AI Process Audit: Which Three Processes Go First
Know Which Three Processes Are Worth Automating First
An AI process audit gives you a ranked shortlist of the three manual processes worth a pilot, with a reason for each. It lists the processes your team does by hand, walks the busiest ones with the people who do them, and scores each on five things: volume, how much judgement each item needs, whether the data can be reached, what a wrong answer costs, and whether someone will own the result.
The top three by score go forward, each with a measured baseline. Everything else is parked with a note saying why, so the same ideas do not come back every quarter.
This guide is for the operations manager, finance lead or IT manager who has been told the business needs to be "doing AI" and has to say what that means. The usual trigger is a board or owner mandate with no named process behind it, or a new AI role being hired with no first project to hand them.
Working out whether this fits your business? Book a free 30-minute review
Why the First Process Is Usually Picked Badly
Australian businesses are already using AI. The National AI Centre's AI Adoption Tracker found 43% of the businesses it surveyed reported some level of AI adoption between December 2025 and February 2026. Most of that is people using general tools. Far less of it is a business process that now runs differently.
The gap opens at the first decision. Three ways of choosing tend to go wrong:
- Starting from the tool. Copilot or ChatGPT licences are bought, and the question becomes "what can we use this for?" The answers are the tasks the tool demos well, which are rarely the ones holding the business back.
- Starting from the loudest request. The executive who asked first gets their process done first, whether or not it has data or volume behind it.
- Starting from the best demo. A vendor shows an agent handling a task end to end, and the business tries to find a process that looks like the demo.
Each one skips the same step: looking at the work itself. Gartner predicted in June 2025 that over 40% of agentic AI projects, its term for AI agent projects, will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls (Gartner, June 2025). Unclear value is what a badly chosen first process produces.
How the Audit Runs
The audit is short and deliberately unglamorous. Most of the time goes on watching people work and reading the systems they work in.
How a process audit turns a long list into three pilots
Collect the long list
Ask every manager one question: which work in your area is done by hand, queues behind one or two people, or gets rekeyed from one system to another? Accept everything. A long list of several dozen candidates is common, and nothing is scored yet.
Walk the busiest ones
Pick the ten or so with the most volume or the loudest complaints and sit with the people who do them. Watch real items, not the procedure document. The procedure says "check the invoice against the PO". The person doing it also checks the supplier's ABN, looks up a spreadsheet only they maintain and rings the site manager twice a week. Those extra steps are where the judgement lives, and they decide whether AI can help.
Check the data
For each walked process, find out where the inputs sit and whether they can be reached. An inbox, a SharePoint library, a line-of-business database with an API and a PDF scanner in the corner are very different starting points. A process whose inputs only exist on paper or in one person's head may need a different first step entirely.
The Five Scores
Score every walked process on the same five criteria, one to five each. Keep the definitions written down so two people scoring the same process land close together.
| Criterion | Scores 1 | Scores 5 |
|---|---|---|
| Volume and backlog | A few items a month, no queue | Hundreds a week, a visible backlog |
| Judgement per item | Pure rules, or a decision only a senior person can make | Judgement a trained person makes from patterns in past items |
| Data reachable | Paper only, or locked in a system with no export | Digital inputs reachable by API, database or export |
| Cost of a wrong answer | Wrong output reaches a customer or regulator before anyone sees it | A person checks every output before it goes anywhere |
| Owner | Nobody will own the result | A named manager wants it and will measure it |
Two of these read the opposite way to what people expect.
The judgement scale is not a straight line from easy to hard. A process that is pure rules belongs in Power Automate or a workflow tool, where it runs the same way every time without AI. A decision that needs a senior person's full context is a poor first pilot. The sweet spot is the judgement a trained team member makes from patterns: which cost code, which category, which priority. Our guide to Power Automate vs AI agents covers where that line falls.
Cost of a wrong answer is scored on the controls around the process. High-stakes work with a person approving every output can score well. The same work with no review step scores badly until one is designed in.
A Worked Example
Consider a professional services firm whose long list has eight candidates after the walks. The scores below are illustrative, to show how the method sorts them.
| Process | Volume | Judgement | Data | Error cost | Owner | Total |
|---|---|---|---|---|---|---|
| Supplier invoice coding to matters | 4 | 5 | 4 | 4 | 5 | 22 |
| Client intake and conflict check triage | 4 | 4 | 4 | 4 | 4 | 20 |
| Timesheet narrative clean-up before billing | 5 | 4 | 5 | 4 | 3 | 21 |
| Board paper summaries | 2 | 3 | 4 | 3 | 4 | 16 |
| Staff leave approvals | 4 | 1 | 5 | 4 | 3 | 17 |
| Tender response drafting | 2 | 3 | 3 | 3 | 5 | 16 |
| Archive file classification | 3 | 4 | 2 | 5 | 1 | 15 |
| Partner strategy memo drafting | 1 | 1 | 3 | 2 | 4 | 11 |
The top three by total are invoice coding, timesheet clean-up and intake triage. Each has a named owner and digital inputs.
Leave approvals scored well on volume and data but low on judgement. That is a workflow job, and it goes to whoever runs Power Automate. Archive classification scored well on judgement but has no owner and paper-heavy inputs, so it is parked with a note. The note matters: when someone raises it next year, the reason it waited is already written down.
What to Do With the Processes That Miss
Most processes fail at least one gate. The audit should say what happens to each one, so that "not now" has a reason attached.
Where a process goes after scoring
The data gate is the one most often missed. A process that scores well everywhere except data reachability is usually an integration project wearing an AI label. Getting the inputs out of the source system comes first, and our system integration work often starts there.
What the Audit Produces
An audit that ends in a slide deck of "opportunities" has not finished. The output should be specific enough that a pilot can start the following week.
A weak audit compared with a useful one
| Metric | Opportunity deck | Useful audit |
|---|---|---|
| Candidates | 20 use cases, unranked | Three ranked processes, the rest parked with reasons |
| Evidence | Workshop sticky notes | Walked items, data sources named, scores recorded |
| Starting point | Estimated benefit | Measured baseline per process |
| Ownership | "The business" | A named manager per process |
| Next step | A strategy phase | A scoped pilot with an accuracy threshold |
For each of the three, record the baseline before any AI runs. These are the numbers the pilot will be judged against.
Baseline to record for each shortlisted process
The accuracy threshold is the owner's call. It is the share of AI outputs that must match what their own team would have done before the process moves from shadow mode, where AI runs beside the team and nothing goes out, to co-pilot mode, where a person approves every output.
Timing and Who Should Be in the Room
An audit for one business unit or one function usually fits inside the four weeks of a Phase 1. A business with several sites or divisions may run one per division, so each shortlist reflects how that division actually works.
A four-week process audit
Three groups need time in it. The people who do the work, because they know the steps the procedure leaves out. IT, because data access decides more scores than anything else. And the managers who will own the results, because a process with no owner does not survive its pilot.
Keep the audit separate from tool selection. If the business has already bought Copilot licences, that is fine, and the audit may well find Copilot covers one of the three. It should not start from the licences. Where staff are already using AI tools informally, run a shadow AI inventory alongside it, because what people use unofficially is good evidence of where the manual load sits.
How This Fits an AI Strategy Engagement
A process audit is the practical core of most AI strategy work. A broader AI strategy consulting engagement adds governance, risk and the operating model around it. The audit is what keeps that strategy tied to named processes and measured baselines.
Once the three are chosen, each one goes through a fixed-price pilot with its own exit point: build, run in shadow mode against the baseline, then decide on the numbers whether to proceed. Our guide to moving an AI pilot to production covers that decision. For a finance example of one shortlisted process carried through, see automated invoice processing in Business Central.
The lead on every Solve8 engagement spent 18 years on operations and data systems for BHP, Rio Tinto and Senex Energy, gained while working with previous consulting employers. These companies are not Solve8 clients. The habit that carries over from that work is simple: watch the process being done before deciding what to build.
When the shortlisted processes need something built rather than configured, that work runs through our custom AI development service, inside your own tenant. More on how Solve8 works across strategy, automation and integration is on our homepage.
Common Questions
What is an AI process audit?
An AI process audit lists the manual processes in a business, walks the busiest ones with the people who do them, and scores each on volume, judgement, data access, the cost of a wrong answer and ownership. The result is a ranked shortlist of three processes, each with a measured baseline.
How long does an AI process audit take?
An audit for one business unit or function usually fits inside about four weeks: one to collect the long list, one to walk the busiest processes, one to score them and check data access, and one to confirm owners and measure baselines.
Why pick three processes and not one?
Three gives you a fallback. If the first pilot stalls on data access or a change in priorities, the next one is already scored, owned and baselined, so the programme keeps moving without starting the audit again.
Should a process that is pure rules be automated with AI?
Usually not. A process with one right answer at every step suits workflow automation such as Power Automate, which runs the same way every time and leaves an audit trail. AI earns its place on the steps that need a trained person's judgement.
Do we need to buy AI tools before the audit?
No. The audit should run before tool decisions, so the tools are chosen to fit the shortlisted processes. If licences are already bought, the audit may find they suit one of the three, but it should not start from them.
Get Your Shortlist
If the business has an AI mandate and no named process behind it, a process audit is the quickest way to give it one. It ends with the first pilot chosen, owned and baselined, so the next conversation with the board is about a result, not a licence.
Book a 30-minute scoping call and bring the three processes your managers complain about most. We will tell you how we would score them. The audit runs as Phase 1 of an AI strategy engagement, and the shortlisted processes carry straight into the build.