
Where AI fits in your business systems
AI fits at the manual steps where someone reads a document or an email and keys what it says into a system you already run. Supplier invoices that do not match the purchase order are the usual example. Others are customer requests retyped into the ERP or job system, quotes built from someone else's document, and tender responses assembled from past bids. Pick one of those steps, measure it as it runs today, and you have a first process you can defend to finance.
This guide is for the applications, business systems, ERP or IT manager at a multi-site organisation in Australia or New Zealand who has been asked by the CFO or the leadership team to "find where AI fits". It covers how to find candidates, four tests for choosing between them, a worked example on supplier invoice matching, and the short case to take back upstairs.
Why the question lands on the systems team
The request usually arrives without a process attached. Leadership has seen vendor demos, staff are already using public chat tools, and somebody has to turn that into a decision the business can fund.
You are well placed to answer it because you already know where the retyping happens. You see the shared mailboxes that feed the ERP, and you know which integrations were descoped and replaced by a person with two screens open.
There is a risk reason to answer it soon. The Office of the Australian Information Commissioner published guidance on commercially available AI products in October 2024. As a matter of best practice, it recommends that organisations do not enter personal information into publicly available generative AI tools. If staff are pasting customer emails into a public tool to save retyping, the manual step you are about to look for already has an unofficial workaround, and it sits outside your controls.
The shape to look for: a document comes in and someone keys it
Before you look at any product, look for one shape of work. A document or an email arrives from outside the system. A trained person reads it, decides what it means, and types the result into the ERP, the job system or the CRM.
That shape suits AI for a practical reason. Language models read inconsistent documents well, and your system of record is good at holding structured data and enforcing rules. A person sits between the two because, until recently, only a person could read a supplier's PDF or a customer's email and make sense of it.
These are the four places we would look first.
Supplier invoices that do not match the purchase order
The invoice arrives by email as a PDF. The ERP will not release it because the price, the quantity or the receipt disagrees with the order, so an accounts payable officer investigates and rekeys. This candidate gets a full worked example below.
Emailed customer requests retyped into the ERP or job system
Purchase orders, service requests and change requests arrive in a shared inbox and are typed in as sales orders or jobs. The person typing also checks the customer, the items and the price as they go. Our answer page on emailed purchase orders into the ERP covers how that step works when AI prepares the order and a person approves it.
Quotes built from someone else's document
An estimator reads a customer's RFQ, a technician's defect report or a set of drawings, then builds the quote lines in the quoting system or the ERP. The reading takes longer than the pricing. We set out what AI can and cannot do there in AI in the quoting process.
Tender responses assembled from past bids
A bid team answers each new questionnaire by searching old submissions and pasting in the nearest answer. The source documents are your own, and a bid manager reviews every answer before it is lodged. See AI for tender and RFP responses for how a first draft is built from past submissions.
Write your own list from your own systems. Shared mailboxes, import spreadsheets and the screens where staff keep a PDF open beside a data entry form are the places to start. Ask the team leaders which documents they dread on a Monday, then check the answer against what the system records.
Four tests that sort the candidates
Most systems teams can name ten candidates in an afternoon, and choosing one is the harder job. We use four tests, and a first process needs to pass all of them.
Volume
The step has to happen often enough that a saving is visible and accuracy can be measured. A few hundred items a month can give you a usable sample within weeks. A process that runs a few times a quarter cannot show you whether the AI is right, however painful each instance is.
Volume also tells you whether the result will carry across sites. A step that every branch performs in roughly the same way can be proven at one site and extended to the others.
A person who approves the output
Someone with authority already checks this work. An accounts payable officer releases the invoice, an estimator signs the quote, a bid manager approves the answer. Keep that person. The AI prepares the work and the approver decides, so a wrong suggestion is caught by the same control that catches a human error today.
This lines up with current Australian guidance. The National AI Centre's Guidance for AI Adoption sets out six essential practices. Two of them are deciding who is accountable for AI and maintaining human control. A process with a named approver gives you a starting point for both.
A measurable before and after
You need a number you can pull from the system today and pull again later: items per week, handling time per item, the share that needs rework, or the age of the queue. If the only evidence available is that people feel busy, choose another process or spend a fortnight counting first.
Prefer a measure the system records by itself, such as a status change with a date, over one that depends on staff filling in a timesheet.
Data that stays in your own cloud environment
The documents in these processes carry supplier bank details, customer names and prices. The first process should be one where the AI model runs inside your own cloud environment in Australia, reads through restricted access, and logs every result. If a candidate only works by sending documents to a public tool, it fails this test.
For New Zealand sites the same test applies. Agree the hosting region with whoever owns security before the process goes on the shortlist.
What is true of the manual step today?
- High volume, a named approver, numbers in the system, data can stay in your cloudFirst process
- High volume but nobody can produce a baselineMeasure it first, then decide
- One right answer at every step, no reading or judgementFix with a system rule or workflow
- Rare, or lives in one person's headDocument it and leave it for later
If you want a scored comparison across a longer list, our post on running an AI process audit sets out a scoring method. For a first decision, the four tests are enough.
Worked example: supplier invoices that do not match the purchase order
Supplier invoice matching is a common first process because it tends to pass all four tests. Volume is high, an accounts payable officer approves every release, the ERP already records dates and statuses, and the documents can be processed inside your own environment.
Consider a typical multi-site distributor. Each branch raises purchase orders and receives goods. Invoices arrive in a central accounts payable mailbox as PDFs, and the ERP matches the clean ones without help. The rest drop into an exception queue. The price moved, the supplier billed a carton where the order says each, a freight line was never on the order, or the branch has not entered the goods receipt.
Each exception then needs a person to open the invoice, open the order, find the receipt, work out which record is wrong and email the branch or the buyer. That reading and comparing is the manual step, and it is where the AI goes.
- Invoice arrivesA PDF lands in the accounts payable mailbox
- Invoice readSupplier, order number, lines, GST and totals extracted
- ComparedLines checked against the purchase order and goods receipt
- Cause proposedPrice, unit, freight or missing receipt, with the evidence attached
- Officer approvesAn accounts payable officer accepts, corrects or rejects the fix
- ERP updatedThe approved invoice is posted with a record of who approved it
What the AI does and what stays a rule
Keep the two apart when you scope this. Tolerances, approval limits and GST checks are rules, and they belong in the ERP or in plain code where finance can read them. The Australian Taxation Office sets out what a valid tax invoice must show, and checking those fields is a rule as well.
The AI's part is narrower. It reads invoices in whatever layout suppliers send, lines them up against the order and the receipt, and proposes the likely cause of a mismatch in a sentence the officer can check. Payment release and supplier bank details stay with people, and the AI has no access to change either.
The baseline to pull before anything is built
Most of the baseline is already in the ERP. For the last quarter, pull the number of invoices received, the number that went to exception, and the days from invoice date to release for matched and unmatched invoices. Add exceptions by cause if your system records one. If it does not, ask the accounts payable team to tag a fortnight of exceptions by hand.
That gives you a count, a delay and a list of causes, which is enough to tell finance where the time goes today. Our post on three-way match exceptions goes through the causes one by one, and the answer page on supplier invoices into the ERP covers the checks in order.
What it asks of your systems
The applications team owns this part, and it sets the pace of the work. You will need:
- Read access to purchase orders, goods receipts, the vendor master and open invoices, through restricted views or the ERP's API.
- A copy of each invoice PDF as it arrives, usually by a rule on the accounts payable mailbox.
- A test company or sandbox where draft postings can be written without touching live ledgers.
- A decision on where the model runs, agreed with whoever owns security.
The work sits beside the ERP and writes through the same interfaces an integration would use, so the list does not include an upgrade or a replacement.
How it earns trust before anyone relies on it
Run it in shadow mode first. The accounts payable team works as normal and does not see the AI's output. Each day, the proposed cause and fix for every exception are compared with what the officer did, and agreement is tracked by cause.
Finance sets the accuracy threshold in advance. When a cause category meets it, the officer starts seeing suggestions for that category in co-pilot mode and still approves each one. Categories that miss the threshold stay manual until they improve.
We use the same sequence in client work. Our fire protection quoting case study describes an engine that is built and tested, with the next stage running it in shadow mode beside the quoting team before anyone relies on it.
Who needs to agree before it goes to the CFO
A case that reaches finance with the process owner unconvinced tends to stall there. Three people should have seen it first.
The process owner comes first, usually the accounts payable manager or financial controller in the invoice example. They confirm the baseline is fair, set the accuracy threshold, and agree to own the result.
The security owner needs to agree where the model runs, what it can read and what is logged. Take them the access list from the section above. A read-only design with restricted views is an easier conversation than a request for broad access.
The people who do the work should see it too. They know which suppliers send unreadable invoices and which branches enter receipts late, and that detail decides whether the baseline is believed.
A one-page case your CFO can approve
Your CFO asked where AI fits. The answer they can act on is one process with a measured starting point, a small first step, and a point where they decide whether to spend more. It fits on one page.
- The processOne, named, with its owner
- The baselineVolume, handling time, rework, queue age
- The first stepPhase 1 on one queue at one site
- The controlA person approves every output
- The decision pointContinue, change or stop, on the numbers
Write the baseline in finance's terms. "Invoices in exception wait longer for release than matched invoices, and here are both figures" is a statement a CFO can test against the ledger. Leave vendor claims about time saved off the page. You do not have your own figure yet, and producing it is the job of Phase 1.
Keep the first step small enough that it does not need a program of work. Phase 1 covers one queue at one site: confirm the baseline, check the data and the access, and agree the accuracy threshold with the process owner. Its output is a go or no-go recommendation, and the CFO decides at that point whether to fund the shadow pilot.
- Before you startPick the processApply the four tests and name one process and one owner
- Phase 1Baseline and designMeasure the step as it runs today and confirm access to the data
- Decision pointCFO decidesFund the shadow pilot, change the process, or stop
- Phase 2Shadow pilotAI runs beside the team on live work until the accuracy threshold is met
- Phase 3Build into the systemSuggestions appear in the ERP with the approval step built in
When we run this work, each phase is a fixed fee agreed before it starts, so the amount committed before each decision point is known. Our guide to getting a first AI project into production covers how the later phases run, and the post on taking an AI pilot to production covers the funding decision in more detail.
What to avoid
Two patterns are worth avoiding because both leave you with nothing to report.
A tool rollout with no process
Licences are bought for a group of users, a few people draft emails faster, and at the next budget review the CFO asks what changed. Nobody can answer, because no process was named and nothing was measured. A general assistant has its uses, but it does not answer the question you were asked, which was where AI fits in the systems the business runs on.
A pilot with no baseline
The pilot runs, the demo looks good, and the team says it helps. Without the starting numbers, finance cannot separate the AI's effect from a slow month, so the pilot is neither extended nor closed.
| Tool rollout, no process | One process with a baseline | Impact | |
|---|---|---|---|
| What is funded | Licences for a group of users | Phase 1 on one named process | Scoped first step |
| Who owns the result | Nobody in particular | The process owner | Named owner |
| Evidence at review | Usage counts and opinions | Before and after figures from the ERP | Testable |
| Where the data goes | Depends on each user | Your own cloud environment | Controlled |
| Next decision | Renew or cancel licences | Fund the pilot, change it or stop | Clear choice |
When the answer is a system fix
Some candidates fail the tests for a good reason: the step has one right answer and the system should enforce it. A mandatory field that was never made mandatory, or a price that should come from the price list, is a configuration or integration job. Those jobs usually cost less than AI and you can often do them in-house.
One of our published case studies shows the split. A major Australian wholesale business asked where AI could improve its operations. We checked how its systems handle orders, purchasing, stock and compliance first, then wrote a staged AI plan around what we found. The result was a clear split between the errors a system fix solves and the ones worth AI. The wholesale AI plan case study has the detail.
For an applications manager this is useful. A shortlist that says two items are system fixes your team can do and one is worth AI is more credible to a CFO than a list where everything needs AI. Where a candidate does need something built around your systems, custom AI development is the route.
How Solve8 can help
Our consultants and engineers work with applications, business systems and finance teams at multi-site organisations in Australia and New Zealand. We help you pick the first process, measure the baseline and build the AI into the systems you already run. The work runs in fixed-fee phases agreed before we start, a person approves every output, and the data stays in your own cloud environment.
The first phase is described on our AI strategy page. To talk it through, book a free 30-minute review and bring the manual step your team complains about most. We will go through the four tests with you and say plainly whether it is worth a Phase 1.
Related Reading
- Power Automate vs AI Agents: Rules Where They Fit - how to tell a rules job from a judgement job before you choose a tool
- Automated Invoice Processing in Business Central - what the built-in invoice features do if Business Central is your ERP
- What an AI Strategy Consulting Engagement Gives You - what the written output of a strategy engagement looks like
- Where to start with AI across many sites - piloting at one site and rolling out site by site
Sources: Office of the Australian Information Commissioner, Guidance on privacy and the use of commercially available AI products, October 2024. National AI Centre, Department of Industry, Science and Resources, Guidance for AI Adoption. Australian Taxation Office, tax invoice requirements.
