Your AI agent works on your machine. I always use Claude Code on the terminal, with full computer use, web access, and access via CLI to the rest of my world. When I asked people to just do as I do, I discovered they were using Claude on the web, with no access to their own computer or even their files.
Nobody had refused to change. The app had been pushed to their machines, but they never looked, never noticed it. People in that organisation use a lot of web-based tools, including Office 365: that was their default.
An agent plan makes my mistake at a larger scale. It names a model, a workflow, and a saving, and it assumes that everyone else will behave differently from the day it goes live. Before you deploy an agent, name everyone who has to behave differently for it to work, why they would, and how they would even know.
Refusal is the easy case
In The Forward Deployed CPTO I wrote that access is the constraint, not talent. I can ship something without asking six different owners for permission first, and almost none of my engineers can. That post was about permission. My just do as I do moment showed me a second group, so two groups have to change for your agent:
- The owners. They hold a system, a dataset, or a budget line, and they may have a good reason to say no.
- The users. They have a default that works for them, and nobody has given them a reason to leave it.
Owners are the easy case, because you can see them. They say no in a meeting, and you can argue with the reason. Users say nothing. They carry on working the way they did, so the plan never hears about it. I only found out because I asked.
The view from the top is the worst place to look from. The more senior you are, the more keys you hold, so your own demo tells you the least about anyone else’s.
The other side has priorities
Cory Doctorow writes: It’s easy to come up with amazing applications for technology if you’re allowed to assume that you can get everyone else in the world to set aside their priorities and their interests in favor of yours.
He was writing about shopping agents, and shopping shows what happens when the other side has a reason. The Verge reported that a Meta agent could keep re-adding The RealReal’s listings to a cart each time a 20-minute hold expired, blocking other shoppers for hours or longer. The RealReal’s chief technology officer then said the company’s bot defences make it much more difficult for all agents to log in, add items to cart, and check out. Amazon began blocking agents from making purchases, while Walmart launched a gateway for approved agents: cooperation on its own terms, through a door it controls.
The same interests exist inside your company, and Doctorow’s engineering point applies there too. Add specialised bots to a system and, he writes, This compounding error rate only increases as you add more and more bots. Say each step works 95% of the time, every step must succeed, and failures are independent: five steps give you about 77% end to end, and ten give you about 60%. Treat that as illustrative: the numbers are mine, and nobody measured them. Checks and retries improve the sum, but they are work the demo left out, which is why the evaluation harness comes before the agent.
People are steps in that chain as well. A colleague who has never opened the app is a step that fails every time.
What assuming success costs
Leaders who assume success can destroy the business. The agent demos well, a pilot number nobody ever observed goes into the plan, and the next decisions stand on it: the budget moves, roles go, and the old process gets switched off. If you remove the ability to do the work the old way and the new way runs into these problems, doing the work is now impossible.
Org changes get undone quickly even when nothing was switched off. Meta set up an applied AI engineering team at 50 people per manager in March, and by September Business Insider reported that it was asking some individual contributors in its Applied AI division to move back into management. Meta has not said why, as I noted in What Middle Managers Are For. It can at least still appoint a manager. A process nobody remembers how to run is harder to bring back.
Defaults do move, and owners do give in. Walmart cooperates, and a chief executive can issue a mandate. But a mandate is not adoption, and my own story shows why. The people I asked were already using Claude, so a mandate to use Claude would have changed nothing, and a usage count would have reported success. The plan has to say who moves, when, and at what price.
Before you deploy
Five questions, in the order I would ask them:
- Who has to behave differently for this to work? Go and look at what they use today.
- What do they lose if it works?
- Why would they change: a mandate, a payoff, or their own interest?
- How would they even know? That takes comms, training, coaching, and empathy.
- What does the plan do when they don’t, and how often does the whole chain succeed?
Your agent works in the demo because it runs on your machine, with your keys. Have you looked at anyone else’s?
Leave a Reply