Which process should you look at first?
Choose one process to start with. The National AI Centre suggests a process that is high volume, so improvements have a real effect; a known pain point for staff or customers; data-rich, meaning it involves information a program could process; and contained, with a clear start and end. Customer enquiries, quoting, client onboarding, invoicing, rostering, appointment scheduling and referral processing are common candidates. Avoid starting with a process that is changing for other reasons, such as a system migration, because you would be mapping a moving target.
How do you map the current process?
Map the process as it actually runs, with the people who run it in the room. For each step, record five things: the action that starts the step, the person, role or system that does it, the inputs needed, the output produced and how long it usually takes. Add the total time per case and how many cases arrive in a typical week. The session is also where staff learn what is being considered. That matters because the people in the process are the ones affected by any change, and they are usually the first to see where a proposed change would break.
Where are the pain points?
Ask the people involved where things get stuck, what workarounds they have built and what information they wish they had at each step. Then label each issue by type, using the types in the National AI Centre's process mapping guide listed below. The labels matter because different types point to different fixes. Retyping data that already exists in another system is usually an integration problem, while inconsistency often calls for a template before anything else is considered.
- Bottleneck: work slows or stalls at a step.
- Rework: the step often has to be repeated.
- Waiting: delays between one step and the next.
- Manual data entry: typing information that already exists elsewhere.
- Inconsistency: the same task done in different ways by different people.
- Information gap or over-processing: missing data that stops progress, or more effort than the task needs.
How do you score a pain point for AI fit?
Score each pain point from 1 to 5 on the five criteria in the National AI Centre's guide: volume (rare scores low, many times a day scores high), repetitiveness (varies each time scores low, consistent scores high), data availability (no digital records scores low), error tolerance (catastrophic errors score low, easily fixed errors score high) and current cost in effort or money. Totals run from 5 to 25, and the guide treats 13 to 25 as high AI fit and 5 to 12 as low. Then rate business impact separately: an opportunity is high impact if a wrong outcome could harm people, produce unfair results or be hard to undo.
When is a simpler fix better than AI?
Low AI fit with high business impact means fix the process first. The National AI Centre lists the usual alternatives: templates and checklists, rules-based automation such as email filters and scheduled tasks, removing unnecessary steps, training, and connecting systems that do not share information. Its key rule is that if a simpler solution delivers 80% of the benefit at 20% of the cost or complexity, start there. Also check the software you already pay for, because accounting, document and customer management products can include AI features that cover the task without a new tool.
What should you write down before choosing a tool?
For each candidate, write a short scope: the intended use and what is out of scope, who is affected, what could go wrong, what would make it worth the effort and what success looks like. Then decide how people will check the output, using one of the four oversight patterns the guide describes below. Before building anything, walk the redesigned process through with sample cases, including missing or wrong inputs, and record the decision to proceed.
- Check before acting: a person approves every output, for high business impact.
- Watch while acting: the AI acts and a person monitors and steps in, for medium impact and high volume.
- Spot-check: a person reviews a sample of outputs, for low impact and very high volume.
- Exception handling: the AI handles routine cases and flags predictable edge cases for a person.