One Wrong AI Sentence Means Rework; One Wrong Automated Email Can Become an Incident
Last Wednesday afternoon, I helped a freelancer friend set up a customer-message routing tool. AI classified “I don’t need your service right now” as “Send me a quote immediately,” so the system sent a promotional email to someone we shouldn’t have contacted. I stared at that email and froze for a moment. I had only wanted to save some time—how did it go so wrong? Later, I realized the problem wasn’t necessarily that AI wasn’t smart enough. We had simply given it too much room to improvise.
A Different Approach: Don’t Let AI Write the Answer—Let It Make a Choice
This approach is often called a “System 1 model.” Despite the name, I think of it as a fast triage worker: it takes an input, selects one option from a predefined list, or returns a score and probability. It doesn’t write long articles or invent workflows on the fly, so it’s often faster, cheaper, and more consistent across repeated runs.
TypeSafe AI co-founder Diogo Almeida showed a similar approach at the AI Engineer Code Summit in New York. His comparison really stuck with me: ordinary AI is like “Einstein and Mr. Bean randomly taking over.” If the business side can’t tell whether a given result is good or bad, it’s hard to trust a program to execute it automatically. That also explains where this approach fits: routing customer messages into presales, support, or hold; sorting orders into standard handling versus manual review; or selecting one of several fixed replies.
Replicate It Today: Build a Minimal Classifier for $0
Money: An open-source model or an existing computer can make this free. Time: Start with 100 real examples, which takes about two hours. Technical barrier: You need to be able to copy a configuration and run the example; coding isn’t required. First step: Open your content tool, find its “classify, tag, or score” feature, and create a classification task with only three options: “Presales,” “Support,” and “Manual Review.” Test it with 10 old messages. Don’t connect it to your website or let it send emails yet—keep the output in a spreadsheet so you can inspect it. Not everyone needs this tool. If you only occasionally write copy and handle fewer than ten messages a day, there’s no rush to try it.
Match the Experiment to Your Current Workload
Just starting out: If you receive fewer than ten customer messages a day, I’d keep handling them manually for now. Once similar messages pile up into the dozens, I’d build a small three-way classifier so the maintenance effort doesn’t exceed the time saved.
Working with one or two clients: If you’ve started thinking, “Copying my previous reply might leave something out, but rewriting it from scratch is slow,” I’d let it handle classification only, while I confirm every final reply. I’d also keep an escape hatch that sends unclear cases to a human.
Scaling up: If several people are repeating the same first-stage screening, I’d record why each manual correction happened, then hand the stable rules to the program. I’d begin with low-risk steps, such as organizing tickets, tagging, and setting priorities, while keeping human review for payments, refunds, and public commitments.