This is the exact sequence we walk clients through on paid engagements — free, because most businesses can do the first four steps themselves.
Before touching any tool, find out where the hours actually go. Most businesses guess wrong.
For one normal week, have each person jot down what they spend 30+ minutes a day on. You're hunting for tasks that are frequent, repetitive, and involve reading or writing — answering the same emails, re-keying data between systems, drafting near-identical documents.
Trap:starting with the tool (“we bought ChatGPT, now what?”) instead of the task. Tools in search of a problem become unused subscriptions.
The best first candidate is high-volume, low-stakes, and easy to check.
From your audit, choose a single workflow where a mistake is cheap and easy to catch — drafting replies to routine enquiries, summarising meeting notes, writing first-pass product descriptions. Explicitly rule out anything involving money movement, legal commitments, or medical/safety decisions for the pilot.
Trap: piloting three things at once. One workflow, done properly, teaches you more than three done halfway — and gives you a clean before/after number.
One assistant licence, one automation tool if needed, human checks every output.
Start with a general assistant (see our reviews for current picks) and write a reusable prompt for your workflow — include two or three examples of great past output. Keep a human approving every AI output for the full 30 days, and log the time both before and after.
Trap:feeding client-confidential data into consumer AI accounts. Use business tiers with training turned off — check our reviews' data-handling scores before you paste.
At day 30, the numbers decide — not the enthusiasm.
Compare hours saved against the fully loaded cost: subscriptions plus the checker's time. A good pilot typically shows 30–60% time reduction on the target workflow. If it's under 20%, the workflow was a bad fit — that's a finding, not a failure. Pick the next candidate from your audit and re-run.
Roll the win out, loosen the checks gradually, and only then look at custom systems.
Extend the working pilot to the whole team, move from checking every output to spot-checking, and connect it to neighbouring workflows with an automation tool. When you hit something off-the-shelf genuinely can't do — high volume, deep integration with your systems, decisions that need your business logic — that's when a custom build is worth pricing.
Trap:jumping to a custom build before exhausting off-the-shelf. Custom systems start around $15k — they're the right answer only when volume justifies it. We'll tell you honestly: ask us.