Cleared to Paid · Methodology
A consulting project, staffed by agents
Cleared to Paid runs like a phased strategy engagement. AI agents gather and check the evidence; I frame the questions and make the calls.
Oct 13th 2026
The FDA has cleared more than 1,600 AI medical devices. Who actually pays for them, and why do so few get paid by insurers? Answering that is a classic strategy question, so the project is run the way a consultancy runs an engagement: hypotheses first, a phased workplan, quantitative and qualitative evidence, a red-team review, and a synthesis graded for confidence. The difference is the team. AI agents do the legwork to written rules, a separate agent tries to break their answers, and I make the judgement calls.
The approach at a glance
| Consulting practice | How it runs here |
|---|---|
| Hypotheses first | The project starts from written hypotheses, scored on a scorecard as evidence comes in. One starting belief, that FDA AI clearances are stagnating, has already been proved wrong. |
| Phased workplan | Five phases, each with its own question, methods and articles (below). |
| Quantitative analysis | Scripts over every FDA AI-enabled device, Medicare claims and payment rules. Labels come from a classifier whose error rates are measured, with 95% intervals. |
| Qualitative research | Case studies of 36 payment arrangements in four countries, a 28-device case review checked against makers, FDA and CMS, and scans of the value chain and of 21 health systems' own AI. Interviews come in a later phase. |
| A sourced fact base | Every claim an agent brings back carries a verbatim quote, a source link and an A, B or C grade. |
| Red-team review | A separate agent tries to break the answers before they are used. |
| Partner judgement | I set the rulings on edge cases and make the final calls. |
| Synthesis with confidence | Each takeaway is rated high, medium or low, with the reason stated. |
Five phases
- Phase 1The moneyUnder way
- Phase 2The buyer
- Phase 3The builders
- Phase 4The verdict
- Phase 5The next wave
Phase 1 maps who pays for every AI device the FDA has cleared. The later phases cover the buyer, the builders, the verdict and the next wave. Each phase ends in a set of articles.
How the evidence is made
Six steps turn public data into a finding, in this order.
- Data pipeline
- Research agents
- Spot-checks
- Adversarial check
- Owner rulings
- Published error rates
1. Data pipeline
Reproducible scripts build the device list, the payment routes and the charts. The device list is the FDA's own: 1,614 AI-enabled devices, one row per FDA submission. Payment data come from CMS fee schedules and claims files, inpatient add-on payment rules in the Federal Register, Medicare contractor documents and makers' coding guides.
A device gets a payment route only when a CMS rule, a contractor document or the maker's own coding material names it for a code. Matching a code by function is not enough, so the counts are a floor. Every correction to the derived data is logged in a change log, row by row.
2. Research agents
About ten AI research agents work in parallel. Each one gets the same written evidence rules:
- Quote the source verbatim, in 25 words or fewer.
- Grade every piece of evidence A, B or C (table below).
- Never bypass a paywall, a login or a bot check.
- Stay away from an excluded-company list.
The agents researched three payment routes: existing service codes, services run by device makers, and AI sold as a scanner option. They also scanned the clinical AI value chain and the AI that health systems build for themselves.
They also checked labels. 238 open labels across 105 sampled devices were checked against the full FDA summaries. Where a summary's text was garbled, the agents rendered the PDF pages as images and read those.
3. Spot-checks
The lead agent re-verifies key claims at the primary source: SEC filings, CMS rules and makers' own pages. A company is named as a leader only alongside a measure and a source, and rows without a source are dropped.
4. Adversarial check
A separate agent tried to break 42 of the answers. None was wrong. One was corrected and two were downgraded to unsure.
5. Owner rulings
Some calls need a person. I set 13 rulings for edge cases, such as how to label a summary that lists only hardware requirements, and I reviewed every label the agents left unsure.
6. Published error rates
The labels behind the findings come from a rule-based classifier reading FDA summaries. Its accuracy is measured against the reviewed labels on a random sample of 100 devices, and published with the results, as in the evidence section of the full report.
Mostly right, and measured
Classifier labels matching reviewed labels, random sample of 100 devices, % with 95% interval
Deployment is not used in any finding. Company type was measured before the fix this check triggered. The modality and training-data rules were fixed after the sample's disagreements had been seen, so those two rates may flatter the classifier slightly.
Show the numbers
| Field | Match | 95% interval | Note |
|---|---|---|---|
| Radiology function | 98% | 93–99% | |
| Model type | 96% | 90–98% | |
| Company type | 94% | 88–97% | Measured before the fix it triggered |
| Clinical area | 85% | 77–91% | |
| Modality | 83% | 74–89% | |
| Training-data source | 79% | 70–86% | |
| Deployment | 50% | 40–60% | Not used in any finding |
Source: Cleared to Paid label review, October 2026Healthcare AI Insights
Evidence grades
| Grade | Basis |
|---|---|
| A | A primary document: a rule, a filing, a regulator's page or the maker's own document |
| B | A company page or reputable trade press |
| C | An inference, a single secondary source, or not verified |
What the numbers do not claim
- Clearance is not use. A device without a payment route may still be widely used and paid for by a hospital.
- No route found is not no payment. Routes are a floor, and private payers are not covered.
- Medicare is not every payer. The claims data cover traditional Medicare only and lag by two years.
- A rate is not coverage. A national price does not mean every Medicare contractor covers the test.
Use of AI tools
The multi-agent research workflows are built with Claude Code. AI assistants helped write the scripts, draft labels and search public sources. Every number in the report is computed by scripts from the data files. Every agent-sourced claim carries a source link and a grade. Claims the agents could not verify are marked as unverified or left out.
Data, code and licences
The code and data will be published in a public repository. FDA and CMS data are US government works. CPT code descriptions are AMA copyright, so codes are described in my own words. No patient-level data is used.