Give agents the repeat work, keep the judgement

I built AI agents that work as digital workers for me. They take the repeat work, so my time goes to strategy, to hands-on model work and to talking with the people behind each project. I think every AI leader should work this way, because running your own agents is the fastest way to learn what the whole company will need.

What my digital workers do

None of it is glamorous. That is the point.

The weekly AI spending report. What we spent, where and on which models.
The weekly AI brief.
Travel booking in Ramp.
My calendar. Organising it, so meetings land where they should.
Deadline reminders. So nothing slips between projects.
The AI governance onboarding plan. Keeping it up to date as departments join.

Together they would take a real share of my week. Now the agents do them, and the time they return goes to the work only a person can do.

What stays human

Agents are good at work that repeats. They are not where judgement should live. The decisions I keep are the ones with no template:

Data architecture under complicated regulations. Senior living and skilled nursing mean personal and health information. How data is structured decides what is safe to automate at all.
Starting AI governance when there is no playbook. On my first day in the role there was none for MCP connections. Someone has to decide where to begin.
Strategy when new models arrive every few weeks. The question is not which model is newest. It is which change matters for our users, and which can wait.
The model lifecycle. How a model is evaluated, released, monitored and retired.
The return an agent can really bring. Not what the demo promises, but what it changes in time, cost or risk.

And the most important work of all: talking with people. Every project has a team behind it, and their worries and ideas never show up in a dashboard.

Least disruption, same direction

When I make a decision, I ask two questions. Which approach disrupts users least? And which gets every department moving in the same direction faster?

Those questions shaped how we governed AI across the company. We moved governance to the connection layer and gave everyone one governed route to AI. As a design partner to Barndoor, the gateway we chose, I give feedback on its product so that governance does not become a barrier to adoption. Barndoor has published a case study of the work.

Today every MCP connection is governed, and everyone in the company, across more than 30 departments, uses AI through that route. My own agents use it too. I own everything about AI in the company, so I hold myself to the same rules as everyone else.

You cannot manage what you cannot see

My weekly AI spending report only works because spend is visible in one place. Without central governance and visibility, you do not even know what each department spends across vendors, or how efficiently it uses tokens. You cannot manage that, and you certainly cannot defend it to a leadership team.

So visibility came first. Now we know AI cost at any time. Every employee has a tool that shows cost in real time and within each session, and we have tracked a 38% reduction in AI cost. Model routing, which sends each task to a model suited to it, is part of that.

People make better choices when they can see what a choice costs. The reason is simple. When AI cost arrives as one bill at the end of the month, nobody who made the choices ever sees it, so nobody can act on it. When each person can see what a session costs while they are working, cost becomes part of the decision. They can choose a lighter model for a simple task, rephrase a request instead of running it again, or stop a session that is going nowhere. The same holds one level up. When each department can see its spend across vendors next to its results, budget conversations start from evidence instead of guesses. Visibility does not police people. It gives them the information to make a good choice themselves.

This is the same lesson my own agents taught me, at company scale. An agent is only as useful as the data and the governance around it.

Why AI leaders should run their own agents

I came to AI through operations research, and I still like to test an idea on myself before I ask anyone else to adopt it. Running my own digital workers does three things a slide deck cannot.

It shows where the friction is. Every place my agents get stuck, on access, on data quality or on unclear rules, is a place a whole department will get stuck too.
It makes governance concrete. I know what a sensible policy feels like from the user’s side, and what makes people want to work around one.
It frees the time leadership needs. Strategy, hands-on model work and real conversations with people do not fit in a week full of repeat tasks.

What to do next

If you lead AI, try this over the next month.

For one week, write down every task you repeat.
Pick three with a clear definition of done and a low cost of error, and give them to an agent.
Connect those agents through the same governed route you expect everyone else to use. Mine do.
Measure the time you get back, and spend it on judgement and on people.

Then use what you learned to set the policy for everyone else.

About the author

Dr. Linda Xu is Director of Data Analytics & AI at MealSuite, where she owns AI end to end. More about her.