Every Inconsistent Sample Is a Hypothesis You Can't Trust
Meghav Verma
Manual sample prep has a way of hiding in plain sight. It's not the part of the workflow anyone complains about loudly; it's just the thing that happens before the "real" science starts. A tech extracts, evaporates, reacts, moves on. It's been done this way for decades, by hand, and it mostly works well enough that nobody stops to ask what "well enough" is quietly costing them.
But “well enough” is exactly the problem when the data downstream feeds an AI model instead of a single analysis.
The Bottleneck You've Already Accepted
Every lab has its version of this: one tech who processes samples slightly faster than another. A pipetting technique that varies by hand, by fatigue, by how many samples came before it in the run. A step that says "process promptly" in the SOP but means something different depending on who's at the bench that day.
None of this looks like an error. It's just how manual work behaves; it drifts. And for a long time, that drift was tolerable, because the science downstream could absorb it. A statistician could flag an outlier. A researcher could sanity-check a result against domain knowledge.
AI models don't do that. They don't know which variation is biology and which is bench technique. They just find the pattern that best explains the data they're given. If your "treatment" samples happen to have been processed, on average, a few minutes differently than your "control" samples, the model may key off that instead of anything real. It doesn't announce the mistake. It just performs beautifully in validation and falls apart the moment it meets new data.
The Real Cost Isn't Time; It's Trust
The instinct is to think of manual prep as a throughput problem: it's slow, it's a bottleneck, it caps how many samples you can run. That's true, but it's not the expensive part.
The expensive part is what happens to the data itself. Every sample that was handled just slightly differently than the one before it is a small, invisible edit to your dataset. One that no amount of downstream analysis can fully undo, because you often can't tell which samples were affected or by how much. Multiply that across a study, across sites, across months of runs, and what you're left with isn't just noisy data. It's data you can no longer fully vouch for.
That's the real definition of the bottleneck: not how many samples you can process, but how much you can trust the ones you already did.
Naming It Is the First Step
This isn't a knock on the people doing the work. Manual sample prep variability isn't a discipline problem; it's a structural one. Asking a human process to be perfectly consistent, run after run, indefinitely, is asking for something human processes were never built to do.
At Axle Research and Technologies, this is the problem we've spent years living inside of alongside the labs we work with. Not as outside observers, but as people who've stood at the bench and felt exactly this friction. We think the field is overdue for naming it plainly: inconsistent sample prep isn't a minor inefficiency in an AI-driven workflow. It's a quiet failure mode sitting right at the foundation of it.
This is exactly the kind of question we'll be unpacking at Future Labs Live 2026. If it sounds familiar, come find us there.
