What not to automate, and how to tell a real boundary from an excuse

A 40-person logistics firm wanted to automate its customer complaint responses. The volume was there. Around 30 a month, all arriving by email, all needing a reply inside a day. On paper it was a clean candidate.

The right call was to leave it manual. Not because the build was hard, but because that reply was the only moment the customer heard a person take responsibility. Everything around it got automated. The complaint itself stayed.

A mature operation has a boundary, not a percentage¶

Nobody brags about how much of their work is still done by hand. They should. A mature operation can tell you exactly what it automated, exactly what it decided to leave alone, and why. The second list is the harder one to produce.

Most advice stops at selection. Identify your first automation covers how to choose a starting point. This is the other half of the same decision. Every candidate you rule out should be ruled out on the record, with a reason someone else can challenge.

Judgement calls where the reasoning matters more than the output¶

Some work produces a decision. Some work produces a decision you can defend. Those are different products.

Pricing an unusual job. Deciding whether to extend credit. Choosing which client gets escalated this week. A tool can produce an answer. What it cannot produce is a named person who thought about it and will stand behind it. If you expect to explain the call six months later, keep the call.

Moments where the human relationship is the product¶

The reply to a complaint. The call after a missed deadline. The first conversation with a new client.

These look automatable because they have a predictable shape. They are not. The value is not the message. It is the evidence that someone chose to send it. Automate the scheduling, the reminders and the follow-up tracking. Leave the moment alone.

Work that is low volume and high variance¶

Automation pays when the same thing happens often. Twelve times a year, each one different, is the worst case you can pick. You build for the pattern you saw once, then spend the next year handling exceptions to it.

A 22-person accounting firm spent six weeks building a workflow for year-end client adjustments that ran nine times annually and needed manual correction on seven of them. They reverted to a documented checklist with a named owner. Same nine jobs, no maintenance burden, fewer errors.

When reviewing the output costs more than doing the work¶

Run the arithmetic honestly. If a task takes four minutes and checking the generated version takes five, you have not saved time. You have added a system to maintain.

This shows up most in short, high-stakes writing. Contract clauses. Regulatory responses. Anything where one wrong word is expensive. Reading closely enough to catch a subtle error is most of the work.

How to tell a real boundary from an excuse¶

From the inside, caution and avoidance feel identical. Three tests separate them.

• Can you name the specific failure you are protecting against? A real boundary names a consequence. An excuse says it is complicated.
• Would you defend it to the person waiting on the output? Say it out loud to someone the change would benefit. Excuses get quieter.
• Have you automated everything around it? A protected step inside an otherwise untouched process is not protection. It is avoidance with better vocabulary.

Fail all three and the problem is not the work. What small businesses get wrong about automation covers what usually sits underneath that.

Then write the boundary down. A decision nobody can find in six months is not a decision.

Keep exploring¶

If you are working out where your own line sits, the AI Readiness Audit produces both lists, what to automate and what to protect, or you can just contact FIT and talk it through.