The most underrated skill in AI-driven R&D is knowing when not to automate
The most underrated skill in AI-driven R&D is knowing when not to automate
There’s a quiet assumption in many AI and self-driving lab discussions: if something can be automated, it should be.
One moment that stuck with me was during a lab visit where a team walked me through an “automated” workflow they had just put in place. Throughput was clearly up. Data was flowing. The setup looked modern and well engineered.
But in the conversation that followed, an important nuance came out: despite the speed, they still wanted a bit more confidence that the outputs were as consistent and comparable as they needed them to be.
Nothing was “broken.” The instruments were running. The pipeline was producing numbers. The challenge was more subtle: the decision rules around the pipeline weren’t stable yet.
Edge cases, messy inputs, and inconsistent interpretation can get scaled just as efficiently as execution.
Automation is powerful. But it works best when the surrounding decisions are stable. If you automate a step that’s unclear, poorly understood, or constantly changing, you don’t remove complexity. You amplify it.
Not every part of R&D should be automated at the same time. Some steps thrive on repetition and consistency. That’s where automation creates huge leverage. Other steps still need judgment and interpretation. They also benefit from a bit of deliberate friction.
Early on, manual intervention can reveal insights that rigid pipelines hide. Later, automation shines. It enforces discipline, repeatability, and scale.
The mistake isn’t automating too much. It’s automating without sequencing and robust decision making.
The best AI-driven labs don’t try to eliminate humans. They decide where humans stay in the loop, and why. Progress isn’t about reaching “full autonomy” as fast as possible. It’s about workflows that know when to accelerate and when to pause.
That restraint is rarely discussed. But it often separates impressive demos from systems teams actually trust.
If this has come up in your team, feel free to drop me a PM - always happy to chat and tell you more about our work at @ReactWise.
