Nobody can prove their accuracy.

Where this started

A new client inquiry, relayed by voice note: an AI tool for medical coding and billing. Three questions came with it.

The problem, first principles

Every US clinical encounter must be translated into codes — ICD-10 for diagnoses, CPT for procedures, HCPCS beyond that — before a claim can be billed. Manual coding is slow, expensive, and error-prone: under-coding leaves revenue on the table, over-coding invites audits and penalties. The tool's job is to turn clinical documentation into an accurate, auditable code set with minimal manual entry.

How the research ran

Four camps, none covering the whole job

The US AI coding market has split into four camps. The closest thing to a verified autonomous, end-to-end, EHR-agnostic coding product does not exist yet.

The player table

25 players, copyable comparison. Scroll horizontally on smaller screens.

Eight gaps a new entrant could be better at

Where the field is weak, unverified, or structurally exposed — ranked by how directly a focused entrant could exploit them.

Who buys, what triggers them, what blocks them

Discovery, led by sparky-sai: buyer personas, buying triggers, objections with answers, a GTM sketch — and five critic flags that corrected the brief's premises.

Critic flags — read first

Buyer personas

The signers

The users

The vetoers

Buying triggers, ranked

1. Denial rate pain. A practice watching 8–12%+ of claims denied or downcoded has a number on the wall. Denial prevention is claimed by many vendors but unevidenced — publishable denial-rate deltas would cut through the noise.
2. Coder shortage / turnover. Certified coders are scarce and expensive; backlogs grow; DNFB days creep up. The staffing trigger that makes “copilot” framing welcome rather than threatening.
3. Audit finding or payer clawback. Nothing concentrates a CFO like an OIG or payer audit. Post-audit buyers demand evidence trails — exactly the differentiation hypothesis.
4. Undercoding revenue leakage. Specialty practices routinely leave money on the table with conservative coding. A tool that finds missed legitimate codes sells on found revenue, not cost savings.
5. EHR migration or upgrade. Rip-and-replace moments reopen every build-vs-buy decision — and the moment native EHR coding features get evaluated. Win the comparison or lose to the bundle.
6. PE-backed practice rollup. Consolidators standardize the revenue cycle across dozens of practices — one decision, many seats. Longer cycle, bigger prize.

Objections — and what actually answers them

GTM sketch

Channels, in order of expected cost-per-closed-deal

1. RCM/billing-company partnerships — revenue-share or white-label. They own the practice relationship; the product rides it.
2. Specialty societies and coder associations (AAPC/AHIMA chapters, HFMA) — education-first: publish the accuracy methodology, not a pitch.
3. Practice-management consultants and MSOs — they advise P1 buyers at moments of pain. One relationship, many doors.
4. EHR marketplaces (athenahealth, eClinicalWorks, NextGen) — meet buyers where they shop; avoid the Epic certification tax early.
5. Pilot-led direct — 3–5 referenceable pilots with published results. Pilots are the marketing engine, not a pre-sales cost center.
6. Physician communities — documentation-burden relief stories travel peer-to-peer. Never sell to doctors; let doctors pull.

Messaging angles (test in discovery, in this order)

1. “Every code, evidence-linked.” Audit defense, not speed — lead with the answer to the compliance veto.
2. “Find the revenue you're already owed.” Undercoding recovery — sells on found money.
3. “Your coders, unblocked.” Copilot framing; backlog and DNFB relief. Never “replace your coders.”
4. “Priced in the open.” Transparent pricing against a field where nobody publishes any.
5. “One specialty, done completely.” Depth over breadth — the anti-EHR-bundle position.

Discovery interview guides

Run 12–20 interviews before build scoping: 6–10 physicians, 6–10 coders/billers, 2–4 administrators or RCM operators. 30–45 minutes each; stop at saturation, not at a fixed number.

+ Physicians — the adoption gate (9 questions)

1. Walk me through your last clinic day — when did documentation happen, and how much leaked into the evening?
2. Who codes your visits today — you, a coder, both? Where does it go wrong?
3. Tell me about the last claim denial or downcode that annoyed you. What happened, and what did it cost you in time?
4. Have you ever worried a visit was undercoded — work you did but didn't get paid for? Give me an example.
5. Has your practice ever been audited, or come close? What changed afterward?
6. What AI or ambient tools touch your notes today? What do you actually trust them with, and what do you double-check?
7. Assist-vs-autonomous probe: “If a tool drafted your ICD-10/CPT codes and you just approved them, would you use it? What if it submitted directly and you only reviewed exceptions — what would you need to trust that?”
8. What would have to be true — evidence, guarantees, who stands behind it — for you to let software touch your coding?
9. Who in your practice would block this purchase, and what would they say?

Close: “If we built the thing you just described, would you pilot it? What would the pilot have to prove in 90 days?”

+ Coders & billers — power users and potential saboteurs (9 questions)

1. Walk me through a chart from note to submitted claim. Where does your time actually go?
2. What are the top three coding errors you see repeatedly — in your specialty, from physicians' documentation?
3. What does your backlog look like right now? What happens when a coder quits?
4. Have you used any AI coding or CAC tools? Where do they get it wrong — give me the specific failure modes.
5. Trust probe: “What would an AI suggestion need to show you — evidence, confidence, what — before you'd accept it without re-checking?”
6. If a tool cleared your routine charts and left you the complex ones, is that a better job or a threat? Be honest.
7. Who is liable, in your view, when an AI-suggested code turns out wrong — you, the physician, the vendor? What should the answer be?
8. What does your compliance team worry about most — upcoding, downcoding, audit trails, something else?
9. For billers: where do denials actually come from — coding, documentation, eligibility, payer behavior? What would “denial prevention” have to do to be real and not marketing?

Close: “If you were designing this tool, what's the one thing it must never do?”

+ Administrators & RCM operators — the signers (6 questions)

1. What does your coding function cost you today — staff, outsourced, or both? What does the backlog cost you in DNFB days?
2. What is your denial rate, and what share is coding-related? What would a 2-point improvement be worth annually?
3. When you last evaluated a coding/AI vendor, why did you say no?
4. How do you buy software — who has to say yes, and how long does it take?
5. What would transparent per-claim or per-provider pricing need to look like for you to trial something without a six-month procurement?
6. Would you pilot with 90-day published results? What metrics would the pilot have to move?

The recommendation: conditional build

Phase 4, by sparky-zero. Build a focused wedge product. Partner for distribution. Pass if the conditions fail. Nothing here authorizes spend — it scopes what a build would take so the decisions below can be made with eyes open.

The seven decisions — builder's recommendations

These are for the client and the co-CEOs to resolve. Below is what the build wants, and why.

MVP scope

Architecture

Phased plan

Team (phases 1–2): 4–6 engineers (backend, NLP/ML, frontend, integrations), 1 compliance/privacy lead, 1 certified derm coding SME (contract), part-time healthcare counsel. Effort: ~18–30 engineer-months to pilot-ready MVP — stated as effort, not dollars; the budget envelope is still an open question.

The research holds

sparky-118 checked 35 competitive claims with named sources. Verdict: the conditional-build recommendation survives — no structural claim is overturned.

What the verification changed or confirmed

1. The GAO finding is the load-bearing one. GAO Science & Tech Spotlight GAO-26-109116, “AI for Medical Notes and Coding” (July 2026): accuracy “may be difficult to verify.” Government weight behind the verification-first thesis — the single most authoritative support in the whole package.
2. The conditional-build window is open but narrowing. Fathom is the nearest end-to-end counterexample; Akasa claims vendor-reported blinded evals. Neither has independent named-source confirmation. Re-check both on every future sweep — either could flip the thesis.
3. EHR absorption verified on all three legs (Epic Penny early adopters, Oracle Upstream AI autonomous profee coding, athenahealth bundling at no extra cost) — hardens the off-Epic specialty wedge as the structural position, not just a preference.
4. No independent head-to-head accuracy audit exists — verified. Independent accuracy-benchmarking has no direct incumbent: the strongest single support for “published accuracy as the trust moat.”
5. SmarterDx sits inside New Mountain's “Smarter Technologies” roll-up — one more data point for the “vendor continuity” objection. Footnote, not a recommendation-mover.
6. Never cite CorroHealth revenue. Neither ~$380M nor ~$2.3B has a credible source. It appears nowhere on this page.
7. Safest external comps to cite: Candid, Suki, Waystar (public), Nabla. Use these, not the unverified ones, in partner and diligence conversations.
8. Do not price on Abridge's $316M extension — the $5.3B valuation is well-sourced; the extension rests on secondary citation of an unseen piece.
9. Corrections applied: Suki is assisted, not autonomous. Abridge's $775M VA figure is a multi-vendor program ceiling, not Abridge revenue. Nabla is independent (Microsoft's deal was Nuance, 2021). The Augmedix–Commure date was the agreement announcement, not closing.
10. Timing risk: Akasa's launch is days old; Epic Penny and Oracle Upstream AI are weeks old. The autonomous landscape is shifting under this verdict — published-accuracy-first design is the correct hedge.

Watch items for the next sweep

• Fathom: any independent validation of its KLAS/accuracy claims.
• Akasa: independent corroboration of its claimed blinded third-party evals.
• Any head-to-head accuracy study published by anyone — it would weaken the moat.
• AGS–Blackstone exact closing date; any credible CorroHealth revenue figure.