Dr. Desmond Daley
Why Healthcare AI Efforts Stall | Dr. Desmond Daley
Move beyond AI pilot projects to real healthcare transformation with this practical framework for implementation.
All articles
By Desmond Daley
Hospitals do not feel like "AI."
They feel like a nurse walking the unit, again, looking for a missing order that "should be in the system."
They feel like a resident re-entering the same medication list into a second screen because the first one "doesn't count."
They feel like a clinician staring at an alert that might be right, but cannot be trusted, because nobody can explain where it came from.
Then you step into a ballroom where the chorus never changes: AI, machine learning, predictive analytics.
That gap, the floor versus the stage, is why so many healthcare AI projects die. Not with a headline. With a calendar invite that stops getting sent. A slow drift into "we'll revisit this next quarter."
Here is the uncomfortable truth: AI does not transform healthcare. People do.
And people do not change behavior because you bought software. They change when a tool fits the work, earns trust under pressure, and gives time back without stealing judgment.
If that sounds less glamorous than the hype, good. Glamour does not move patients through a system safely.
Name the monster: pilot purgatory
Most healthcare AI initiatives do not fail because the models are weak. They fail because the system mounts an immune response.
Healthcare is a high-stakes, interruption-heavy environment. The default state is overload. So when a new tool shows up that adds clicks, adds uncertainty, or adds one more place to look, the system does what it is designed to do. It rejects the foreign body.
The rejection looks polite on slides. In real life, it looks like:
usage spikes during go-live week, then flatlines
"training" becomes theater
workarounds become the real workflow
a dashboard stays alive while the product dies
Nobody announces it. The pilot just quietly stops getting mentioned.
The real failure mode: we treat AI like a purchase
"We need AI" is not a strategy. It is a vibe.
A strategy sounds like:
reduce time-to-disposition in the ED without increasing risk
reduce diagnostic delays in one defined pathway
reduce documentation burden without breaking billing integrity
reduce avoidable readmissions for a specific cohort
When you start with a measurable problem, the tool choice narrows fast.
When you start with "AI," you spend months debating features that do not move an outcome.
Stop buying models. Start owning workflows.
Trust is not a UI preference. Trust is the product.
A model can be statistically impressive and operationally useless.
In healthcare, a recommendation you cannot interrogate is a liability. Clinicians know it. The first question is not "how accurate is it?" The real question is:
What happens if I follow this and it is wrong?
That is not negativity. That is professional ethics under load.
So trust has to be designed, not requested. It is built through:
workflow truth (what actually happens, not what the policy says happens)
governance (clear guardrails so people feel safe to use it)
ownership (a named adoption owner with authority, not vibes)
feedback loops (the tool learns from reality, not committee opinions)
Training helps. Training alone is a fantasy. Behavior change happens under load.
Data quality is not a technical footnote. It is the foundation.
Every AI conversation becomes a data conversation because healthcare data is messy in the exact ways that matter:
fragmented across systems
inconsistently coded
missing at the moment decisions are made
"available" in theory, inaccessible in practice
Then the model becomes the scapegoat for a weak foundation.
You do not fix this by chanting "interoperability." You fix it by narrowing scope: one workflow, one dataset, one decision point, then make it reliable.
A framework that survives reality: Foundation → Pilot → Scale
If you want AI that does not die quietly, you need an execution path that respects healthcare's constraints.
Phase 1: Foundation (Weeks 1 to 6)
Pick one high-impact workflow where pain is real and measurable. Map the real workflow, not the brochure version. Audit data for that exact path. Define success like an operator, not a marketer.
If you cannot answer "who uses this at 3 a.m.?" you are not building. You are presenting.
Phase 2: Pilot (Weeks 7 to 16)
Start small. One unit. One population. One moment of use. Co-design with the people who carry the risk. Iterate fast based on real friction, not meeting-room preferences.
The goal of a pilot is not a press release. The goal is proof of trust under pressure.
Phase 3: Scale (Weeks 17 to 52)
Scale what works. Kill what does not. No zombie pilots. Integrate into the workflow instead of bolting on another screen. Build internal capability so you are not permanently vendor-dependent.
This is slower than buying a tool.
It is also how you avoid wasting a year.
What real success looks like (and why it is rarely flashy)
The wins that matter do not look like sci-fi. They look like less chaos.
A health system wants to reduce readmissions. The reflex is to shop for an off-the-shelf model. The operational reality is local: their patterns, their discharge constraints, their failure points.
So the work starts where it should:
find the smaller cohort driving most avoidable returns
build a focused signal for that cohort
land the recommendation inside discharge planning, where action is possible
Because in healthcare, prediction without action is just a score.
The path forward: less hype, more operational courage
Healthcare does not need AI slogans. It needs leadership willing to do three hard things at once:
demand measurable outcomes (not vibes)
protect clinicians from useless complexity
do the unglamorous work of workflow change
AI will not "disrupt" its way into hospitals.
It will earn its way in, workflow by workflow, trust by trust, outcome by outcome.
Build for 3 a.m., not demo day.
The One-Page Play (run this quarter)
Pick one workflow. Not a department. Not a platform. A workflow.
Define the receipt (choose one primary metric):
time-back (minutes per clinician per shift)
throughput (time-to-disposition, LOS, turn time)
safety (avoidable harm proxy tied to the workflow)
cost (specific avoidable utilization, not "AI ROI" vibes)
Design rules (non-negotiable):
The recommendation must land inside the existing work, not on a new dashboard.
A clinician must be able to ask "why" and get a usable answer.
The tool must reduce cognitive load, not redistribute it.
If you bought it, it is yours. No black-box dependency that blocks ownership.
Decision gate at 90 days:
If frontline usage is real under load and the primary metric moves, scale.
If usage is forced, explanations are weak, or the workflow gets slower, kill it and fix the foundation.
That is the whole game. Not hype. Not vibes. Execution.