An AI agent that can take an action shouldn’t automatically be allowed to. Autonomy should be earned through evidence, not granted because the technology is capable of taking the action.
I build Agentic Delivery Agents: operational agents that work inside set boundaries. Most automation talk aims for full autonomy. My work aims for the right amount. I first walked through this in Episode 6 of Divya Does AI.
Rules first, model second
A typical AI agent starts with what the model can do. It reads information, reasons through options, and uses tools to finish a task.
The Agentic Delivery Agents I build start with the business. Rules, delivery context, policies, and the outcome the team expects. Then they look for the problems that matter: a staffing conflict, a project running past its planned hours, approved time off missing from the schedule.
Not every job needs a large language model. Sometimes a fixed rule and a clear threshold do the work better. The goal isn’t the most advanced tool. It’s the right amount of intelligence for the problem.
Four stages, in order
I give an agent more authority one stage at a time. Each stage has to hold before the next one opens.
- Observe and identify. The agent spots what needs attention. Someone over capacity. Planned hours that don’t match actual hours. A stale project status. It decides nothing. It just saves people the digging.
- Interpret and recommend. The agent explains what the problem means and suggests a fix. Flagging that someone is overloaded is easy. Deciding who takes their work means knowing skills, availability, and client commitments. Data alone can’t answer that. Domain knowledge can.
- Execute within boundaries. For low-risk, repeatable work with clear rules, the agent can act. It knows when to proceed, what result to expect, and how to check the result stayed in bounds.
- Bring the human back in. Decisions that touch people, client commitments, budgets, or scope go to a person. The agent still gathers the facts and lays out the options. A person makes the call.
I think of it like power of attorney. You wouldn’t hand it to a stranger on the street. You’d watch how they work first, and wait until they deliver the same result every time. An agent earns trust the same way.
What I’ve seen
For one client, we needed 6,860 brand files tagged, logged, and checked. Doing it by hand, we estimated about 245 hours, or roughly six weeks for one person. My agent finished in about 2.5 hours. It was repetitive, low-risk work with clear rules, which is exactly what stage 3 is for.
Rule or judgment
Counting automated tasks is the wrong scorecard. The better one asks three things. Are the agent’s actions trustworthy? Are its results predictable? Do its limits match what the business is responsible for?
Some agents should only watch. Some should recommend. Some can act alone inside clear limits. Others should know when to hand the decision to a person.
The question I keep coming back to: is this decision governed by a rule, or does it actually require judgment?
Try this week: list five tasks you’d hand to an agent. Mark each one “rule” or “judgment.” Only the rule tasks are candidates for automation. The judgment tasks get an agent that prepares the decision, not one that makes it.
Adapted from my episode "AI Agents vs. Agentic Delivery Agents" on Divya Does AI.