token spend tied to a financial outcome climbs from about 5 percent with no cost allocation to 23 percent with showback and 32 percent with chargeback. Only 7 percent of organizations use chargeback today.

The report itself says chargeback alone is unlikely to deliver the same results in a less mature organization. And nobody wants a physician rationing an AI scribe to protect a department budget. For clinical tools, showback is the safer first step. Revenue cycle and IT are better places to start billing.

Send each task to the cheapest model that can do it

Accenture sorted 9,368 occupational tasks from the O*NET database and concluded that about 10 percent need a frontier model. Frontier models cost 10 to 20 times more per token than capable mid-tier ones, and 54 percent of AI requests already go to a higher tier than the task requires. As one technology executive told the researchers, “Your Toyota Camry will do just fine. You don’t need to upgrade to a Ferrari anytime soon.”

The caveat is that Accenture’s analysis set aside physical work and anything requiring human review and sign-off. Clinical work that needs a signature sits outside that math. The report also concedes that a cheaper model can need more human review, and when the reviewer is a clinician, that is not a saving.

Require a value case before production

Before a workload goes live, the report wants three answers on paper: what the process costs today, what financial outcome AI will improve, and how that outcome will be measured in dollars. A proposal without answers to these questions is an experiment, and should be funded like one, with a scope limit and a stopping point. Just 35 percent of organizations can calculate cost per business outcome for even their largest use case.

The report measures customer operations by cost per resolved interaction and back-office work by hours per transaction. In a health system, read those as patient access contacts and claims.

Clinical documentation is the hard case. Turning saved clinician minutes into a dollar figure finance will sign is another matter, and the report admits no amount of measurement will isolate the value that belongs to tokens alone.

Make the right choice the easy one

Some 65 percent of companies have introduced user training, but the routing gap remains. The report’s answer is to set the default by policy at the gateway or the desktop. No clinician should be weighing model tiers between patients, so the tool must make the choice for them.

The report also wants finance to make measurable value a condition of investment, HR to build cost awareness into expectations, and every workload to have a named business owner. In a health system, that owner is an operational or clinical leader, not IT.

The tokenomics payoff is part avoided cost and part better reporting

The headline payoff is $5.6 million for every $10 million of annual token spend. About $4.4 million is cost growth avoided by lifting optimization from today’s 23 percent average to the 31 percent needed to hold spend flat. The other $1.15 million is existing spend whose return becomes provable once chargeback and workload-level measurement are in place.

One figure is money not spent. The other is money already spent that can finally be explained.

Organizations in the survey that are still piloting have trimmed about 15 percent of their token use. Those with AI running at scale are at about 28 percent. Holding the bill flat takes 31.