Every AI agent gets a launch. A review, a demo, and a name on a slide that says who owns it.
Then the launch ends and the agent keeps working, every day, making small decisions nobody reads.
Ivanti asked 1,500 IT professionals about this earlier this year. 85% said every agent they run has a named owner. 42% said that ownership is clear. That gap is the distance between who owns the agent on paper and who is watching it on a Tuesday afternoon.
The failure worth worrying about is not the agent that breaks loudly. That one gets fixed. It's the agent that keeps running and starts drifting. It might return something that looks right, and quietly skip the next step in the pipeline.
Most of us, including me, have learned this one the hard way.
In a system I built for a client, an approval step stopped working because two settings didn't match. No error, no alert. The work just stopped moving, and I only found it by going to look.
Now, every agent I use, build, or deploy gets three things on day one:
1. Every action is captured and traced.
What came in, what the agent decided, which tools it touched, and what went out. Every step, every time.
The record is not the agent's job. It's written by the system around the agent, in plain code that runs the same way every time. An agent can forget to report. A trace built into the plumbing can't.
It also shouldn't depend on where the agent was built. Whether it lives in an app you bought, a framework, or code your team wrote, the record of what it did has to work the same way. The platforms will keep changing under you. The record has to outlast them.
2. Failures and escalations show up where your people already are.
When the agent fails, gets stuck, or reaches something it shouldn't decide on its own, it says so, on the apps and devices your people already use. An alert that lands in a separate tool nobody opens gets ignored.
3. Approvals take one step.
Somebody has to be able to move the work forward quickly: read what the agent wants to do, then approve it, correct it, or take it over. Again, this has to be where people are working. If approving means logging into different tools, the work piles up behind the gate, or people start approving without reading. Either way, the gate is gone.
And a bonus: build trust levels into the tools themselves.
Give every tool and every workflow a trust level, with guardrails: read, draft, ask, act. Draft means it writes and you send. Ask means it sends once you say yes. A new agent starts at the bottom. As the trace shows it getting things right, you turn the dial up one tool at a time, and you can turn it back down in a minute when something changes.
That's how you dial AI in to work with you, instead of choosing between all and nothing.
What it looks like for us today
For the agents we build, all three run through Slack. Every step lands in a thread, escalations ping a phone, approvals are a tap, and the same thread doubles as the place I coach the agent when it gets something almost right. It works, though I think something better could exist. That's a future post.
This is ownership after launch: deciding what gets recorded, who gets called, and who moves the work forward. It's a leadership job, and it's the one most teams skip, because a launch is visible and day 31 isn't.
This week, ask your team one question: where would we go to see what our agents did yesterday? If nobody has an answer, that's the first thing to fix.
If you're working that out, I'm glad to compare notes.
Could your team answer it today?