Blog
|

In a health system, AI spends most of its time:

  • Documenting a visit into a structured note
  • Condensing a chart before a consult
  • Drafting a prior authorization letter
  • Suggesting billing codes
  • Reformatting a referral
  • Answering a portal message

Thousands of these simple tasks a day in a mid-sized system. It’s transcription, summarization, and reformatting, which language models handled competently three generations ago.

However, a smaller set is more challenging. A differential when labs and imaging disagree, or a history buried in 200 pages of outside records. That work earns the more expensive AI.

…a routine visit note and the hardest diagnostic question can cost the same to answer.

Without deliberate model routing, a routine visit note and the hardest diagnostic question can cost the same to answer.

New to some of these terms?

  • Token: The unit AI models read and write in. Ordinary prose runs about three-quarters of a word per token, but clinical text is denser, since drug names, lab values, and dosages fragment into several tokens each. Output tokens usually cost several times more than input tokens.
  • Inference: The act of running a model to get an answer. Training builds the model; inference is what you pay for every day afterward.
  • Model Routing: Sending each request to the cheapest model likely to handle it well, and escalating to a more expensive one when the cheap answer isn’t good enough.
  • Open-weight model: A model whose trained values are published, so your organization can run it on hardware you control instead of paying a vendor per token. Open weights don’t mean open licences, and running the model yourself trades token costs for infrastructure costs.
  • SLM: A model with far fewer parameters than a frontier model. Cheaper and faster to run, and competitive on narrow or well-defined work.
The Stakes

Rising Spend and Unfunded Controls

An unbudgeted AI bill can be the difference between a profitable quarter and a loss.

Icon for Thin Margin

0.4%

Median health system operating margin through May 2026

Strata Decision Technology

Icon for Rising Spend

85%

Healthcare organizations planning to increase AI budgets in 2026

NVIDIA healthcare and life sciences survey

Icon for Shrinking Capital

41%

Health system executives expecting capital investment to decline over the next two years

Sage Growth Partners

Icon for Unfunded Controls

26%

Hospitals planning to raise AI governance and safety budgets by two or more percentage points

Black Book Research

Adoption can outrun the funding behind it. The clearest example on record is outside healthcare. Uber gave 5,000 engineers access to Claude Code in December. Usage went from 32% to 84% by March, at $500 to $2,000 per engineer per month. AI budget was gone four months into the year. Uber absorbed it.

Model Triage

Routing Works and the Savings Are Measured

Routing happens to be one of the better-studied problems in applied AI.

RouteLLM, out of UC Berkeley, predicts how hard each request is before choosing a model. On one benchmark it routed 14% of queries to the expensive model, cut cost by more than 85%, and recovered most of the quality difference. Stanford’s FrugalGPT tries cheaper models first and escalates only when the answer falls short, reporting reductions of 59% to 98% while matching the accuracy of the best single model.

Most requests don’t need your best model.

For a health system, we’d start with three tiers.

Local

Open-weight or small models, on hardware you run

  • Ambient documentation
  • Chart summaries
  • Referral reformatting
  • Billing code suggestions

Mid-Tier

A commercial model, not the flagship.

  • Prior authorization drafting
  • Patient portal replies

Frontier

The best model available that month

  • Differentials where findings conflict
  • Synthesis across hundreds of pages
Spending Caps

Budgets Belong Where the Work Happens

Icon for Budget

Budgets by Department

Radiology, revenue cycle, and ambulatory operations have different AI workloads and different tolerance for spend.

Icon for Caps

Caps Before the Invoice

A hard limit per user, per project, and per model, enforced at the platform rather than reconciled afterward.

Icon for Monitoring

Token Monitoring in Real Time

Consumption by person, team, project, and model, visible today rather than next month

Icon for Agents

Circuit Breakers for Agents

Agentic tools consume far more than chat. Anthropic reports that agents use roughly 4x the tokens of a chat interaction

Shadow AI

There’s Another AI Bill Arriving

The Shadow AI Problem Is Already in Your Environment

Unapproved tools appear in no budget and no log.

Read the shadow AI analysis

Governed but not restricted

Where Sovereign AI Fits

Armor built Sovereign AI as a private AI work platform for regulated industries. Every request from your people to any model, frontier or local, passes through one governed control plane. A model router assigns work by tier so routine tasks stop hitting premium endpoints. Spend is metered and capped by user, team, and organization before the bill rather than after it. Policy runs on every request, and every prompt, response, and agent action is written to an audit log you can export.

One control layer, and it answers finance and security with the same evidence.

See how it works at sovai.com

About Sovereign AI

Sovereign AI is the governed AI work platform for the whole company, built by Armor. One control layer for every model, every team, and every dollar, inside your walls, under your rules. Armor has spent 17 years securing regulated industries, protecting over 1,700 organizations across 40+ countries held to the highest compliance bars in the business. Sovereign AI is what that experience looks like as a product. Learn more at sovai.com.

About Armor

Armor to protect regulated data in the public and private cloud. AI is the next risk, so Armor built Sovereign AI to bring that same protection and compliance to how organizations use AI in the workplace: a fully governed platform that lets them leverage AI without creating undue risk to their data and their regulatory obligations. For more information, visit armor.com and sovai.com and follow us on LinkedIn.