Skip to content

AI, decoded

What does it actually take to ship AI in a regulated industry?

Provenance on every component, not a single passing benchmark. Corti CEO Andreas Cleve's requirement is that you can explain what each part of the pipeline did, why it changed, and what went into it. He also doubts that public leaderboards are good evidence: he observes that noncommercial models lead the big medical benchmarks and suggests, with a hedge, that this is because they are trained to do well on tests rather than in practice. Cleve told Chain of Thought that in Corti's workflow a doctor reviews the output, which gives the company what he calls a golden label for whether each result is complete and concise enough to use.

· Updated · Chain of Thought

Level 3: Leading with AI · 3.5 Guardrails, risk and regulation

Enterprise AIAI Evaluation & Reliability

  • 4 episodes
  • 4 guest clips
  • 8 glossary terms
A food-style ingredients label ticking off data in, model card, where it's hosted, encryption and why it changed, beside the words: show your work, every component, every change.

In detail

1. The slowness is a bar, not a blind spot

Andreas Cleve, CEO of Corti, rejects the premise that regulated industries are behind. On healthcare he is blunt that the laggard story is “super boring,” (3:59) and offers the counter-observation that a clinic packs more devices per room than his own machine-learning office does (3:33). His reading: “they adopt tons of technology, and they do it quite fast, but they have a pretty high bar for doing it” (3:33, 3:59).

That reframes the problem. The obstacle is not appetite, it is the evidence you have to produce before anyone will let the thing touch a patient.

2. Every component needs its own benchmark

The concrete requirement is provenance across the pipeline rather than a score at the end. Cleve describes it as a factory: “In the factory, every component that’s interacting has to have benchmarks, so we need to understand remove data, what happens, where does it happen?” (30:43)

And the artifact regulators and buyers want is an explanation, not a number: “we need to be able to actually explain what we do, why we did it, why we changed it, when we did, what’s in the model cards, like what’s in the food we’re serving, what’s the ingredients we used, why do we use them, when do we move them, how do we launch it, where is it hosted, at what level and what state are we encrypting” (31:45).

For high-risk systems, the EU AI Act writes part of this into law: Article 11 requires technical documentation and Article 12 requires the system to technically allow automatic logging of events over its lifetime.

A suggested evidence packet, not a regulatory submission:

RecordWhat it should let a reviewer answer
Intended use, users and limitsWhat the system is meant to support, and which uses were never tested
Inputs, component versions and settingsWhich data and software produced the result under review
Test cases, reference method and known failuresWhat was measured, on which population, and what is still uncertain
Human review and action controlsWhat the reviewer can see, correct or stop, and who signs off the consequence
Change history and monitoringWhat triggers a fresh review, and who owns it

For medical devices, the good machine learning practice principles from the MHRA, FDA and Health Canada (updated 5 October 2026) ask for much of this: data representative of the intended patients, test sets independent of training, reference datasets whose limits are understood, and testing of the human–AI team rather than the model alone. They are medical-device principles, not a checklist for every regulated field.

If you are scoping a regulated build, here is his summary of the shape of the work: “you can’t win health care by being good at one thing. You have to be good at many things” (32:16).

3. Public benchmarks are weak evidence, by his own argument

Cleve makes an observation about medical leaderboards that cuts against the usual proof point. On the big public health datasets, “the leaders are noncommercial models” (31:14). His inference is hedged and he says so: “It’s probably because it’s not very useful in practice… there’s probably really good reasons why it was trained in a way that makes it really good at tests but not very good in the wild” (31:14).

What replaces the leaderboard is layered checking. He describes “a series of models as judges passing judgment on whether or not this reasoning pipeline gets it” rather than batch processing (39:35), plus a large stock of holdouts. See what is LLM-as-a-judge for when that check can be trusted, and how do you evaluate an AI system when the output isn’t deterministic for the surrounding method.

4. The workflow can hand you a label for free

Corti’s product sits in a workflow where a clinician has to approve the output. That approval is a graded signal, not just a UI step. “a doctor has to say, what you did was cool. So we actually at the end get a quite cool golden label of whether or not it’s as complete or as concise as somebody would actually use it for” (41:24).

If a regulated workflow already requires human sign-off, that gate can feed an evaluation set. It isn’t one automatically. An approval click can mean “good enough to send” rather than “correct,” and reviewers miss errors. Record what the reviewer checked, against which criteria, and what they corrected. Keep the cases you learn from separate from the ones you test on, so the final check uses data the system hasn’t been tuned on. What is LLM-as-a-judge covers the same split for model graders.

5. Compliance turned out to sell

Cleve reports the inversion directly: “Being from Europe, we’re really good at compliance… If I were doing AI for many other things in the world, I wouldn’t be saying this. It would be an inhibitor” (34:51). Instead a large US customer with a long purchasing cycle came to them, and part of what they liked was alignment with the EU AI Act and GDPR, on the buyer’s expectation that “no matter what the market will drift towards some of this stuff” (35:15).

He does not oversell it. He grants that some of the red tape “we should remove,” while holding that under it sits “a nucleus of truth” about needing to trust the system in a high-stress environment (35:15).

6. Time saved is not the same as work removed

The counterweight in the episode is about AI scribes. Corti surveyed 2,000 early adopters in the US, and Cleve reports “they told us on average they spend at least three hours a week just, like, vetting and correcting their AI,” alongside a standing worry that something was misrecorded (37:32). His verdict on that pattern is that “we’re creating workflows here that shouldn’t be there” (38:25).

A tool that shifts effort from writing to checking has not necessarily paid off yet. That is the same measurement trap covered in how do you measure whether AI is actually paying off.

Why it matters

Regulated buyers are not asking whether your model is good. They are asking you to show your work per component, and to keep showing it after every change. Teams that treat compliance as a launch gate discover it is a build constraint; teams that build for it early, on Cleve’s account, occasionally find it closing deals instead. Where the data itself cannot leave the building, on-premise deployment is the adjacent question.

Disclosure: Episode 34 was sponsored by Galileo while Conor worked there; he refers to Galileo’s approach to custom benchmarks and metrics at 33:48.

Shipping AI in a regulated industry: show your work, per component Corti CEO Andreas Cleve describes the AI pipeline as a factory where every component has its own benchmark: was the data that came in good, was it understood, did it translate well, was it transcribed well. For each component you must be able to explain what it did, why it changed, when, and what went in. At the end, a clinician has to approve the output, which gives a golden label on whether it is complete and concise, labels that can feed an evaluation set once the review criteria are recorded. Weak evidence: one score on a public medical leaderboard, where by his account the leaders are noncommercial models, probably trained to test well rather than work in the wild. What replaces it: buyers want benchmarks per component, shown again after every change, and Corti adds models as judges and a large stock of holdouts. Show your work, per component Andreas Cleve, CEO of Corti, episode 34: the pipeline as a factory FOR EVERY PART, EXPLAIN what it did · why it changed · when · what went into it Data in benchmarked Understood? benchmarked Translated? benchmarked Transcribed? benchmarked CLINICIAN SIGNS OFF A golden label: complete? concise? "what you did was cool" labels that can feed an evaluation set WEAK EVIDENCE One leaderboard score On the big public health datasets, "the leaders are noncommercial models." His guess, hedged: probably trained to be "really good at tests but not very good in the wild." WHAT REPLACES IT Layered checking Buyers want benchmarks per component, shown again after every change. Corti adds models as judges on its reasoning pipeline and a large stock of holdouts.
Corti CEO Andreas Cleve’s factory picture from episode 34: benchmark every component, explain every change, and let the clinician’s required sign-off double as a label. The leaderboard point is his own hedged inference. Download the image

Hear it from the guest

“You can't win health care by being good at one thing. You have to be good at many things”
“So a doctor has to say, what you did was cool. So we actually at the end get a quite cool golden label of whether or not it's as complete or as concise as somebody would actually use it for.”
“We call this fiduciary-grade AI, where these professionals really have a duty to be correct.”
Watch on YouTube
“The thing you hear a lot about is super intelligence. I am more focused on super reliability.”
Watch on YouTube

Quotes lightly edited to remove filler words.

From the conversation

This explainer is drawn from these episodes — each carries its full transcript.

Concepts in this explainer

AI Impact AssessmentAI LiabilityAutomated Decision-MakingHigh-Risk AI SystemISO/IEC 42001NIST AI RMFEU AI ActHuman in the Loop