Three Questions to Answer Before You Automate

Imagine asking five people on the same team how they handle a customer refund.

One follows the 30-day deadline. Another makes exceptions for loyal customers. A third considers the account’s value. A fourth asks their manager. The manager says, “It depends.”

Everyone believes they are following company policy.

Now, ask an AI system to handle those refunds. Which version of the policy should it follow?

This hypothetical example captures a challenge that deserves more attention in enterprise AI adoption: work can get done by people for years without everyone agreeing on how it should happen. Automation makes those differences harder to leave unresolved.

Leaders need to bring three questions into the conversation earlier.

1. Is This Worth Doing?

An AI project begins with a manageable goal, like helping employees process refund requests faster.

Then the suggestions arrive. Could it personalize responses? Predict complaints? Identify opportunities to retain customers? Perhaps the whole customer service process needs a redesign.

Each idea has merit. But together, they can turn a useful improvement into a project too large to finish.

The desire to make something better can obscure the importance of the work itself. A task may be frustrating, visible, or interesting to automate — without being where the organization most needs to invest.

Even when the task matters, it is easy to improve the wrong part. If a refund takes five days because it waits for approval, drafting the response faster will barely change the customer’s experience.

Even when the task matters, it is easy to improve the wrong part. If a refund takes five days because it waits for approval, drafting the response faster will barely change the customer’s experience. Removing an unnecessary handoff might.

Before building, establish what a worthwhile improvement would look like: less waiting, fewer errors, lower effort, or a better customer outcome. Use that goal to judge proposed additions.

The question “Could we also do this?” needs a companion: “Would it matter enough?”

2. What Should the AI Do?

Once the goal is clear, the team needs to agree on the decisions the system will make.

This sounds straightforward — until people start comparing answers.

In our refund example, one employee prioritizes consistency. Another prioritizes preserving the customer relationship. Both believe they are exercising good judgment. A manual, human-powered process may accommodate both approaches quietly.

But automation requires clearer boundaries. When should the deadline apply? When is an exception justified? When must the system ask someone?

These questions involve authority as well as policy. Choosing an approach affects whose judgment carries weight, who can approve exceptions, and who bears responsibility for the outcome. 

That is where organizational politics can start to impact a seemingly technical discussion. Another round of requirements may reflect an unresolved disagreement about who gets to decide.

Before development, ask the relevant teams to work through a few concrete cases — including difficult ones. Compare their answers, and identify who has the authority to settle differences.

You do not need to anticipate every exception. However, you do need agreement on what the AI may decide, and when it must defer to a person.

3. How Will People Work With It?

Now, imagine the system is ready. It handles straightforward refunds and sends unusual requests to employees.

Who receives those requests? Is reviewing them part of their job? Can they override the system? Who updates the system’s instructions when the policy changes?

Without clear arrangements, automation can move a bottleneck rather than remove it. Routine processing becomes exception handling, but in many cases, nobody has allocated time for that new work. And employees may continue performing their old checks, because they do not know which ones the system has replaced.

‘Keep a human in the loop’ is only useful if that person knows what they are expected and authorized to do.

“Keep a human in the loop” is only useful if that person knows what they are expected and authorized to do.

Before an automated system is deployed, involve the employees whose work will change. Assign responsibility for reviews, overrides, and ongoing updates. Test the handoffs alongside the system’s output.

The pilot should demonstrate that people can work with the AI — including when it is uncertain or wrong.

Start With the Work

Answering these three questions can avoid a lot of wasted time, energy, and money.

They can also make technical development more purposeful. Builders get a clearer goal. Employees help shape their future responsibilities. Leaders can see which disagreements require their attention.

AI creates an opportunity to reconsider how work happens. Taking that opportunity seriously means making decisions that a familiar process may have allowed everyone to postpone.

It’s worth having those conversations while designing the system, because they get much harder when upset customers start asking why the AI did what it did.

Daniela Gorza is Co-Founder and COO of Maigent, a Boston AI company building personal AI that runs on the user’s own device. She holds an MBA from MIT Sloan.