1. Describe one task, not an entire department
Write down the trigger, the input and the expected output. “Improve sales” is too broad. “Prepare a CRM update from a completed intake form for a person to review” gives you a process you can test.
2. Establish the manual baseline
Observe a representative sample. Record frequency, time per task, exceptions and the corrections needed. Separate hands-on work from waiting time. A quicker draft is not automatically a faster completed task if review takes longer.
3. Assess access and risk
Check whether the necessary data is available, whether the tools can connect and who may authorise access. Ask what happens if the output is wrong. A reversible draft is usually a simpler starting point than a payment or a decision affecting a person.
4. Define acceptance before building
Specify what a successful output contains and the errors that block release. Include duplicates, empty fields and unavailable services in the test set. Keep a manual fallback and name the person who can pause the workflow.
5. Compare, then decide
Run the same kind of work through the pilot and the manual process. Count total review time, failures and maintenance effort. Continue, change the scope or stop based on the evidence. A pilot that rules out a poor investment has still been useful.
Frequently asked questions
Should the first workflow use an AI agent?
Only if choosing between actions adds value. A simpler rule-based workflow may be easier to verify and maintain.
What should I bring to an automation agency?
One task description, example inputs and outputs, the tools involved, frequency, time spent and any access or confidentiality constraints.