Earlier this week, Anthropic said its Claude model had spent 21 hours searching a bacterial DNA database without anyone watching over its shoulder, and had surfaced a previously undetected enzyme system with CRISPR-like characteristics. A Stanford bioengineering professor called the pattern-spotting "incredibly exciting." At least one microbiologist added, fairly, that it isn't yet a rival to CRISPR the technology - more a lead worth someone's proper attention.

Most of the coverage stopped there, at the biology. That's a shame, because the more useful story for most businesses has nothing to do with enzymes.

Here's the part worth sitting with: this is what genuinely autonomous AI work looks like now. Not a chatbot answering a question in a few seconds. Not a tool waiting patiently for the next prompt. Twenty-one hours of unsupervised searching, ending in something a human expert then decided was worth following up.

That's a capability, and it raises a question most businesses haven't properly sat down and answered yet. If an AI tool can work for hours without anyone checking in, how many of those hours are you actually comfortable handing over?

It's tempting to answer that the way most AI adoption conversations start: with the tool. Which model, which plan, which integration. But the more useful version of the question isn't about the tool at all. It's about decision rights - specifically, which decisions inside your business you're willing to let something run on for an extended stretch without a person in the loop, and which ones need a person checking every step.

Those two categories aren't fixed. They shift by task, by risk, and by how well you've actually tested the tool on your own work rather than taken someone else's word for it. Twenty-one hours unsupervised is one thing when the worst case is a wasted afternoon of compute. It's a different conversation entirely when the unsupervised task touches client communications, financial commitments, or anything that leaves your business once it's sent.

This is precisely the gap a proper AI readiness assessment is built to close, and it's usually smaller and more mechanical than it sounds from the outside. Most businesses we work with haven't drawn this line anywhere. AI use has grown task by task, person by person, without anyone stepping back to ask which of those tasks now run with real autonomy, and whether that autonomy was actually agreed or simply accumulated.

Drawing the line doesn't require slowing everything down. It requires being specific. Which tasks can run unsupervised overnight. Which need a person to review the output before it goes anywhere. Which shouldn't be automated at all - not because the AI can't do them, but because the cost of it getting one wrong is too high to hand over. That's a short, concrete conversation, not a lengthy one, and most businesses have never actually had it out loud.

Claude's twenty-one hours in a DNA database is a striking example precisely because the stakes were low and the result was genuinely useful. Most businesses won't get that combination by accident. They'll get it by being deliberate about where autonomy is earned rather than assumed, testing on real work before extending the leash, and treating a good result as a reason to ask the next question rather than a reason to stop asking.

The scientists who reacted to this story didn't get excited because a computer worked unsupervised. They got excited because it produced something worth a human's attention, and someone was there to give it that attention properly. That's the model worth copying, whatever your business does: give AI room to work, and keep a person genuinely in the loop for whatever it hands back.

Go wisely.