The Problem Is Not the Model. It’s the Meeting After the Model.
AI is already producing useful predictions, what organizations still don’t know how to produce are decisions
The room had everything it needed to make a decision.
Spatial transcriptomics data. Histological analysis. Months of work compressed into a presentation that showed everything, clearly. The genes were where they were supposed to be. The model worked.
Then someone asked the question.
So what do we do now?
No one answered. Not because they didn’t know. But because the question had no owner. The decision that should have followed from that data (advancing to the next clinical stage, redesigning the protocol, changing the patient selection criteria) wasn’t assigned to anyone in that room. The model had done its job. The organization hadn’t.
What happened next I’ve seen repeated across different formats, different companies, different technologies: more evidence was requested. More data. More layers of analysis. Not because the existing data was insufficient. But because producing more data is a socially acceptable way of not deciding.
THE PROBLEM NO ONE NAMES
70% of labs have the hardware. 10% actually use it.
70% of clinical laboratories already have whole slide imaging scanners
10% of pathologists use them routinely in daily diagnostic work
That gap, between owning the technology and using it to change decisions, is not a technical problem. It’s an organizational one. And it repeats at every layer of the system: in clinics, in hospitals, in translational medicine departments at large pharmaceutical companies, in Series B biotechs that just closed a round on an AI pitch.
The industry invests massively in building models. It invests very little in building organizations capable of acting on them.
The bottleneck in medical AI is not in the server room. It’s in the conference room.
TWO MECHANISMS, THE SAME RESULT
How projects that technically work still fail
I’ve identified two distinct mechanisms that produce the same outcome: a valid model that changes nothing.
In most modern organizations, information is distributed across multiple teams. But the authority to act on it remains fragmented or ambiguous. AI accelerates this tension, because it produces outputs that require immediate institutional coordination, inside organizations designed to dilute risk through consensus.
The result is a silent paradox: no one decides, because the decision is distributed. But accountability for the fact that nothing moves forward has no owner either.
Mechanism 1: Paralysis by evidence
The model delivers a clear insight. The data is solid. The scientific team trusts the results. And the organizational response is to ask for more data.
Not because there are genuine doubts about the quality of the analysis. But because the decision that should follow from that data implies risk, implies a change of direction, implies that someone becomes responsible. Accumulating more evidence is safer than deciding with the evidence that already exists.
The project doesn’t get cancelled. It gets postponed indefinitely under the banner of “we’re still analyzing.”
Mechanism 2: Intent without authority
Here the problem is different. The team wants to move forward. The input is good. The motivation exists. But the go decision never arrives. Not because no one wants to make it, but because it’s not clear who has the mandate to do so.
The project stays alive on paper. In practice, the people who built it start losing faith. What was once an interesting opportunity becomes routine maintenance work. And when someone finally makes the call (if it ever comes) the team is no longer energized to execute it.
This mechanism has a name in implementation science: decisional ownership failure. And according to Definitive Healthcare data, it’s the second most cited barrier by imaging and diagnostics leaders when explaining why AI doesn’t get adopted in their organizations.
AI projects rarely die from technical failure. They die from organizational exhaustion.
THIS IS NOT ANECDOTAL
The Epic Sepsis Model: when alerts don’t translate into action
Michigan Medicine deployed a predictive sepsis model integrated directly into their electronic health record system. The algorithm worked: it calculated real-time risk scores for thousands of patients simultaneously.
During the first wave of COVID-19, alert volume increased 43% while overall hospital occupancy dropped 35%. Nursing teams were receiving hundreds of alerts daily with no governance framework telling them how to translate a risk alert into a specific therapeutic decision.
The response was to turn the alerts off.
The model didn’t fail. The organization didn’t know what to do with what the model was telling them.
This isn’t an isolated case. 80% of AI projects in healthcare fail to move beyond the pilot phase. Between 49% and 96% of clinical decision support alerts are ignored or actively disabled by clinicians. Only 5% of pharmaceutical companies extract real competitive value from their AI investments.
IT ALREADY HAS A NAME
The AI Translation Gap: a clash of architectures
In 2023, researchers at Duke and Mayo Clinic formalized this phenomenon under the name AI Translation Gap: the chasm between the successful technical validation of a model and its inability to be systematically integrated and used by human professionals to change the course of real decisions (if you want to go deeper on this, I wrote about it here).
The definition is precise. But the problem it describes runs deeper than the name implies.
For years, the industry assumed that the main obstacle was building models good enough to enter medicine. That once an algorithm demonstrated clinical performance, adoption would follow naturally.
That assumption was wrong.
The real problem may be something else entirely: human organizations were never designed to operate around probabilistic outputs produced by machines. Their governance structures, their chains of authority, their institutional incentives, all of it was built for a world where knowledge arrives filtered through a human expert who assumes responsibility for the interpretation.
AI doesn’t arrive that way. It arrives as a probability distribution. As a score. As a prediction with no assignable human authorship. And organizations don’t know what to do with that in practice, because no one designed the processes to handle it.
What we’re witnessing is not a technology adoption problem. It’s a collision between inherited institutional architecture and a new computational epistemology. And that collision is not resolved with better algorithms.
The discovery didn’t die from lack of evidence. It died because no one had the mandate to act on it.
WHAT COMES NEXT
This post describes the problem from the inside: what it looks like, what it feels like, why it happens.
Next week: is the pharmaceutical industry actually ready for AI or is it just investing in it? There’s an important difference between the two. And the data suggests most organizations still don’t know which category they’re in.
The next competitive advantage in AI will not be algorithmic.
It will be organizational.
IF YOU RECOGNIZE THIS PROBLEM IN YOUR ORGANIZATION
I work with translational medicine teams and scientific leadership in biotech and pharma to diagnose and resolve the AI Translation Gap, from auditing the governance of existing models to designing decisional structures that turn technical outputs into real clinical actions.
If you want to explore whether it makes sense to work together: book a 30-minute call
References:
Overgaard et al. (2023). Implementing quality management systems to close the AI translation gap. npj Digital Medicine. doi:10.1038/s41746-023-00968-8
Naicker et al. (2026). Implementing an AI Decision Support System in Radiology: NASSS Framework. JMIR. doi:10.2196/80342
Wong et al. (2021). External Validation of the Epic Sepsis Model. JAMA Internal Medicine. doi:10.1001/jamainternmed.2021.3333
Definitive Healthcare AI in Imaging Survey (2025). Cloudera Life Sciences AI Report (2026). McKinsey Global Survey on Life Sciences AI (2024).