AI adoption
AI trust is earned
AI trust must be earned and never assumed.
AI trust must be earned and never assumed.
Slowly removing humans
I recently discussed this with a senior member of a global software company. We talked about their content creation engine.
They’ve created a team of agents, including reviewers, grammar-checkers, data retrieval, and fact-checking agents. Each agent has a single job and they are all overseen by humans. There are humans-in-the-loop at every step.
Their deliberate intention is to slowly remove some, not allAnthropic found new Claude Code users auto-approve about 20% of the time, veterans over 40%, yet they interrupt more. Oversight shifts from approving to intervening, of those human checks. They’re focused on how agents and humans collaborate and how that relationship is going to evolve.
To reduce risk, they’ve chosen localised long-tail content rather than their flagship content. That containment of the blast radius of any potential issues is important.
Keeping humans-in-the-loop is a trade-off. It makes things slower, but it also reduces risk.
Positive friction
Humans provide positive friction and consideration: a change request, a pull request, publishing workflows. These all make people stop, deliberately slow down and give something more consideration.
It always makes me think of the Japanese technique of shisa kankoliterally just pointing and saying. But this simple act makes you stay more mindful and engaged, to stay mindful.
You need to find the sweet spot: enough oversight to manage the risk without pulling agents back to human scale or human speed.
Expect it to go wrong
When working with agents I always like to keep Murphy’s Law in mind: everything that can go wrong will go wrong.
Only trust AI with what you can afford to go wrong. Not less, but definitely not more.