Foreword

Why this book, why now

A weekend before a contract

Picture this: a founder in a seed-stage medtech writes me on a Friday afternoon. They have a vendor proposal on the table for 180,000 euros to build an AI module for their next clinical study. The vendor demo looked clean. The model “works” on the cohort they were shown. Monday is the contract signing. Can I have a quick look over the weekend?

Two hours into the proposal, I count seven questions the founder has not asked. By Sunday night, the deal is paused. By the end of the next month, the entire scope is rewritten, the budget cut in half, and one of the original POC milestones is dropped as unbuildable with the data they actually have. No one is angry. Everyone is relieved.

This book exists because the same conversation happens, in different forms, three to five times a month.

The landscape, and what is missing

The anti-hype literature on AI is good. Cathy O’Neil, Arvind Narayanan, Bob Wachter, Eric Topol have all written books that anyone serious about AI in medicine should read. They cover bias, snake oil, deployment failures, clinical context, and the social politics of automation. They are right about most things.

What they do not cover is the founder of a fifteen-person biotech in 2026, with a Series A runway, a vendor email open in another tab, and a board meeting on Thursday. That founder cannot read a 400-page critique and translate it into a contract clause. They cannot afford a Big Four advisory engagement. They cannot wait six months for academic consensus. They need a decision framework, today, written by someone who would also be willing to build the system if asked.

That is the gap this book tries to fill.

Why me

I am Saber Graf. I am Pharmacist by training (PharmD, Université de Constantine), neuroscientist by PhD (Université de Bordeaux, defended December 2024 on self-supervised learning applied to neural signal classification), engineer by daily practice. For more than twelve years I have worked at the intersection of biological complexity, signal processing, and software systems, in early-stage biotech and CRO environments where the data is messy, the deadlines are real, and there is no army of analysts behind you.

In 2025, my work on self-supervised learning to correct label noise in electrophysiology was published in Scientific Reports (Graf et al., 2025; DOI 10.1038/s41598-025-90380-x). The method moved a classifier from 73% to 84% accuracy on a notoriously noisy dataset, validated across two cross-domain benchmarks. That paper is part of why I felt I had standing to write this book. Not because publication is a credential, but because the work itself required me to live inside the contradictions that healthtech AI projects keep producing: confident models on dirty data, metrics that flatter, ground truth that is not really ground.

Outside of my own research, I architect and maintain a Django + Next.js + Azure platform for a CRO every day. I write the code I recommend. Under the name SG AI Solutions, I audit healthtech AI projects for founders and operators in biotech, medtech, and digital health. I have learned that most of what kills these projects is not the models. It is the questions that were not asked at the start.

This book is the seven questions, plus what comes after them.

For whom, not for whom

This book is written for founders, CSOs, CTOs, CMOs, and Heads of Data or R&D in biotech, medtech, and digital health at seed to Series A. Fifteen to eighty people. Two to fifteen million euros raised. A concrete AI project on the table, or coming in the next six months. Pharma R&D early-stage decision-makers are welcome too.

It is not written for academics looking for a literature review. It is not written for large pharma corporate buyers who already have an in-house AI division and a procurement process designed for nine-figure deals. It is not written for ML engineers who want a tutorial. And it is not written for the AI-skeptical reader who wants confirmation that none of this works. Some of it works very well, when scoped correctly. That is the entire point.

What this book is, what it is not

This book is a decision guide. It is structured around the choices a founder actually makes: whether to build, buy, or partner, how to read a vendor proposal without being deceived, what to look for in a team, what to expect in production, and where the regulation is heading.

It is not a technical manual. There are no Python snippets, no PyTorch tutorials, no architecture deep dives. Better books exist for that.

It is not legal advice. Where the EU AI Act, the FDA Predetermined Change Control Plan, or the EHDS rules matter, I flag them and refer you to your lawyer.

It is not a prediction book. I am not going to tell you what will happen in 2030. I am going to tell you what to do this quarter, and how to recognize a few patterns that have been quietly killing healthtech AI projects for the last five years.

How to read it

The book has three parts: Demystify, Decide, Real Cases. If you are already comfortable with what AI can and cannot do, skip Part 1. If you are evaluating a specific vendor or POC right now, go straight to Chapter 4 (the seven questions) and Chapter 6 (reading proposals). If you want stories from the field first, start with Part 3.

Total reading time, end to end, is around four hours. Each chapter is self-contained. You can read them in any order. There is no chapter dependency that will break the rest of the book if you skip it.

Before you sign

If you have a POC contract sitting on your desk this week, do not sign it before reading Chapter 3. That is not a sales line, it is the cheapest thing this book can do for you. The seven questions in Chapter 4 will tell you whether the project is real or theatre, and Chapter 6 will help you read the proposal as someone who could also build the thing.

A final note on humility. I will be wrong on some predictions in this book. Some vendors I treat skeptically will turn out to ship excellent products. Some technical patterns I praise will age badly. That is the cost of writing about a moving field in motion. Tell me where I got it wrong. The next edition will be better.

Saber Graf Aix-en-Provence, June 2026