AI Smart Health Report
A generated health report people could actually read, that became a key USP in five enterprise deals.
- Role
- Product Analyst, HCL Healthcare
- Team
- Cross-functional, with engineering and design
- Scope
- Product requirements · UX flow · personalisation logic · cross-sell placement
15%
Incremental revenue
From cross-sell placed inside the report, not around it.
- 5+
- Enterprise closes
- 0→1
- Built from nothing
I owned
- The product requirements, the UX flow, and the personalisation logic that decides what an individual report says
- The decision that a health artifact would carry a commercial ask at all, and where that ask could defensibly sit
- Framing the report as something an enterprise buyer could point at, not only something a user reads
We shipped
- The AI Smart Health Report, taken 0→1
- Cross-sell hooks inside the report
I did not own
- The sale. The report was cited as a key USP in enterprise closes; I built the artifact, I did not close the deals.
- The clinical content standards. What is and is not safe to say about a result is not a call I would make alone.
A page of numbers nobody could read
People completed a health check-up and received their results: a column of clinical values, a column of reference ranges, and no way to tell which of it mattered.
“My haemoglobin is inside the range. Is that the same as fine?” is not a question a table answers. It is the question the product had to answer, and it is a product problem rather than a clinical one. The information was already correct and already delivered.
- InputCheck-up resultsClinical values a person cannot interpret on their own
- GeneratePlain-language interpretationPersonalised to the individual reading it
- ActA next step worth takingOne prompted action, not a list of twelve
- OutcomeEvidence the employer can point atThe artifact that makes the programme legible at renewal
Three people had to say yes, and only one of them reads it
This is the part of the problem I find genuinely interesting, and the part most consumer product work never contains.
The beneficiary
Can I understand my own result, and is there one thing worth doing about it?
The employer
Did the programme I bought produce something my people actually valued?
The approver
Does this create a liability I have to sign my name to?
The beneficiary reads the report. The employer buys the programme. And someone internal has to be willing to put their name on what it says. A report that delights the reader and worries the approver does not ship. One that satisfies the approver and bores the reader does not get read, which means the employer has nothing to point at either.
The decision I actually made
Personalisation was the mechanism, not the point. The point was that a report has to be worth finishing, and a generic report is not, so the logic decides what an individual sees based on their own results rather than showing everyone the same page with different numbers in it.
Then the commercial question. A health artifact carrying a commercial ask is a real trade-off and I want to be plain that I chose to make it, because the alternative, a beautiful report the business cannot justify building again next year, is not obviously the more ethical option.
That line is a layout, not a copy guideline, which is the whole reason it holds. Here is the hierarchy it produces:
Annual health check · March
Your results, explained
| Test | Result | Usual range |
|---|---|---|
| Haemoglobin | 14.2 g/dL | 13.0–17.0 |
| HbA1c | 5.9 % | 4.0–5.6 |
| LDL cholesterol | 96 mg/dL | 0–100 |
| Vitamin D | 22 ng/mL | 30–100 |
What this means
Two results sit outside their usual range. Your HbA1c of 5.9% is above 5.6, which is the figure a doctor would want to look at alongside your history. It is not a diagnosis on its own. Your vitamin D of 22 is below 30, which is common and usually straightforward to correct.
Everything else on this page is inside its usual range.
Next step
Book a 15-minute call with a doctor to go through these two results.
Book a consultationRead the order. The value, then what it means in a sentence a non-clinical adult can finish, then a rule, then a different ground, and only there, the ask. The next step is a consultation with a doctor, which is what an out-of-range HbA1c actually warrants. If that block had offered a supplement instead, the report would have become a funnel with a stethoscope on, and no amount of careful wording would have fixed it.
Layout and hierarchy accurate; every name, number and date is invented. No real user appears here.
The constraint is structural: the commercial block cannot appear inside the interpretation block, so it cannot become the answer to a worrying result even if someone later rewrites the copy.
What happened
I took the AI Smart Health Report from nothing to shipped: the requirements, the flow, the personalisation logic and the cross-sell placement.
Those hooks drove 15% incremental revenue, and the report was cited as a key USP in five or more enterprise closes. I built the artifact; I did not own the sale. What it demonstrably did was give a salesperson something concrete to put in front of a buyer, in a category where most of what gets promised is a programme rather than a thing.
What I would do differently
I would build the evaluation before the feature, not after. The first question anyone asks about a generated health report is how it avoids giving medical advice, and the honest answer is that I designed the product surface and not the enforcement layer underneath it.
So I went and built what I think that answer should look like. Grounded scores generated health text against a visible rubric. Whether every claim traces to a source value, whether it crosses into diagnosis or dosage, whether an out-of-range value routes to a clinician, and whether a non-clinical adult can read it. It runs in your browser with no server and no key.
That is the version of this case study that should have existed the first time.
How I worked this out
Where a commercial ask can and cannot sit inside a clinical artifact
The placement that converts best is the placement that costs the most trust, and trust is the entire asset. A report nobody believes is worth nothing to the reader and nothing to the employer who bought the programme. The version I would defend puts the interpretation and the ask in visibly different registers, keeps the ask out of the sentence that explains a result, and never lets a commercial prompt be the answer to a concerning value. That is a design constraint, not a copy guideline: if the boundary depends on whoever writes the next line, it is not a boundary.
The question an AI PM will ask about this, and where my answer stops
The first question anyone will ask about a generated health report is how it avoids giving medical advice. It is the right question. My published record covers what I designed (the requirements, the flow, the personalisation logic and where the commercial ask sits) and not the enforcement layer underneath it. What I can show is what a defensible answer looks like, which is why I built Grounded: an evaluator that scores generated health text against a visible rubric for grounding, scope, escalation and readability. That is in the Lab, and it runs in your browser.
Results
Measured outcomes
15%
Incremental revenue
From cross-sell placed inside the report, not around it.
5+
Enterprise closes
Where the report was cited as a key USP. I built the artifact; I did not own the sale.
0→1
Built from nothing
Requirements, UX flow and personalisation logic.
Next case study
Step SyncingA fifteen-second launch cut to under two seconds, and step-sync completion up 35%.