Your AI Strategy Is a Procurement List, Not a Redesign

Years ago I watched a Stripe dashboard cross a million dollars in a single day.

I remember where I was standing. I remember refreshing it like an idiot. What I don't remember is anybody in that room asking whether the product was actually good, because it wasn't, and we found out eighteen months later in the way you always find out, which is all at once.

Scale doesn't fix a broken design. It funds it for a while and then presents the bill.

I think about that a lot when health system executives walk me through their AI strategy, because the strategy is almost always a procurement list. Ambient scribe, imaging triage, a chatbot on the website, a pilot in one service line with a champion who leaves within a year.

Why doesn't buying AI tools change anything?

Because the tools sit on top of an operating model built around the scarcity of clinician time, and they never touch it. BCG made this argument better than I can in AI Won't Fix Your Health System. Redesigning It Will., and the line that stuck with me is that the most valuable AI in healthcare won't always have a human in the loop.

That's uncomfortable and it's correct. Every advisory tool you deploy still terminates in a person who has to do the thing. Add ten advisory tools and you've built ten new inboxes for the same forty clinicians.

Capacity doesn't come from better suggestions. It comes from work leaving the queue entirely.

What does redesign actually look like?

It looks boring, which is the good news. The most convincing recent example isn't even generative AI. A multi-site study in npj Digital Medicine looked at nine hospitals that moved discharge tasks off bedside nurses onto remotely located virtual nurses. Matched cohorts, 4,662 encounters on each side. Thirty day ED readmissions came in at 3.7% for virtual nurse assisted discharges versus 13.3% for traditional ones.

Nobody invented a new clinical insight there. They changed where the work happened and who did it. Then the outcome moved.

That's redesign. Everything else is a plugin.

Which work should leave the queue first?

Start with work that is high volume, protocol-shaped, and currently being rationed. In almost every system I've looked at, that's outbound patient contact: follow-up after discharge, pre-op prep, medication adherence, recall for care gaps, all the calls your coordinators would make if the day had thirty hours in it.

It gets rationed silently. Nobody writes "we contact 18% of our post-discharge patients" on a board slide. But that's the number, and it quietly caps every downstream metric you do report.

Hand that entire lane to a system that runs it end to end and escalates only the exceptions. Not a suggestion engine. The lane. Our case studies are all versions of the same move.

What breaks when you do this at system scale?

Governance, mostly, and that's fair. Once software talks to patients without a clinician mediating every sentence, you have to answer for what it said, in which language, to whom, and what happened next.

Which is why HANA is fully open-source and self-hostable, running inside your infrastructure, with no hard dependency on a single US model vendor. Your compliance team can read the prompts. Your CISO can see where the audio goes. Across more than a million interactions in five countries we've had zero critical adverse events, and the research page has the boring detail on how that gets measured, because "trust us" isn't a governance model.

The systems moving fastest here aren't the ones with the biggest risk appetite. They're the ones that made the risk legible.

What's the real competitive question?

Easier access pulls more patients in. BCG's point about always-on access increasing pressure on who delivers care and where is the part most executives skip past, because it's the expensive part. Make it trivially easy for a patient to be contacted and to respond, and more of them will surface real problems, and those problems land on your capacity.

So the honest sequence is: automate the contact lane, watch demand rise, then redesign who absorbs it. Doing step one without planning for step three is how good pilots die quietly in year two.

We built our pricing around that reality rather than per seat, because per seat licensing punishes exactly the behaviour you're trying to create.

Key Takeaways

AI procurement isn't AI strategy. Tools that advise clinicians leave the operating model untouched, and the operating model is the constraint. The interventions that actually moved outcomes recently moved work, not information: virtual nursing cut 30-day ED readmissions from 13.3% to 3.7% across nine hospitals by relocating discharge tasks. For most systems, the first lane worth handing over completely is outbound patient contact, because it's high volume, protocol-shaped, and already rationed in silence. Do it on infrastructure you can inspect and host yourself, and plan for the demand that easier access will surface.

FAQ

What's the difference between AI augmentation and AI redesign in healthcare?

Augmentation gives a clinician a better input and leaves the workflow intact, so capacity stays roughly fixed. Redesign removes a category of work from the human queue and rebuilds the process around that, which is where capacity gains actually come from. Most health systems have bought a lot of the first and very little of the second.

Which clinical workflows are safest to automate end to end first?

Protocol-driven, non-diagnostic, high-volume contact work: post-discharge follow-up, pre-procedure preparation, medication adherence checks, and care-gap recall. These have clear scripts, obvious escalation triggers, and are already under-delivered, so the counterfactual is silence rather than a clinician. Our use cases page breaks them down by specialty.

Why does self-hosting matter for health system AI?

Because governance requires inspection. Self-hosted, open-source infrastructure lets your compliance and security teams read what the system says to patients, control where protected health information travels, and avoid a hard dependency on one model vendor's pricing and policy changes. If you want to see how that deployment works in a hospital environment, grab a slot and I'll show you.