Accountability does not delegate
An agent can own a task. It can’t own the consequence.
Oct 16, 2026 · @Akshay
Analytics and AI have climbed a ladder: descriptive, diagnostic, predictive, prescriptive, generative. Each step has been more useful than the last. Agentic is different, because for the first time the system acts.
There’s a growing belief that autonomous agents can be switched on and left to run. For a large enterprise, in October 2026, I don’t think that’s true. An agent can own a task. It can’t own the consequence.
When we let it run
With a national grocer, we built a set of agents to run promotional campaigns. They could do nearly the whole job: set the objective, pick the target customers, write the briefs, simulate redemption and margin uplift (around 2.2–2.3% on one campaign), recommend the channels and create the content.
At first, we let them run. No step had an owner, and nobody owned the system as a whole.
On one campaign, the agent recommended four channels, and the forecast assumed all four. What it couldn’t see was that the company was mid-negotiation with the media partner behind one of them. The campaign ran on three, results came in well below forecast, and because nobody owned the call, nobody caught it.
An owner at every step
After that, every step got an owner. The campaign analytics lead signed off the target customers, and if they were wrong, it was on them. The head of digital marketing owned the whole system.
Get Akshay Bardia’s stories in your inbox
Join Medium for free to get updates from this writer.
It slowed things down. It also gave us our best results.
People hold the thresholds
“Keep a human in the loop” isn’t enough on its own. Ask someone to approve hundreds of agent actions a day and they’ll start clicking yes. The real design question is which few decisions a person holds, and where the thresholds sit. Below them, the agent acts. Above them, its owner decides. As the agents learn, the thresholds can move.
Three rules
- One senior owner for the whole system.
- A named owner for every step that changes what customers see or what the company spends.
- Thresholds set in advance, and reviewed as the agent earns more trust.
An agent can own a task. The consequence needs a name next to it.
