• Regulation & Compliance

I understand the appeal of handing more work to AI. When you run a business, something is always waiting: an unanswered email, a report to finish, a customer who needs a reply. Getting some of that time back matters. 

At Rowvar, I want AI to earn its place in the everyday running of an organisation. Before a system goes live, I keep coming back to one question: what, exactly, are we giving it permission to decide? 

 

Decision one: what AI may do on its own

Take a customer complaint. AI might summarise the correspondence, draft a reply, send it or approve a refund. These can appear as options in the same product, but they carry different responsibilities. A tool’s features should never become its job description by default. 

My starting point would be a task with a clear purpose and measurable results. Drafting replies for someone to check is one possibility. Automatic sending may be appropriate for some routine messages, once the risks and controls have been tested. 

How easily we can undo a mistake matters too. A draft can be corrected before it goes anywhere. A promise already made to a customer may need a manager to put things right. 

Let the system prove useful within those limits. Give it more freedom when the evidence justifies it, and reduce that freedom if problems emerge. Keep the decision open to revision. 

 

Decision two: how a person checks the work

Now imagine a customer who has already tried the suggested solution twice. The system sends another polite reply explaining the same steps. The wording is excellent. The customer is getting nowhere. 

“Refer complex cases to a human” is easy to put in a policy. I want to know what triggers that handover, whether the process catches those cases and who actually receives them.

Then there is the reviewer’s afternoon. If they have a growing queue and a target that rewards speed, how carefully can they check each answer? Do they have the original information, or only the AI’s summary? Can they reject a suggestion without having to defend the delay? 

People make mistakes too. Review helps only when someone has the time, information and authority to do it properly. I would examine a sample of approved answers as well as the changes reviewers make. An approval rate alone tells us little about what was checked. 

 

Decision three: who answers for the result

Before launch, someone should own the outcome and have the authority to pause the process, investigate a problem and arrange corrections. Staff need to know who that person is and who covers when they are away. 

Responsibility also extends to the customer. Someone who disputes a decision should be able to reach a person with the power to reconsider it. Automation should not leave people with fewer ways to resolve a problem. 

I would also count the work mistakes create. Suppose AI saves three minutes on each of 1,000 replies a month: 50 hours. If 50 replies are wrong and each takes 45 minutes to put right, that creates 37.5 hours of rework. Only 12.5 hours of the saving remain, before other costs. That is a different picture from the headline figure. 

For a small business, I would begin by recording these three decisions on one page. Work affecting health, livelihoods or access to services needs closer scrutiny, but everyone operating the system should understand its boundaries. 

Then test that page. Bring together the people who will run the process and follow one realistic mistake from the customer’s first message to the moment it is resolved. See where it gets stuck. Fix those gaps before expanding the system’s role. 

I want businesses to feel confident using AI well. For me, that confidence comes from knowing its limits and being able to act when it falls short. A customer shouldn’t need to understand our technology to get a problem fixed.

---

Read More Regulation

Comments are closed.