Siloed AI is a risk you can't see

We have been yelling at each other about this one for weeks, which is how you know it belongs on the internet. When does a personal Claude skill or ChatGPT project stop being enough, and a team need a shared agent working from the same data? Anthony’s test: find the people who already share your level of data access, the ones you’d happily screen share with. That’s where the agent goes. And if you catch yourself asking a coworker to repeat something that was already captured somewhere, the silo is the problem.

Joe tells us how his SOW project left his personal ChatGPT, moved into Bruce, and now writes most of them for him, which raised the question of what Joe is even here for. (Shaking hands. Kissing babies. Lunch.) We get into prototyping alone before handing a workflow to the team, raising the permissions without changing the workflow, and the theory that how much data your agents share is a proxy for how much your team trusts each other.

Then the cautionary tales: the ChatGPT suggestion Joe clicked that emailed Ryan and Anthony on its own, and the case study Bruce wrote with a little too much project detail in it, which is why Rudy, our website agent, is a different agent. Plus Anthony on deterministic workflows, why big orgs should hand out the good tools before people bring their own, and why everyone wants to approve every agent action for about two days.

We closed by announcing the wrong domain. It’s weshouldtalkontheinter.net. It also gets published from Ryan’s laptop, which makes it a silo. We know.