6 HUMAN IN THE CORE
The book delivers on it in three movements. The first three chapters establish
the foundation: what AI's category-of-event shift means for a regulated enter-
prise, how the major regulators have already converged on the architectural
commitments that follow, and how this analysis groups those commitments
into five categories. The next five chapters walk the categories (data, thread, in-
telligence, team, and process), naming the responsibilities, the controls, the
failure modes the regulators have already cited, and the convergent witness the
advisory firms provide. The final two chapters are crosscutting. Chapter 9 lifts
the categories that have assembled, and treats mutual validation, lifecycle gov-
ernance, and cybersecurity posture as obligations that run across every layer.
Chapter 10 closes by addressing the reader's first 365 days inside the analysis,
with reading paths by role and a calendar from Day Zero forward.
1.3 The Stakes and the Imperative
The opportunity is real. McKinsey's life sciences analysis suggests that up to
95% of life sciences roles may have agentic teammates, and that two-thirds of those
roles will involve building, managing, and supervising agents. That is not a small re-
organization. It is the second-largest workforce reshaping in the industry's his-
tory, after the move to global outsourcing of clinical and manufacturing opera-
tions. The pattern repeats, with different artifacts, in every regulated vertical.
The companies that get this right will run their pipelines faster, write their sub-
missions cleaner, monitor their post-market signals earlier, and operate their
plants with fewer deviations. They will also, and this matters, be better places
to work. The routine cognitive load that exhausts scientists and clinicians will
move to systems built for it, and the humans will be freed to do what only hu-
mans can do.
The risk is also real, and Gartner has named it sharply: over 40% of agentic AI
projects will be cancelled by the end of 2027, due to escalating costs, unclear business
value, or inadequate risk controls. The same forecast that gives us the opportunity
tells us that nearly half of the attempts will fail. Some of those failures will be
benign. Abandoned pilots, sunk-cost write-offs, embarrassed sponsors. Some
will not. In a regulated industry, an inadequately controlled AI project does not
just fail commercially. It produces warning letters, triggers consent decrees,
causes recalls, loses trial sites. In the worst cases, it harms the patient at the
end of the chain. And in this way, perhaps the projects that don't get cancelled
could do the most harm.
There is a middle path between the 95% opportunity and the 40% failure rate,
and it is not a strategy of caution. The companies that will win this transforma-
tion are not the ones that move slowly. They are the ones that move quickly with
The book delivers on it in three movements. The first three chapters establish
the foundation: what AI's category-of-event shift means for a regulated enter-
prise, how the major regulators have already converged on the architectural
commitments that follow, and how this analysis groups those commitments
into five categories. The next five chapters walk the categories (data, thread, in-
telligence, team, and process), naming the responsibilities, the controls, the
failure modes the regulators have already cited, and the convergent witness the
advisory firms provide. The final two chapters are crosscutting. Chapter 9 lifts
the categories that have assembled, and treats mutual validation, lifecycle gov-
ernance, and cybersecurity posture as obligations that run across every layer.
Chapter 10 closes by addressing the reader's first 365 days inside the analysis,
with reading paths by role and a calendar from Day Zero forward.
1.3 The Stakes and the Imperative
The opportunity is real. McKinsey's life sciences analysis suggests that up to
95% of life sciences roles may have agentic teammates, and that two-thirds of those
roles will involve building, managing, and supervising agents. That is not a small re-
organization. It is the second-largest workforce reshaping in the industry's his-
tory, after the move to global outsourcing of clinical and manufacturing opera-
tions. The pattern repeats, with different artifacts, in every regulated vertical.
The companies that get this right will run their pipelines faster, write their sub-
missions cleaner, monitor their post-market signals earlier, and operate their
plants with fewer deviations. They will also, and this matters, be better places
to work. The routine cognitive load that exhausts scientists and clinicians will
move to systems built for it, and the humans will be freed to do what only hu-
mans can do.
The risk is also real, and Gartner has named it sharply: over 40% of agentic AI
projects will be cancelled by the end of 2027, due to escalating costs, unclear business
value, or inadequate risk controls. The same forecast that gives us the opportunity
tells us that nearly half of the attempts will fail. Some of those failures will be
benign. Abandoned pilots, sunk-cost write-offs, embarrassed sponsors. Some
will not. In a regulated industry, an inadequately controlled AI project does not
just fail commercially. It produces warning letters, triggers consent decrees,
causes recalls, loses trial sites. In the worst cases, it harms the patient at the
end of the chain. And in this way, perhaps the projects that don't get cancelled
could do the most harm.
There is a middle path between the 95% opportunity and the 40% failure rate,
and it is not a strategy of caution. The companies that will win this transforma-
tion are not the ones that move slowly. They are the ones that move quickly with