Practical guide
Before you add AI, walk through the work
Find one task worth improving, then count the effort it takes to check the result.
Show me the last one
Before choosing an AI tool, show me the last support request, proposal or report someone completed. Walk through it with the person who did it. A process diagram helps, but it often misses the workarounds that keep the process running.
Where did they find the information? What did they copy between systems? Which answer needed checking with a colleague? Where did the task sit waiting? Write down what actually happened, including the steps that never made it into the process document.
Find the part worth changing
Pick one recurring difficulty. Maybe someone spends an hour assembling information before making a five-minute decision. Maybe the actual delay is a missing approval. These need different responses. Adding a drafting tool will not fix an approval nobody owns.
Ask how often the task happens, what a good result looks like and how someone recognises a mistake. Could a template, clearer instructions or a change to the existing software solve it? Compare those options before adding another tool to maintain.
An example to test
Consider a hypothetical team preparing weekly customer updates. Each person searches project notes, collects changes and writes a draft. An AI assistant could prepare the draft from an agreed set of notes, link each claim to its source and flag missing information.
The person responsible still checks dates and commitments before sending it. The experiment ends at a reviewed draft. Sending messages, changing project dates and making new commitments would each require a separate decision about authority.
Try it on work you understand
Take five completed examples you are permitted to use. Include a routine case, a messy one and one with missing information. Record how the task works today, then run your proposed approach against the same examples. Keep sensitive material within the tools and access arrangements your organisation permits.
Count the time spent preparing inputs, reviewing and correcting the result. Record what the assistant got wrong and whether the reviewer caught it. Agree in advance who owns the trial, what would justify continuing and what failure would make you stop. Five cases can reveal problems worth investigating. They cannot establish that the workflow is ready for everyone.
When I would pause
If nobody agrees what a correct result is, sort that out first. If a reviewer cannot detect a mistake that matters, a small drafting trial cannot establish that the workflow is safe to rely on. And if the task rarely happens, maintaining the new workflow may cost more than it saves. Leaving it alone is a valid decision.
Written for this site, 4 October 2026.