The most common AI mistake we see: someone tries to automate an entire job. 'Have Claude handle customer support.' 'Use ChatGPT for sales calls.' Nine times out of ten, it fails — not because the model can't, but because a job is many loops, and only some of them should be automated.
What a loop is
A loop is a repeatable unit of work with a clear input and a clear output. 'Reply to a lead' is not a loop — it's a job. But 'classify this reply as interested / rejected / unclear' is a loop. Input: email. Output: enum. That's automatable.
Zoom in on the job you'd like to automate. Break it into loops. Look at the loops. The right AI candidate is the loop that (a) happens most often, (b) has the tightest input-output shape, and (c) is the least fun to do manually.
A worked example
Our agency's sales job has maybe fifteen loops. Finding leads. Sending outreach. Reading replies. Classifying them. Following up. Booking calls. Doing calls. Sending offers. Closing deals. And so on. Which of these did we automate?
Finding leads: yes — a nightly Playwright scraper. Highly templatable, no judgment involved.
Sending outreach: yes — Loops transactional API fires the moment a lead is scraped. Zero human judgment per lead.
Classifying replies: yes — Claude reads each reply and returns { classification, summary } in Slovene. This is the highest-leverage automation we've built. It removed two hours a day of email triage.
Booking calls: no. This needs judgment — timing, tone, availability. A human does it in thirty seconds.
Doing calls: obviously no. This is what the whole system exists to protect.
The rule
Automate the loops that make the humans faster at the parts only humans can do.
What to avoid
AI-generated customer support that can't say 'I don't know'. AI-drafted contracts nobody reviews. AI decisions nobody audits. The rule: if the loop's failure mode is embarrassment or legal exposure, keep a human in it. If the failure mode is 'we missed a signal', let the AI take a first pass and log everything.
That last point is why our AI classifier logs every classification alongside its outcome. Six months later, we can join Claude's decisions to actual revenue and prove which classifications convert. The AI isn't a black box — it's an auditable pipeline stage.