The best first candidate is often work someone repeats every week: copying enquiries, assembling documents or chasing missing information. Start with one such task rather than rebuilding the whole business.
Where work starts repeating
An illustrative example: a customer submits an enquiry through a form. A colleague copies it into a spreadsheet, assigns a salesperson and sends a confirmation. A missing phone number or attachment starts another round of emails. This is a model process, not a result from a particular client.
Break it into four steps: receipt, validation, assignment and confirmation. Each needs a clear outcome. “Handle the enquiry” is too broad; “record a complete enquiry and assign it by region” can be checked.
Choose a process with clear rules
A good candidate repeats, has accessible inputs and produces an outcome you can verify. Assigning an enquiry by region needs a rule. Deciding whether to accept an unusual project needs human judgement.
Watch the exceptions. If every other enquiry needs special handling, improve the form or agree on rules first. Automating an unclear process only makes the confusion faster.
Calculate time without promising miraculous savings
Example: 80 enquiries per month × 6 minutes of copying = 480 minutes, or 8 hours. This is the time spent on the current task, not a guaranteed saving. Allow for checking, exceptions and maintenance.
For one working week, record case volumes, working time and corrections. Separate active work from waiting for customers. Otherwise you will attribute savings to the solution that it cannot deliver.
What a small first pilot looks like
The form creates a record, checks mandatory information and assigns an owner. The customer receives a confirmation. Incomplete enquiries remain visible in a queue for follow-up; the system must not silently discard them.
Run the pilot on a limited scope first. Compare original submissions with records, test duplicate submissions and a destination-system outage. Agree who receives alerts and how processing is retried.
When we would hold off on automation
For three enquiries a month taking two minutes each to copy, a simpler form might be a better investment. Also wait if rules change frequently or nobody can say who owns the outcome.
Define access and the scope of sensitive data before connecting tools. Not every process needs AI: predictable integrations are often better suited to fixed rules.
What to bring to the first conversation
Bring an anonymised input example, the desired output, your tools and the three most common exceptions. Add monthly volume and someone who knows the process.
This lets us define a bounded pilot: what it handles automatically, what it passes to a person and how you decide whether to continue.