Everyone's watching the clinic. The biggest AI wins in healthcare are showing up somewhere nobody's covering.
Ask someone where AI is changing healthcare and they'll say diagnosis, drug discovery, decision support at the bedside. That's where the ambition sits, and it's not wrong to be excited about it.
It's just not where the results are yet.
Look across roughly thirty AI deployments in active use across pharma, biotech, medical devices, and health plans, and the value clusters somewhere far less interesting to talk about: verification datasets, audit evidence, traceability, test coverage. The paperwork. The reason is simple once you see it – this is text-heavy, rule-governed, high-volume work with a defined structure and a human expert on the other end who can review a finished draft far faster than they could ever write one. That's exactly the job current AI is built for.
What the data shows:
- Value concentrates in regulated evidence work, not clinical decisions – the reverse of where the industry's attention is.
- Thirty separately funded use cases turn out to be four repeatable mechanisms wearing different names.
- Reported gains cluster at 50-90% effort reduction, and that's the wrong number to be leading with, which is a big part of why these programs stall after pilot.
- The largest untapped opportunity is in the clinical and regulatory document chain – needs no new invention, just a mechanism these same organizations have already proven works.
- The one capability every other capability depends on, assurance of the AI systems themselves; is also the least built.
Findings drawn from an aggregate analysis of AI use cases across multiple HCLS organizations. No individual program is identified; figures are portfolio-level.
The same four problems keep showing up
Diagnostic instruments, therapeutics, health plans – different businesses, same friction, every time.
Experts are hand-building evidence that doesn't need a human hand
Trace matrices, verification datasets, test protocols, audit packs – assembled manually, often mapping thousands of requirements against thousands of specifications. It's not that this takes effort. It's that it can only be done by the small number of people qualified to do it, so every program queues behind the same handful of experts.
People are doing the job of an integration layer
An engineer manually reconciles five monitoring platforms. A quality lead builds a picture of complaints, deviations, and supplier risk by hand-crossing four separate systems. The logic that connects it all lives in someone's head, takes hours to run, and offers no guarantee the important signal got seen.
Testing effort follows habit, not evidence
In this portfolio, 12% of test cases were near duplicates. Some functional areas were tested heavily and produced no defects; others carried real defects with no coverage at all. And two-thirds of logged defects resolved to "not a bug" or "can't reproduce”; meaning most of the effort went into proving there was nothing wrong.
Everyone's waiting for someone else
New hires wait on expert time to get up to speed. Developers file tickets to discover interfaces that already exist. It's the same bottleneck each time: another person's attention, and it doesn't show up on any project plan.
None of this is unique to healthcare. What makes it a healthcare problem is that these are regulated artifacts – so a mistake here isn't expensive. It's disqualifying.
Thirty projects. Four machines.
By name, this portfolio has thirty distinct initiatives – thirty sponsors, thirty business cases. By function, it has four.
The same artifact-generation engine was independently built six separate times under six separate names, because each of six different teams needed a slightly different output from it. A trace matrix here, audit evidence there. Six invoices for one machine.
That's not a story about wasteful teams. It's a story about funding the wrong unit. Pay for use cases, and every business owner buys their own bespoke version of a general capability – the organization pays the integration cost over and over and never ends up with a platform. Pay for the mechanism instead, and the artifact type becomes a configuration choice, not a rebuild.
Why the efficiency pitch keeps failing
Nearly every deployment here leads with an effort-reduction number – 50 to 90%. The number's real. It's just the wrong argument for this sector.
Cost savings compete for budget against every other cost-saving idea in the building, and most of those have shorter payback and less technical risk. A complete, auditable trace matrix doesn't compete for budget, it pre-empts a regulator's question. Same deployment, same underlying work, but described as an efficiency gain it fights for airtime; described as closed exposure, it doesn't.
Worth being honest about the limits here too: none of this is flawless. A generative system in this portfolio reached 98% completion on a verification dataset, not 100%, with hallucination as the documented cause. That's not a failure – it's a design constraint, and it's exactly why the working version of this mechanism keeps a human reviewing the finished draft rather than trusting the model to sign its own work. That's not a caveat bolted onto the value proposition. It's the reason a regulator will accept any of this at all.
What nobody's built yet
Two gaps stand out, and both are within reach – because the mechanism to close them already exists inside these same organizations.
The clinical and regulatory document chain
Protocol authoring, site selection, clinical study reports, pharmacovigilance triage, submission assembly – all high-volume regulated text with a defined structure and a regulator downstream. It's the exact same shape of problem the software and quality teams have already solved. What's stopped it isn't technical difficulty. It's that a different sponsor, budget line, and risk committee sit between a working engine and the regulatory affairs team that needs it.
Assurance of the AI itself
Any generative system inside a regulated workflow becomes a system that needs validating – model qualification, drift monitoring, error-rate bounds, auditability. As AI-specific regulatory guidance matures, this stops being nice to have and becomes the license to operate. It's also the least developed capability in the entire portfolio: mostly stated intent, rarely deployed practice. That's the sector's real risk, the one capability everything else depends on is the one nobody's finished.
A few smaller gaps are worth naming too: manufacturing quality is automated at the reporting layer but not the investigation layer, payer operations remain largely untouched even in organizations built around them, and computer vision – in a sector defined by imaging – is nearly absent.
The takeaway
Fund the mechanism, not the use case – thirty initiatives paying for the same integration thirty times is the actual waste here. Make the case in terms of exposure closed, not hours saved. And build assurance before you scale, not after, because that's the constraint that will bind exactly when you try to move from pilot to production.
The clinic still gets the headlines. The evidence chain – quieter, less glamorous – is where the value already is.