One of the ironies of our new machine age is that it still seems to involve an awful lot of meetings. At Anthropic, where artificial intelligence (AI) adoption is understandably high, the median employee now spends 20 per cent more time in meetings than a year ago.
But while AI hasn’t superseded the need for human-to-human interaction at the company, it has changed the nature of it. Namely, people are now under pressure to communicate with each other in a machine-readable way.
In an interesting report about how employment is changing inside Anthropic, the company wrote that “private work, in unlogged meetings and direct messages, is invisible, confusing and discouraged in an AI-heavy workplace” because it means AI agents cannot gain up-to-date context about everything that is happening. That can lead them down blind alleys.
One poor soul, whose name was on a months-old document about Claude merchandise, found that AI agents kept channelling requests to them for T-shirts and mugs, even though their job responsibilities had changed.
READ MORE
As a result, it is much better for work to happen in the open: agent-visible Slack channels; recorded meetings. Anthropic is not the only AI-heavy workplace to have come to this conclusion. Shopify has built an AI coding agent that lives in the company’s Slack and only works with people in open channels, for example.
What are the advantages of working this way? Tobias Lütke, Shopify’s chief executive, has argued that it creates a sort of “Lehrwerkstatt”, or “training workshop”: everyone can watch each other working with the AI tool and so learn by osmosis. “It does not require a curriculum, a training plan or a manager,” Lütke has written. “It just requires everyone’s work to be visible to the maximum extent possible.”
The more visible the work, the more information AI agents can absorb, too, of course. This raises the prospect of AI not just as a tool but as a vast sponge of corporate information, including previously uncodified “tacit knowledge” absorbed by listening in to every meeting and Slack conversation.
[ When human judgment is not all it’s cracked up to be ]
There could be efficiency gains to this. Human co-ordination has its limits, after all. Institutional knowledge and memory often exist inside different people’s heads. They can warp or erode, especially when key people leave. In contrast, a fully queryable and constantly up-to-date repository of all that knowledge could be immensely useful, especially for new starters.
But knowledge is power, and centralising it in an AI system would redistribute that power in potentially disruptive ways.
It is one thing for Anthropic to feed Claude (its own product) as much information about its organisation as possible; it would be another thing entirely for a company to cede so much power to a system it did not own.
Workers, too, would lose the monopoly on something that is currently valuable (the tacit institutional knowledge built up inside their heads).
But more than that, I suspect the biggest disadvantage of working this way would be that many people wouldn’t like it.
When the academic Ethan Bernstein studied open workplaces, he found that too much transparency led to “counterproductive inhibitions” and stifled experimentation. I can relate to this. At the Financial Times, we can choose to write our articles directly into our publishing software (which makes the work-in-progress visible to editors and colleagues) or we can write on our own devices and then copy over a finished draft. I always choose the latter.
I find it uncomfortable to work if I think someone – or something – is looking over my shoulder. And I suspect I am not the only one.
As for meetings, the pandemic caused a lot of debate about the value of impromptu work conversations sparked by chance meetings by the water cooler. Whether these informal channels are as important as their advocates claim, I doubt that the requirement for every work-related encounter to be recorded and codified for posterity would make communicating with colleagues more enjoyable.
Indeed, one danger would be that many private conversations would continue to happen – and before long the AI agents’ picture of the organisation would diverge from reality in subtle but potentially problematic ways.
That doesn’t mean more companies won’t try it. There has been a lot of talk about AI models as “black boxes” whose inner workings we cannot understand. But rather than try to make them more legible to us, we appear to be moving towards a world in which we make ourselves more legible to them. – Copyright The Financial Times Limited 2026














