Everybody Wants to Be an AI-First Health System. Almost Nobody Wants to Redesign the Work.

Years ago I watched a Stripe dashboard cross a million dollars in a single day. My company. My name on it. I remember refreshing it like an idiot, and I remember the specific flavour of the feeling, which was not joy. It was something closer to vertigo, because I knew the product wasn't good. We'd found a channel that worked and poured petrol on it, and the volume was hiding every broken thing underneath.

Scale doesn't fix problems. Scale hides them, right up until it can't.

I think about that constantly now when I read health system AI strategies, because BCG's argument that most provider AI efforts stall for structural rather than technical reasons is correct and also slightly too polite. Providers keep layering intelligence onto workflows that were designed for a fax machine. Then they're confused when the pilot doesn't scale.

What does AI-first actually mean for a health system?

It means the work gets redesigned, not decorated. In an AI-first model the default assumption flips: routine coordination, information synthesis and follow-through happen without a human in the loop, and your clinical staff spend their day on exceptions and judgment. That's the whole thing. Everything else is vocabulary.

Most organisations do the opposite. They keep the existing workflow intact and bolt an AI onto the end of it, which produces a tool nobody's job depends on. Those tools die in pilot. Every time.

Why do most health system AI programs stall?

Because nobody was asked to give anything up. A pilot that adds a step, adds a login, adds a dashboard someone has to check, is a pilot that competes with clinical work for attention and loses. The programs that survive are the ones where a task genuinely left the building.

Labor is 50% to 70% of provider cost structure and rising, demand is climbing, and pricing power is basically fixed. You cannot incrementalise your way out of that arithmetic. Which means the only interesting question is which categories of work you're willing to stop doing manually, and that's a leadership question, not an IT one.

I had a nervous breakdown in a meeting once. Crying, couldn't stop, brain completely fried after eighteen months of holding too much. The body keeps score, and so do care teams. When you tell a nursing staff that's already at capacity that the AI initiative is additive, you're not being encouraging. You're being expensive.

Which workflow should you automate first?

The one that's high volume, low judgment, and currently done badly because nobody has time. Post-discharge follow-up. Transitional care outreach. Chronic care check-ins. Pre-procedure preparation. Recall for lapsed panels. The specific use cases that carry the most trapped value are almost never the ones that photograph well.

Diagnostics get the headlines. Coordination gets the money. Boring, repetitive, relentless follow-through is exactly what software is good at and exactly what human staff burn out doing.

I crossed the Australian desert with a circus once, which is a longer story than this post has room for, and the thing I took from it was that the acts drawing the biggest crowds were never the most technically difficult. Fire chains beat aerial silk, consistently. Accessible beats impressive. That's a procurement principle, honestly.

What does the evidence say about automating follow-up?

It says the delivery model matters enormously, and that's easy to miss when everything gets lumped together as "digital health."

A multi-site study across nine hospitals found that virtual nursing assisted discharges cut 30-day ED readmissions from 13.3% to 3.7%, with comparable baseline risk after matching. Big effect. Meanwhile a randomized trial of remote monitoring after sepsis and respiratory infection across 19 hospitals found no improvement in days at home, and a worse result for patients over 65.

Both are "automation." One redesigned who does the discharge conversation. The other added a device and a questionnaire to a workflow that stayed the same, and only 60% of assigned patients even enrolled.

That's the whole lesson in two studies. Automation that absorbs the work produces outcomes. Automation that requests patient effort produces enrollment charts. Our own outcomes data across more than a million interactions points the same direction: 85% weekly engagement against a 15% to 20% baseline, because the system reaches out rather than waiting to be opened.

How do you buy AI infrastructure you won't regret in three years?

Ask three questions and watch people squirm.

Where does inference run, and who else sees the data. If the answer routes through a third-party model API, you've made a permanent dependency decision on someone else's roadmap and pricing. HANA is open-source and self-hosted with no OpenAI dependency, and that's a deliberate architectural choice rather than a marketing one.

Can you leave. If the workflow logic, prompts and integrations live in a vendor's black box, you're not buying capability, you're renting a hostage situation.

Does it survive contact with your EHR. Not "do you integrate," everyone says yes. Look at the actual integration and deployment documentation and see whether an engineer on your side could read it and know what happens.

Then look at unit economics rather than platform fees. The pricing model should let you compute cost per completed patient interaction, because that's the number that scales with you. We're seeing 31:1 return in deployments where the automated workflow sat next to revenue, across five countries and three languages.

Key Takeaways

AI-first is an operating model decision, not a technology purchase, and the tell is whether any work actually left your organisation. Programs stall because they're additive to already saturated clinical teams, so the first question isn't what to build but what to stop doing manually. The evidence separates cleanly along one line: automation that absorbs effort from patients and staff produces real outcome movement, while automation that requests effort produces enrollment metrics and disappointment. And on infrastructure, self-hosted and open beats black-box convenience by year three, because the constraint you'll feel isn't capability, it's dependency.

FAQ

Do we need an AI strategy before we start, or can we pilot our way there?

Pilot, but pilot in a workflow you'd be willing to run permanently. The failure mode isn't starting small, it's starting in a place where success wouldn't change how anyone works.

Won't clinical staff resist automated patient outreach?

Less than you'd think, if the automation removes work rather than adding oversight. Resistance almost always tracks who has to check a new dashboard. Staff who get a hundred routine calls off their plate become advocates fast.

How do we evaluate vendors when everyone claims the same outcomes?

Ask for weekly engagement among enrolled patients at week four, not at week one. Ask where inference runs. Ask what happens to your workflow logic if you terminate. Three questions and the field narrows quickly.

If you're mapping out the first workflow to hand over and want a second opinion from someone who has broken this a few times, book a slot and let's go through it.