Which business workflow deserves your first AI test?

Imagine a small service business with a crowded enquiry inbox. Staff read each message, ask for missing details, decide who should respond and draft the reply. It looks like a job for AI. But the contact form asks only for a name and an open text box. The first improvement might be three better questions and a clear routing rule.

That example is hypothetical. It shows the decision an owner needs to make before choosing a tool: which complete job is worth improving, and what kind of change would improve it? If the form fix removes the friction, there may be no AI project. If varied exceptions remain, a small, human-reviewed test could be justified.

NAVE's five decisions before buying another AI tool help when a product is already on the table. This step comes earlier. It helps you choose the work, including the possibility of choosing no AI use yet.

Start with the whole job

Write down where a workflow starts and when it is genuinely finished. For enquiry handling, “AI drafted a reply” is not the finish line. A useful end point could be a complete, accurate response sent by a person and recorded in the right place. Count the missing-information chase, checks, corrections and handoff too.

The National AI Centre's process-mapping guide recommends starting with a contained process that happens often enough to matter and has a real pain point. Ask the people doing the work where requests wait, where information is missing and what they already do to work around it. That conversation is more useful than collecting a long list of AI tools.

Shortlist no more than three candidate workflows. For each, note how often the job happens, what delay or rework it creates, and who is affected if it goes wrong. Use the same unit for each candidate: one complete enquiry, quote, booking or document request, rather than a single step in isolation.

Put the owner decision on one page

The shortlist does not need a numeric score. For each candidate, answer these questions in plain language:

The National AI Centre's opportunity guide also asks businesses to compare AI fit with impact and to try simpler changes where they make more sense. The NAVE shortlist turns that general advice into a decision for one owner with limited time and review capacity. It is a prompt for judgement, not a formula that makes the choice for you.

Let the simpler fix compete

Return to the hypothetical inbox. A structured contact form could ask for service needed, timing and location. A response template could help staff request missing information consistently. Ordinary rules could route messages by service area. These changes may be easier to maintain and explain than an AI system.

If the remaining work still involves varied descriptions, unclear intent or information spread across several places, there may be a case for a narrow AI test. For example, AI could draft an internal summary of a non-sensitive test enquiry for a person to check. It would not decide whether to accept the customer or send a reply on its own. The owner would compare the complete handling time and quality with the improved non-AI process, not with the old broken form.

That comparison matters. An AI demo can look efficient because it skips the work a real team has to do around it. A form fix and a supervised trial should compete on the same end result, including checking and rework.

Check the boundary before a tool trial

Ask what information the proposed workflow needs. If customer or staff details would enter a product, assess the data flow and product before using real records. The OAIC's guidance advises due diligence and particular caution with personal information in publicly available generative AI tools. The right obligations and controls depend on the business and the use.

Use the National AI Centre's screening questions early to spot uses that may need stronger controls. They are a starting check, not an approval. Decide who reviews outputs, what happens when the inputs are incomplete, and who can pause the test. The Centre's workflow guide recommends testing exception cases before building further. business.gov.au likewise advises checking that an AI tool works as expected and understanding how it handles data.

If you cannot describe a safe test boundary or assign a reviewer, defer the AI trial. That is a useful decision. It gives the business time to repair the workflow, clarify information handling or obtain specialist advice where needed.

Choose one of four next steps

1. Test AI when the remaining job is repeated, variable enough to benefit, safe to trial with approved information, and small enough for a person to check.
2. Fix the process first when missing information, unclear ownership or unnecessary steps create most of the delay.
3. Use ordinary automation when clear rules can move work between people or systems without asking a model to interpret it.
4. Defer when the consequence of error, information handling or supervision needed is not yet understood.

Set a date to revisit the chosen workflow. If you do run a pilot, record the starting point and use NAVE's four-week pilot scorecard to judge the complete result. The first AI test is worthwhile when it helps the owner learn whether one real job can improve under control. A decision to use a simpler method can be equally valuable.