Why AI Errors in Biotech Surface Months Too Late
An AI error in biology is expensive because of when it is found, not because of what it is. A model returns a confident, wrong answer at a desk in month one. The work built on top of it fails at the door of a manufacturing facility in month nine. Ryan Cawood, Co-Founder and CEO of Lab Thread, has spent years on the data layer underneath laboratory science, and this is the part of AI adoption he thinks is underrated.
the real danger though is temporal
Most discussion of AI risk in drug discovery imagines a dramatic failure. What he describes is quieter and much harder to catch: nothing looks broken at any point until the moment it is too late to fix cheaply.
Where does an AI error actually enter a lab?
Not at the bench, and this surprises people. Asked what a failure looks like in a working lab, Cawood corrects the premise before answering it.
AI I wouldn't say is actually in the physical lab yet
There are no robots misreading a pipette. The error enters upstream, in the ordinary act of asking software to interpret something.
it's in the office where the mistakes are made ... you're uploading a file and you're saying interpret this data without guardrails in place
Someone uploads a file without telling the model what the data is. The model assumes a standard curve that is not a standard curve, and returns a number. The number is plausible. It goes into the record, and the next experiment is designed against it. Stating what the data is before asking anything of it is the first move in how to prompt AI for scientific research.
Why does the failure take nine months to appear?
Because biology runs on a slow clock. Cawood walks through the specific chain, and it is worth following because every step is ordinary.
A piece of DNA is selected. It goes into a cell line. That line is grown and passaged, sometimes fifty to a hundred times, over the better part of a year, toward a handoff into a GMP environment. If the original sequence was never validated, nothing in that process surfaces the problem. Every downstream step succeeds on its own terms.
the cell line you've spent 9 months making is now no longer viable in a GMP environment
The cost is not the wrong answer. The cost is nine months of correct work performed on top of it, which is why the same error that would be trivial in a software company is severe in a biological one.
Is the paperwork the same problem?
It is the same shape, and in practice it bites more often. Chain of custody for a cell line is still frequently kept by hand.
that poor technician or that poor scientist that's in the lab has to write down every batch, every lot, every every single aspect of what they're doing to that cell line
A single missing batch record, discovered at handoff, can invalidate the package. And because the records are reviewed periodically rather than continuously, a gap can sit open for weeks before a month-end check surfaces it. That is the same latency problem wearing different clothes: the work was fine, the documentation of the work was not, and nobody knew for months.
Can AI fix the problem it is creating?
This is where Cawood is genuinely optimistic, and the distinction he draws is the practical one. He does not want the model generating the conclusion. He wants it checking whether the inputs are complete.
Asking software to read nine months of batch records weekly and flag what is missing is a task with a verifiable answer, no judgment, and a fast feedback loop. It is the inverse of asking it to interpret an unlabeled data file. But the value collapses entirely if the underlying records are wrong rather than merely absent. Running that check continuously is a recurring compute bill rather than a one-off, and the price of that input is less settled than it looks, which we take up in why AI is so expensive to run.
if you give them flawed data, you will get a flawed answer
Which returns to the constraint underneath all of it. A model trained on the collective published record is only as useful as the specific, local, messy data you hand it, and biology produces exactly the kind of data that resists being handed over cleanly. Materials change. Cells propagate and are destroyed. A sequence becomes a virus becomes a line. If none of that was captured as it happened, the conclusion at the end is built on nothing, however fluent it sounds.
What does this change in practice?
It moves the checkpoint earlier. If the expensive error is the one discovered at month nine, the useful investment is validation at month zero, on inputs that feel too trivial to verify. Cawood's own company exists to make that record continuous rather than reconstructed, which is a bet that the documentation problem and the AI problem are the same problem.
The broader pattern here is familiar to anyone who has watched cost accumulate in this industry. The expense of cell and gene therapy is not one large line item but a chain of steps that each have to go right, a dynamic we looked at in why cell and gene therapy is so expensive. Data provenance behaves the same way, and it raises the same awkward question about who is holding the record in the first place, which we covered in who owns the data in a collaboration agreement.
The convener's read
Cawood's account is not a warning about AI. It is a warning about latency, which existed before AI and which AI accelerates by producing plausible inputs faster than anyone can check them. The organizations that will get hurt are not the ones using models aggressively. They are the ones whose records cannot answer a question asked nine months later.
Finding the error late is one problem. Owning it is another, and the FDA is blunt about who does: who is actually liable when a model produces the data.
The full conversation is on the episode page, along with more from Ryan Cawood. If your organization wants to reach the people making these calls, partner with Open Door Salon.
Methodology: every quotation above is drawn verbatim from the recorded, on-the-record conversation between Ryan Cawood, Raphaël Ognar and host Lori Ellis on Open Door Salon, and was checked against the episode transcript before publication.