Match model tier to workflow ambiguity
Your token bill spikes when every workflow hits the same flagship model, even the ones that only extract codes or normalize forms.
Sean Gates
June 4, 2026 · 6 min read
The board deck shows a hockey-stick token line. Nursing documentation, revenue-cycle bots, and the member FAQ all hit the same flagship model. Error rates on simple tasks barely moved. Call it what it is: an ambiguity tax. You paid frontier prices for work that was never ambiguous.
The conventional bet
Health systems buy one enterprise AI relationship and point every workflow at the newest frontier model. Patients are complex, compliance is scary, and "best model" feels like responsible governance.
Half your volume is structured anyway: eligibility checks, code extraction, form normalization, routing patients to the right queue. Those jobs need reliable parsing at a fraction of frontier pricing. Route them like a shuttle between terminals; save long-haul capacity for the trips that need it.
Tiered intelligence
In the Eight Words framework, Model is the brain you rent: capability and cost profile. Strategy means an intelligence portfolio: deliberate tiers behind a Gateway that sends each request to the cheapest brain that still clears the bar.
- High ambiguity, high consequence: patient narrative that could be chest pain; payer policy edge case with conflicting clauses; clinician-facing draft that shapes next steps.
- Lower ambiguity: summarizing a fixed template; mapping ICD suggestions from structured fields; internal FAQ with citations locked to a knowledge corpus.
Frontier models earn their invoice on the first bucket. The second bucket is where finance bleeds.
A decision your CFO can defend
Ask delegates to map your top ten AI workflows on two axes: consequence (patient, clinical, financial impact) and ambiguity (how open-ended is the input?). Each cell gets a default tier.
| Low ambiguity | High ambiguity | |
|---|---|---|
| High consequence | Smaller model + mandatory human review | Frontier + human review |
| Low consequence | Smaller model, automate | Mid-tier or frontier with sampling |
You are asking finance to stop subsidizing vendor marketing on work a sub-dollar call can handle. Clinicians still review what matters; the change is which model drafts the structured steps.
Consistency comes from orchestration, system messages, knowledge citations, and sign-off rules. One weight file for every prompt was never consistency. It was one invoice.
What changes when routing is real
Vendor negotiations shift from site licenses to routing economics. Gateway policy belongs on the executive dashboard instead of a buried integration ticket. Staff learn when to escalate to frontier versus accept a draft from a smaller model. That is operational literacy.
The clinical pushback is predictable: "We cannot have two standards of care." Undifferentiated frontier spend does not create one standard. It creates one bill. Nurses still paste into consumer tools when the sanctioned stack is slow or expensive. Tiering plus governance beats performative uniformity.
Delegate deliverable this week
One page: workflow list, matrix tags, proposed default tier per row. Day one often needs routing policy more than a new vendor.
Tokenizer economics are a separate line item from which brain earns the work—alphabet charges before reasoning even starts.