
Date:
Author:
A2.Ai
001
Shortlist three recurring tasks
Start with a repeated piece of work your team can explain and measure. Examples include routing an enquiry, collecting required documents or preparing a draft from approved information.
Describe each candidate in one sentence: when this arrives, we take these steps to produce this outcome. Name its owner, the systems involved and what happens when information is missing. If the current process is unclear, map it first.
Business.gov.au recommends identifying the problem first and trialling AI in one or two areas. Read its practical introduction to AI.

002
Score effort and readiness
Score each task from 0 to 2 on the five factors below, then add the points out of 10. This is our planning aid, not a validated assessment, safety approval or forecast of savings.
Frequency: less than monthly = 0; monthly = 1; weekly or more = 2.
Weekly handling time: under 1 hour = 0; 1–5 hours = 1; over 5 hours = 2.
Rules: unmapped = 0; partly defined = 1; documented and repeatable = 2.
Exceptions: unknown = 0; recognised but no clear owner = 1; recognised and routed to an owner = 2.
Review: difficult to verify = 0; needs specialist judgement = 1; a reviewer has a clear, repeatable check = 2.
Apply a separate risk check first: unresolved privacy risks, harmful errors or irreversible actions stop the pilot, whatever its score. Keep payment approval and final acceptance with the responsible person.
Worked example: two hypothetical tasks
Enquiry routing: 30 enquiries × 3 minutes = 90 minutes of handling a week. With documented rules, owned exceptions and a clear review check, scores of 2 + 1 + 2 + 2 + 2 total 9/10.
Document drafting: 10 drafts × 40 minutes = 6 hours 40 minutes a week. With partly defined rules, unclear exception ownership and specialist review, 2 + 2 + 1 + 1 + 1 totals 7/10.
In this invented comparison, routing is the more ready first pilot despite the smaller workload. These are starting workloads, not hours saved. Measure actual review, correction and maintenance effort before estimating any benefit.

003
Define the human review boundary
Use a straightforward rule where it is sufficient. AI may help interpret unstructured information, but that uncertain step should have its own checks.
In our recent work, connected reports helped prepare a draft and supporting document pack. Reviewers could request corrections and receive a resubmission before approved handover. Late information and some exceptions still needed manual attention.
Before building, agree what information is required, what sends work to a person, who owns that queue and how the action can be corrected or stopped. The National AI Centre’s foundations guidance includes accountability and human oversight appropriate to the use.
Read the National AI Centre guidance on AI adoption foundations.
004
Test one bounded workflow
Observe a representative sample of the current process, including exceptions. Record handling time, waiting time, corrections, completion and review effort. Compare similar work in the pilot.
Agree success and stop criteria in advance. Include software, training, review and maintenance costs, not just time saved in one step. Keep a route back to the previous process. The National AI Centre’s ROI guide explains what to measure.
Choose one owner and one clear outcome. Expand only when the pilot shows the workflow is useful and manageable.
See how to handle incomplete paperwork
Book a free 30-minute discovery call to choose a practical starting point.



