Executive summary
Huna had a rare technical asset: a proprietary machine learning model, inherited from an already-validated research line, capable of identifying cancer risk from routine blood tests. It also had the classic problem faced by anyone selling AI to healthcare in Brazil: the asset wasn't turning into a business.
I joined as AI Lead Product Manager to define the strategy from scratch. Discovery with health insurers overturned the premise the company was operating on. The obstacle to adoption wasn't model accuracy, it was that the data needed to run it didn't exist in a usable format inside the client. That shifted the entire strategy: from selling prediction to selling the capability to extract, structure, and act on the data the insurer already had but couldn't use.
Three decisions came out of this repositioning: couple instead of replace, sell to whoever pays rather than whoever feels the pain, and treat clinical oversight as a trust strategy rather than a technical limitation. The result was a quantifiable business case and two pilots with insurers. The product never launched, and the final section explains where it stopped and why.
The premise discovery overturned
Healthtechs selling AI to insurers usually pitch the wrong challenge. The company believes it competes on accuracy. The buyer isn't evaluating accuracy, it's evaluating whether it can actually operate the thing.
Throughout 2023 I ran continuous discovery with health insurers. The same pattern repeated across all of them: even at a sophisticated insurer, around 20% of hospital admissions had no structured data. Reports in PDF, closed medical records, clinical history locked in a legacy system nobody wants to touch. It's not a lack of data, it's a lack of usable structure.
The strategic consequence is harsh for whoever built the model: the algorithm's predictive power is irrelevant to a client who can't feed it. Whoever solves the extraction and integration bottleneck gets there before whoever optimizes one more point of accuracy. That reading, not the quality of the model, is what came to define the roadmap and the sales narrative.
The three strategic decisions
1. Couple, don't replace
Insurers don't want to swap out their medical records system, they want to add a capability on top of what they already use. I designed the entire integration proposal as a complementary layer on top of existing systems (medical records, app, and dashboards) instead of asking for a replacement.
The trade-off was real. Replacement gives you more control over the experience, more product surface, and a bigger contract. It's also the proposal that stalls B2B sales in sectors with entrenched legacy systems, because it turns a purchasing decision into a migration project, with a different sponsor, a different budget, and a different risk horizon. I chose the smaller contract with a sales cycle that actually closes.
2. The client is whoever decides and pays, not whoever feels the pain
The ultimate beneficiary was the patient with delayed screening, around 5,000 women in just one of the pilot insurers' base. But the patient doesn't buy screening software, and the clinical team using the model's signal day to day doesn't sign the contract either.
The product was designed for three different readers at once: whoever decides the purchase, and needs a business case; whoever uses the signal, and needs a workflow that doesn't add work; and whoever audits, and needs traceability of the decision. Treating this as a single user would have produced an interface that pleases the clinical team and doesn't close a sale, or the reverse.
3. Clinical oversight as strategy, not limitation
I structured hybrid workflows where the model flags and prioritizes risk, but the final clinical decision stays with the health professional, by product decision and not by implementation accident.
There was an easier path: automate more, build faster, and sell it as "AI that decides." In oncology screening, a mistake has a consequence that isn't fixed in the next release, and a sector that already distrusts black boxes, rightly, doesn't buy opacity. Leaving the decision with the clinician stopped being a constraint and became the sales argument, provided the interface could communicate confidence, uncertainty, and the algorithm's limits to whoever was going to act on that signal. In this context, the "how much to trust it" interface matters as much as accuracy.
The business case that made the conversation possible
A repositioning without a number is just an opinion. The conversation with the insurer only moved past general interest once it was compared against the process it already runs.
The calculation started from a number the insurer itself provided, gathered before any pilot ran: around 5,000 women in its base had overdue mammograms. Based on the monthly cost of screening to clear that list, model-supported screening represented an estimated saving of R$315k versus the insurer's standard process.
Building the argument from the client's own data, not ours, is what changed the conversation. Not "our model has high accuracy," but "this list, which is yours, costs you this much more under your current process to reach the same place." The buyer can take that inside the organization; accuracy, they can't.
In parallel, I worked on building a data partnership with Fleury, with a mapped base of around 200,000 women, which was still under discussion by the end of my time at the company. Strategically, this partnership attacked exactly the bottleneck diagnosed above: access to already-structured data, at volume, without depending on each insurer's integration maturity.
What the strategy required in practice
On a lean, early-stage team, strategy doesn't separate from execution. Besides defining the positioning, I was responsible for full prototyping and UX of the platform, from the internal triage flow to how a risk result reached whoever needed to act on it. There was no dedicated designer, and treating "product decision" and "interface decision" as separate things would have broken exactly the third decision above.
I ran two pilots with health insurers to test and validate the breast cancer screening support algorithm, with joint technical and clinical validation. I advanced discovery further in October 2023 to assess expanding the model to other cancer types, mapping the limitations of the screening methods then available, such as the high cost and low specificity of PSA testing. I also opened and maintained conversations with labs, insurers, and cancer institutes.
I led Huna's participation in InovaHUPE 2023, the acceleration program run by Hospital Universitário Pedro Ernesto, part of UERJ, whose edition that year focused on Technological Innovations for Oncology. Among six finalist initiatives, Huna was selected as one of the top three at the final ceremony, with its solution supporting breast cancer diagnosis from routine tests. The program included one-on-one mentoring and biweekly workshops, plus clinical mentoring with the hospital's medical staff. Announcement on Huna's site ↗
Where it stopped, and the strategic reading of that
The product never launched. The barrier wasn't commercial or positioning-related: it was that the necessary technical validation, with prospective and retrospective studies, wasn't completed within the timeframe. In a clinical product, that gate isn't negotiable, and no amount of commercial traction replaces it.
The takeaway I carry from this is about sequencing. The repositioning solved the adoption problem and produced a business case, pilots, and a meaningful data partnership. What it couldn't solve, and what I couldn't accelerate by product decision, was the maturity of the clinical evidence. In AI healthtech, market discovery and scientific validation discovery run at different paces, and when the second one is the bottleneck, commercial traction turns into debt: it creates expectations in a client that the evidence can't yet support.
What I take into other contexts
- The AI bottleneck is rarely the model. It's the extraction and structuring of data that lives in PDFs, closed medical records, and spreadsheets. Diagnosing that early changes the product you build, not just the sales pitch.
- Coupling beats replacement. In entrenched legacy systems, the proposal that adds closes deals; the one that asks for a full swap turns into a migration project and dies at the door.
- Whoever feels the pain and whoever signs the contract are rarely the same person. In healthcare B2B, designing for one and forgetting the other produces a product people love that doesn't sell, or a product that sells that nobody uses.
- A well-chosen constraint becomes positioning. Keeping the decision with the clinician was the slower path; it's what made the product credible in a sector that distrusts opaque automation.
- Commercial traction doesn't get ahead of evidence. When the real bottleneck is scientific validation, speeding up the funnel just increases the liability.
Tools and methods used
- HubSpot, for partnership and sales pipeline
- Miro, for product backlog and opportunity tree
- Continuous discovery and personas, to guide prioritization
Note on the data: the R$315k saving in the Unimed VTRP pilot is a comparative estimate against the insurer's standard screening process, not an audited figure. Base and insurer numbers were handled under confidentiality.