# Clinical AI in late 2026: measured savings, shared liability and an AI scribe shakeout

Health systems are publishing what AI gave them back, the AMA says physicians should not carry clinical AI risk alone, and two AI scribe companies are cutting staff. The technology works in places. The hard questions are now about evidence, money and who answers when it fails.

Three kinds of clinical AI news arrived in the last two weeks, and they fit together. On 29 October Becker's Hospital Review collected the returns 14 US health systems say they got from technology bets, with CommonSpirit claiming "north of $100 million" a year from AI and automation. The same day the American Medical Association (AMA) published a framework arguing that physicians should not be the only safety net when clinical AI goes wrong. Then two well-funded AI scribe companies, Heidi and Abridge, announced job cuts.

On 9 November the American Heart Association released results from a randomised trial at Kaiser Permanente Northern California. Cardiologists finished AI-drafted echocardiogram reports in about half the time, with edit rates no worse than for human drafts. The result is preliminary and not peer reviewed, but it is the kind of evidence the field has lacked.

The market is growing up. This article explains what happened, how the tools work, how to read savings claims, how to evaluate clinical AI, and the basics of liability and governance. It is about technology and operations, not medical advice.

## What happened

### Becker's tallies what 14 health systems got back from tech bets (29 October 2026)

Becker's published "$100M, $40M, $20M: What 14 health systems got back from their tech bets". It gathers figures from different years and sources; it is not a new survey. The caveats matter more than the headlines:

- **CommonSpirit Health**: its CIO, Daniel Barchi, said the system was "probably generating north of $100 million in annual savings" from AI and robotic process automation. Becker's notes this comes from an earlier interview, not a 2026 update. The system had 242 AI applications live as of February 2026.
- **Mount Sinai**: roughly $50 million in estimated AI-related benefit for 2026. Its chief digital officer, Lisa Stump, called it a mix of revenue, avoided costs and productivity, "not a direct bottom-line impact".
- **UNC Health**: about $20 million, reported by its CIO in October 2025, of which around $6 million came from ambient documentation through less night and weekend charting.
- **Tampa General**: $40 million in savings in the 14 months after its command centre launched in August 2019, per a 2020 release. The centre reportedly cost about $13 million.
- **Baptist Health (Jacksonville)**: about $924,000 in estimated annual labour savings from Epic's AI Professional Billing Coding Assistant, which cut coders' chart review time by 20%.
- **St. Luke's Health System (Idaho)**: $13,049 in added revenue per clinician per year from Ambience Healthcare's ambient documentation tool, validated by the research firm KLAS in 2025.

The list also includes a vendor-reported figure, Beacon Health System's $95 million from Xsolis.

### AMA sets out who is liable when clinical AI fails (29 October 2026)

The AMA published an issue brief, "Liability for Clinical AI". Its core idea is that responsibility should follow control. Physicians, it says, should not be held liable for AI performance problems "they did not know of and had no reason to know of".

- **Developers** answer for design choices, training and validation data, and intended use. They must disclose performance and limits; warnings alone are not enough. "Shared responsibility must never become a way to avoid liability," the brief says.
- **Health systems** must validate tools in their own setting, monitor safety and performance, and be able to restrict or switch off a tool.
- **Physicians** answer for the judgement they can meaningfully exercise at the point of care.
- **Patients** harmed by defective AI should have real routes to a remedy that disclaimers cannot remove.

The AMA's March 2026 survey of 1,692 physicians backs the push: 31% ranked clear liability frameworks as their most important regulatory action, the top choice of seven. The same report found 27% had received no AI training and 92% wanted more.

### Heidi cuts about 18% of staff as two co-founders leave (5 November 2026)

Heidi, a Melbourne-based AI scribe company, told staff it would cut about 18% of its global workforce, as reported by Startup Daily and Medical Republic. With more than 600 employees, that is likely more than 100 jobs. A day earlier, co-founders Yu Liu (CTO) and Waleed Mussa (CFO) said they were leaving.

In September Heidi had raised a US$100 million Series C led by Blackbird, plus a US$240 million debt facility for customer acquisition. Medical Republic, citing the Australian Financial Review, reported that chief executive Dr Thomas Kelly said Heidi had been "too optimistic" about how quickly large health systems would buy, and that investors wanted a "more forecastable" business.

### Abridge confirms layoffs in a "realignment" (9 November 2026)

Abridge, a Pittsburgh-based AI documentation company, confirmed layoffs to KRON4 by email without saying how many people were affected. It called the move an "organization-wide realignment" made "from a position of strength" to flatten the organisation, and said it "remains very well capitalized", holding more than half the capital it has raised. Its 2025 Series E raised $300 million at a $5.3 billion valuation.

### AI-drafted echocardiogram reports tested at AHA 2026 (9 November 2026)

At the American Heart Association's Scientific Sessions, Kaiser Permanente researchers presented what they call the first randomised trial of an AI tool that drafts a full echocardiogram report, including measurements and interpretation. It covered 1,200 exams at 21 Kaiser Permanente Northern California centres, with 9 sonographers and 20 cardiologists.

Cardiologists reviewed drafts without knowing whether the AI, a sonographer or another cardiologist wrote them. They substantially changed 50.6% of AI drafts and 49.7% of human drafts, meeting the non-inferiority goal. Finalising took 4.6 minutes with an AI draft against 8.5 minutes without, about 46% less. The AHA stresses this is a conference abstract, not peer reviewed, with no patient outcome data. Co-senior author Dr Keane Lee said "the cardiologist remains ultimately responsible" for the final report.

## How ambient scribes and drafting tools work

An ambient scribe is like a skilled note-taker sitting in the corner of the consulting room. It listens, then hands the clinician a draft to correct and sign. The clinician still owns the note, as a manager owns a letter an assistant drafted.

Most scribes follow the same pipeline:

1. **Capture.** A phone or computer records the visit, with consent where required.
2. **Speech to text.** A speech recognition model, of the kind described on the [Whisper](/tools/whisper) page, produces a transcript and tries to label who is speaking.
3. **Summarise into a note.** A large language model turns the conversation into sections such as history, examination, assessment and plan. Templates and [structured outputs](/tools/structured-outputs) keep each section in the right field.
4. **Suggest codes.** Many products propose billing codes from the note. Much of the reported money comes from here.
5. **Review and sign.** The clinician edits the draft in the electronic health record (EHR) and signs it.

The echocardiogram tool works the same way with different inputs: it reads scan images and measurements and drafts a report. Coding assistants, like Epic's tool at Baptist Health, read documentation and propose codes for a coder to validate.

The shared shape is "AI drafts, human finalises". It puts a trained person at the last step, which is why it is also the source of the liability question. If the human must catch every error, they need the time, information and authority to do so. See [human-in-the-loop](/tools/human-in-the-loop) for the general pattern.

Language models can also write fluent text the source does not support, such as a symptom nobody mentioned, so review must be real, not ceremonial.

## How to read a clinical AI ROI claim

A return-on-investment figure is a recipe, not a fact. A shop's "£20 saving" on a sofa you did not plan to buy is a real discount but not a saving for your budget. The same logic applies here.

### Who measured it

The Becker's list shows several kinds of source:

| Source type | Example | Weight to give it |
|---|---|---|
| Vendor press release | Beacon's $95 million, from Xsolis | Low. A lead, not evidence. |
| Health system statement | CommonSpirit's "north of $100 million" | Medium. Ask for method and period. |
| Third-party validation | St. Luke's $13,049, validated by KLAS | Higher, but check who arranged it and what was counted. |
| Projection | Penn Medicine's expected $105 million through 2028 | A plan, not a result. |

Validations are often arranged with the vendor's cooperation, so they are not the same as an independent audit.

### What counts as "savings"

Most claims mix four things:

- **Hard cost savings**: money that no longer leaves the building, such as fewer agency nurses.
- **Labour time reclaimed**: freed hours, which only become savings if redeployed or not rehired. Baptist Health's $924,000 is "estimated" labour savings.
- **Added revenue**: more income per visit. According to Fierce Healthcare, $9,685 of St. Luke's $13,049 came from better capture of Hierarchical Condition Categories, diagnosis groups used in risk-adjusted payment; $1,907 came from evaluation and management coding and $1,456 from extra patient volume.
- **Avoided costs**: harm or waste that did not happen, such as fewer pressure injuries. Real, but hard to prove.

Better coding can mean the record finally shows how sick patients are, or more aggressive coding that payers may challenge.

### When, against what, and net of what

Check the period and the baseline. Tampa General's figure covers 14 months from 2019; CommonSpirit's comes from an earlier interview. A good claim states what else changed in between. Then look for net figures: Tampa General's $40 million sits beside a build cost of about $13 million, but most claims omit licences, integration, training and governance staff. The article [a score is only as good as its report](/articles/a-score-is-only-as-good-as-its-report) makes the same point for model benchmarks.

## How to evaluate clinical AI before and after go-live

### Non-inferiority, explained

Most drafting tools are not meant to beat people on accuracy. They are meant to be about as accurate while saving time. A non-inferiority trial tests exactly that. Researchers fix a margin in advance, the amount worse the new method may be and still count as acceptable, and check that results stay inside it. Think of a store-brand painkiller: you do not need it to beat the branded one, only confidence that it is not meaningfully worse.

In the Kaiser Permanente trial the outcome was the share of drafts cardiologists changed substantially, and Healio reports a prespecified margin of 5%. In practice the decision rests on the confidence interval: the whole plausible range of the difference must sit inside the margin, not just the point estimate. The [statistics](/tools/statistics) page explains confidence intervals.

### Human edit rate as a quality signal

The edit rate is a metric any health system can track. If clinicians substantially change half of all drafts, as in this trial, the draft is a starting point, not a finished product. That is fine if review is real.

Edit rates mislead in both directions. A falling rate may mean the tool is improving, or that tired clinicians are accepting drafts unread, called automation bias. Pair edit rates with audits in which a second reviewer checks signed notes against the source recording or images. The [evals](/tools/evals) page describes how to build such test sets.

### Drift monitoring after go-live

A tool that worked in the pilot can degrade. Usually the world changed, not the model: a new EHR template, a clinic with different accents, new equipment, a vendor model update. This is drift. It is like a car's wheel alignment slowly going out after potholes; you only notice if you measure.

The AMA brief asks health systems to monitor safety and performance after deployment, and says developer duties continue while they control the technology or see performance data. A basic plan tracks edit rates, error reports and timings by site and specialty, alerts when they move, and re-tests after every vendor update. The [model monitoring](/tools/model-monitoring) page covers the mechanics.

### Triage tools need a different test

Not all clinical AI drafts documents. The University of Texas Medical Branch (UTMB) uses AI agents to review specialty referral backlogs, according to a September 2026 Healthcare Innovation interview with its chief AI officer, Dr Peter McCaffrey. Instead of first in, first out, the agents tell schedulers how urgent and complex each referral looks; McCaffrey stressed they do not make clinical decisions. He said queue time for some referrals fell to "sometimes as low as half", without giving a baseline. For triage, the key test is whether urgent cases are found faster without others being unfairly pushed back, which needs outcome tracking by patient group.

## Liability and governance basics

Liability in clinical AI resembles a crash involving a car's driver-assistance system. The driver, the carmaker and the garage that last serviced the car may each carry some blame, depending on what each could know and control. The AMA proposes the same logic: several parties share liability in proportion to their ability to prevent harm.

Three consequences follow for deployers:

- **Contracts matter.** The AMA says developers should not use contract terms to shift liability for defects they control. Read indemnity clauses closely.
- **Documentation matters.** Developers should state intended use, performance and limits. A [model card](/tools/model-cards), or equivalent, is where this belongs.
- **Conditions for review matter.** If a clinician has 30 seconds to check a long draft, oversight is a formality. The AMA says physician oversight does not replace adequate evidence, sound design or conditions that make review meaningful.

Governance is the machinery that makes this real: an inventory of AI tools, an approval process, named owners, monitoring, incident reporting and an off switch. The [AI governance](/tools/ai-governance) page sets out a general framework, and in Europe the [EU AI Act](/tools/eu-ai-act) adds duties for high-risk systems, which can include AI in medical devices.

## What it means in practice

**The pitch has moved from "time saved" to "money proven".** Buyers want net, validated figures, and vendor press releases will carry less weight.

**Scribes are becoming a feature.** Medical Republic notes that Epic launched its own AI scribe in February and argues that access to the patient record is the real advantage. Standalone scribe firms compete with tools built into EHRs that hospitals already pay for.

**Enterprise sales are slow.** Heidi's chief executive said the company was "too optimistic" about how fast large systems would buy. Procurement, security reviews and integration stretch sales cycles, and burn rates must match.

**Liability is being spread by design.** If the AMA framework shapes law and contracts, developers and deployers carry more duties, which rewards vendors with strong evidence and monitoring.

## What to do now

1. **Keep an inventory of every AI tool in use**, with owner, vendor, version, intended use and the data it touches.
2. **Ask vendors for the method behind every ROI figure**: baseline period, comparison group, what counts as savings and who validated it.
3. **Report hard savings, reclaimed time, added revenue and avoided costs on separate lines** in your own business cases.
4. **Pilot locally with pre-agreed measures**, such as time to sign a note and a maximum substantial-edit rate.
5. **Audit a monthly sample of signed AI-assisted notes** against the source recording or images, to catch automation bias.
6. **Monitor for drift by site and specialty**, and re-test after every vendor model update.
7. **Review contracts** for liability, indemnity, data use, post-deployment duties and the right to suspend the tool.
8. **Give clinicians time and training to review drafts**, and a simple way to report errors.
9. **Check coding suggestions with compliance staff** before counting coding uplift as a benefit.
10. **Plan for vendor failure**: know how to export notes, templates and settings if a scribe company cuts support or shuts down.

## Risks and open questions

**The echocardiogram evidence is preliminary.** It is an abstract from one health system, without patient outcomes, and cardiologists knew they were being studied.

**Many savings figures are old or mixed.** Several Becker's numbers date from 2019 to 2025, some are vendor-reported, and net returns are rarely shown.

**Coding uplift can become a compliance problem.** Gains later judged excessive may be clawed back.

**Automation bias is hard to see.** A falling edit rate looks like progress even when it reflects less careful review.

**Vendor stability is uncertain.** Heidi and Abridge say they are well funded, yet both are cutting staff. The AMA also leaves open when patients should be told AI was used.

**Liability law has not caught up.** The AMA framework is a policy position, not law.

**Watch next:** the peer-reviewed echocardiogram paper, the new structure Heidi has said it will outline, how many staff Abridge cut, whether the AMA framework reaches legislation or contracts, and whether health systems publish net, audited AI returns.

## Frequently asked questions

### What is an ambient AI scribe?

Software that records a clinical visit, turns speech into text and drafts a structured note for the clinician to edit and sign. The clinician remains responsible for the final note.

### Are the reported savings from clinical AI real?

Some are, but read them carefully. Many figures mix hard savings with reclaimed time, added revenue and avoided costs, and some are vendor claims or projections. Ask for the method, the baseline and the net figure after costs.

### What does non-inferiority mean in an AI trial?

The trial tests whether the new method is worse than the current one by no more than a margin agreed in advance. It suits tools that aim to save time while keeping quality about the same, like the Kaiser Permanente echocardiogram trial.

### Who is responsible when clinical AI makes a mistake?

Under the AMA's proposed framework, responsibility is shared according to who could know about and control the problem: developers for design and disclosure, health systems for selection and monitoring, and physicians for judgement they can meaningfully exercise. It is a policy position, not settled law.

### Why are AI scribe companies cutting jobs if the technology works?

Selling to large health systems is slow, competition is strong, and EHR vendors such as Epic now offer their own scribes. Heidi's chief executive said the company had been too optimistic about sales speed, and Abridge called its cuts a realignment to flatten the organisation.

## Summary

Clinical AI entered a more sober phase this autumn. Becker's gathered returns from 14 health systems, from CommonSpirit's claimed "north of $100 million" a year to Baptist Health's $924,000 in coding labour savings, but many figures are old, vendor-reported or mix revenue with savings. The AMA proposed that liability be shared among developers, health systems and physicians according to who controls the risk, backed by a survey in which 31% of physicians named clear liability rules their top priority. AI scribe companies Heidi and Abridge cut staff as enterprise sales proved slower and EHR vendors moved in. A Kaiser Permanente trial found AI-drafted echocardiogram reports nearly halved finalising time with non-inferior edit rates, though it is not yet peer reviewed. The lesson for buyers and builders is consistent: demand measured, net evidence, keep review meaningful, monitor for drift and write responsibility into contracts.

## Sources

- [Becker's Hospital Review: $100M, $40M, $20M: What 14 health systems got back from their tech bets](https://www.beckershospitalreview.com/healthcare-information-technology/digital-health/100m-40m-20m-what-14-health-systems-got-back-from-their-tech-bets/)
- [American Medical Association: AMA advances framework for clinical AI liability](https://www.ama-assn.org/press-center/ama-press-releases/ama-advances-framework-clinical-ai-liability)
- [American Medical Association: Issue brief, Liability for clinical AI](https://www.ama-assn.org/system/files/liability-clinical-ai-issue-brief.pdf)
- [American Medical Association: 2026 Physician Survey on Augmented Intelligence](https://www.ama-assn.org/system/files/physician-ai-sentiment-report.pdf?stream=top)
- [Startup Daily: Heidi cuts headcount by 18%, as two cofounders depart](https://www.startupdaily.net/?p=141486)
- [Medical Republic: Heidi and Canva cuts change the game for all of local health tech](https://www.medicalrepublic.com.au/heidi-and-canva-cuts-change-the-game-for-all-of-local-health-tech/130732)
- [KRON4 via Yahoo Finance: $5+ billion AI healthcare company confirms layoffs](https://finance.yahoo.com/healthcare/articles/5-billion-ai-healthcare-company-190358679.html)
- [American Heart Association: AI-interpreted echocardiograms comparable to human reports, cardiologists' time cut by about 50%](https://newsroom.heart.org/news/ai-interpreted-echocardiograms-comparable-to-human-reports-cardiologists-time-cut-by-about-50)
- [Healio: AI-interpreted echocardiograms comparable to human reports, may halve reading time](https://www.healio.com/news/cardiology/20261109/aiinterpreted-echocardiograms-comparable-to-human-reports-may-halve-reading-time)
- [Fierce Healthcare: Ambience delivers strong financial returns for St. Luke's Health System](https://www.fiercehealthcare.com/ai-and-machine-learning/ambience-klas-roi-report)
- [Becker's Hospital Review: Baptist Health AI coding cuts chart review 20%, saves $924K](https://www.beckershospitalreview.com/healthcare-information-technology/ai/baptist-health-ai-coding-cuts-chart-review-20-saves-924k/)
- [Healthcare Innovation: How UTMB's Chief AI Officer views agents' roles in health system workflows](https://www.hcinnovationgroup.com/analytics-ai/agentic-ai/article/55407181/how-utmbs-chief-ai-officer-views-agents-roles-in-health-system-workflows)