Healthcare AI Costs: Why Agent Workflow Design Matters More Than Model Price

Prior authorization runs on staff time, and a lot of it. According to the American Medical Association’s 2025 prior authorization physician survey, physicians and their staff spend 13 hours a week on prior authorizations, completing an average of 40 per physician. That's the baseline an AI agent has to beat, and it raises a practical question for healthcare leaders: Will the cost of running an agent come in far enough below the cost of today's process to justify the investment?

Per-token prices have dropped fast, which makes it tempting to skip that question and go straight to Scaling AI Without Scaling Costs Starts With Token Awareness | CDW. But a model's cost only covers one step, and agents rarely stop at one.

Healthcare is especially sensitive to agent design. A single prior authorization draws on multiple data sources, payer requirements change often and clinical documentation is long and complex. Human review is frequently required, and compliance adds steps of its own.

Each of those hurdlers gives an agent more to read and more to redo. The same step that runs cheaply on a clean case can cost many times more when the chart is incomplete or the payer's rules have changed. Enough of those spikes can wipe out the return on investment.

Lower prices made multistep workflows affordable, so agents are getting bigger jobs. Economists call it the Jevons paradox: When a resource gets cheaper to use, people tend to use more of it and total spending can go up anyway.

AI tokenomics follows the same pattern. That's how savings at the model level can disappear at the workflow level.

Picking the right model still matters. Many organizations do it by hand, choosing from a drop-down menu or assigning each agent a fixed model. A routing layer built into the agent architecture makes that choice dynamic, matching each step to the model that best balances capability and cost. But routing decides which model runs. Agent design decides how many times it runs.

In a work queue, a prior authorization review looks like one task. Inside the agent, it's a chain of them: retrieve the patient's records, check the payer's rules, summarize the findings, retry a failed retrieval and escalate to secondary review when something doesn't add up. Depending on the case, that's several model calls for what a budget spreadsheet counts as one review.

Those calls also get heavier as they go. Models don't remember earlier steps on their own, so, many agent designs resend the full history with each call. By the seventh step, the agent may be paying to reread instructions and records it first saw at the start, and costs can climb faster than the step count. IT gets a monthly bill that's hard to explain, and clinicians can end up waiting longer on authorizations for the care they've ordered.

Three design choices drive most of that total: orchestration, retrieval and how the agent handles reasoning and retries.

Take a routine outpatient MRI request, the kind a prior authorization team handles often.

Left to figure out its own path, the agent pulls the full chart, then loads the payer's entire policy document to find the one section that applies. When the call to the electronic health record times out, it tries again and again, carrying everything it has gathered into each attempt. Then the reviewer notices the summary skipped the prior treatment history the payer asks for and sends the agent back for another full pass.

A well-designed workflow handles the same request differently:

• Orchestration. The steps for this type of request are known, so standard software runs them in order and the model steps in where judgment is needed, like reading clinical notes against the payer's criteria. That keeps token use roughly in line with the number of steps.
• Retrieval. The agent pulls only the records the payer's criteria call for, such as relevant diagnoses, prior imaging and previous treatments, and builds a structured summary that later steps reuse. Payer rules for common services are saved once and refreshed when the payer updates them. Pulling fewer records also lines up with HIPAA's minimum necessary standard.
• Reasoning and retries. A routine reauthorization gets a lighter review than a complex case. Retries are capped, so a second failed retrieval gets handed to a person. And because the summary is built around the payer's criteria, reviewers aren't sending the agent back for missing pieces.

Human review is a compliance requirement, but it also works as a cost guardrail. A clear handoff rule gives the agent a place to stop: When its confidence in a clinical detail falls below a set threshold, or the documentation is missing, the case goes to a clinician instead of into another round of retrieval and reasoning.

Clinical and operations leaders set those thresholds and priorities, and written escalation criteria help keep routine cases out of the clinician queue.

Cost per token tells you what the model charges. It doesn't tell you what a prior authorization costs. For that, track three things:

1. Cost per workflow. This is everything one authorization takes, retries included.
2. Cost per successful outcome. Approval rates hinge on payer decisions the agent doesn't control, so define success as a complete, accurate submission that doesn't bounce back for missing information.
3. Retries and handoffs per case. Tag each agent run by workflow and step. If retries climb for one payer after a policy update, that points to a retrieval or rules problem, not a model problem.

Tracking these numbers is central to TokenOps, which brings FinOps discipline to AI token spend. It also shows your token yield rate, the share of tokens that went toward a complete, accurate submission instead of retries or rework. Better agent design raises that rate and brings cost per workflow down.

Workflow numbers help with forecasting, too. You likely already know your prior authorization volume. Multiply it by the cost of a completed authorization, and you can forecast AI spend with far more confidence than a per-token estimate allows.

A token budget for each workflow keeps that forecast honest when a case runs long. Then compare your forecast with the staff hours the current process consumes. That gives your finance team an ROI model it can follow, built on outcomes instead of token math.

Model price is one line item. Agent design determines how many times you pay it.