In a growing business, AI earns its keep in the back office where the input is messy and the work repeats: reading supplier invoices, delivery dockets and purchase orders out of PDFs and emails, sorting inbound requests to the right team, matching documents to jobs or orders, drafting summaries and exception reports, and checking records for gaps. It does the first pass and a person approves. It is the wrong tool where a simple rule already gives the right answer every time, and it should never be the one approving a payment, settling a dispute or replying to a customer on its own. The practical way in is one well-chosen task, connected to the systems you already run, with your IT team involved from the start and a clear before-and-after measure.
The week it's meant to fix
Picture the accounts payable inbox on a Monday: supplier invoices as PDFs and phone photos, delivery dockets from drivers, a customer purchase order typed into the body of an email. Someone opens each one, types the supplier, amounts and order number into the finance system, then checks the job or order in the operations system. When it doesn't match, an email chain starts. The order number already exists in your system, and someone is re-keying it from a PDF. That's double entry (the same information typed into two or more systems), and we explain it in plain terms in what is double data entry. If it sounds like your office, our page on manual data entry killing productivity covers the wider pattern.
Where AI does the job well today
Current AI tools are genuinely useful here, provided a person checks the result.
- Reading documents. Supplier invoices, delivery dockets, purchase orders, remittances and statements, whether PDF, scan or email body. AI pulls out the supplier, dates, order numbers, line items and totals, and copes with layouts older template-based scanning couldn't.
- Sorting and routing inbound email and requests. Invoice, quote request, delivery query or order change? AI tags it and sends it to the right queue, leaving anything it's unsure about for a person.
- Matching documents to jobs and orders. Once a docket or invoice has been read, AI can suggest which job, order or purchase order it belongs to when the reference is missing or written differently ("PO 4471", "your ref 4471", "Smith St job").
- Drafting summaries and exception reports. "Here are this week's invoices that don't match a purchase order, and why." A manager edits a draft instead of starting from blank.
- Checking data for gaps. Jobs with no hours booked, supplier records missing an ABN, timesheets that don't match the roster.
If you work in production, our practical AI uses for manufacturers goes into the factory-floor versions of these tasks.
Where a simple rule beats AI
Plenty of what gets sold as "AI" is better done with a plain rule: cheaper, faster and the same answer every time.
- Anything with a fixed answer. If the invoice total is within an agreed tolerance of the purchase order, it matches. If a timesheet has more than the rostered hours, flag it. You don't need AI to compare two numbers.
- Data that's already structured. If an order arrives through a portal or electronic feed, move it with an integration, not a model that reads it back.
- Calculations. Pricing, GST, freight, pay rates and margin are formulas. AI should never be doing your arithmetic.
A good design often uses both: AI reads the messy document, and rules check what it read. That combination is what we usually mean by workflow automation.
What should stay with a person
Some decisions should keep a named person approving them, however good the tool gets.
- Approving payments and changing bank details. AI can prepare the run and flag anything odd; a person approves it. A changed bank account should always trigger a call to a known number.
- Coding to the accounts. AI suggests the account, job or cost centre; a person confirms it.
- Replying to customers. AI can draft a reply for staff to read, edit and send. We don't build customer-facing bots, and we don't recommend letting AI send anything to a customer unread.
- Judgement calls. Credit decisions, disputes, staff matters and anything with a legal or safety consequence.
What larger teams need to sort out first
With an IT team, a finance controller and a board involved, answer these before anything goes live.
Where the data goes. Every AI tool sends your data somewhere to be processed; find out where, and whether you can choose the region. The Office of the Australian Information Commissioner's guidance on commercially available AI products (October 2024) advises organisations not to enter personal information, and particularly sensitive information, into publicly available AI tools, and to do proper due diligence on any product you choose.
Business plans versus consumer plans. This matters more than which brand you pick. Consumer accounts (the free and personal paid plans) can generally use your conversations to improve the vendor's models unless someone switches that setting off. Business and enterprise plans are usually covered by commercial terms that exclude your data from training by default. According to the vendors' own published terms at the time of writing, that's the case for OpenAI's ChatGPT Business, Enterprise and API, Anthropic's Claude Team, Enterprise and API plans, and Microsoft's Copilot for work accounts with enterprise data protection. Terms change, so your IT team should check the current version for whatever you use. The real risk is staff pasting contracts or customer details into personal accounts because the business hasn't given them a proper one.
Security review. Bring your IT team in at the start. They'll want to know what the tool connects to, whether it can write or only read, how it signs in, where logs are kept and how access is removed when someone leaves.
An approval step that's recorded. For anything touching payments, coding or customers, the system should show who approved what and when. That's an audit trail, and your auditor will ask for it.
Documenting how it works. Write down in plain English what the tool does, what a person checks and what to do when it's wrong, so it doesn't depend on whoever set it up.
Picking the first task
Pick a task that ticks all three:
- It repeats. The same kind of document or request, many times a week.
- The input is messy. PDFs, scans, emails, free text. That's what AI is good at, and why the work is still done by hand.
- A person can easily check the result. Someone already knows what right looks like, and checking takes seconds rather than redoing the work.
Supplier invoices against purchase orders and dockets against orders are common first choices.
Measure before and after. Before you start, count the documents, the time spent on them, the errors found later and how long each takes to reach the right system. Count again after a few weeks. If nothing moves, stop or change the task. Your own count matters, not the vendor's.
Connect it, don't add another island
AI most often disappoints by becoming one more separate tool: the invoice is read beautifully, then someone copies the result into the finance system by hand.
The AI step should sit inside the flow of work: the document arrives, AI reads it, rules check it against the order in your operations system, a person approves the exceptions, and the approved record lands in your finance system without being retyped, whether that's Xero, MYOB, NetSuite, an ERP or a job management system. Each piece of information is captured once and flows to where it's needed, so the numbers people report from come from one agreed place, a single source of truth (one place everyone takes the number from). We explain that idea in single source of truth in plain English. Without the connections, you end up with data silos (information stuck in one system that the others can't see), which we cover in why your systems disagree.
What it looks like when it's fixed. On Monday morning, the accounts team opens a short list of invoices that didn't match a purchase order, each with the reason beside it, instead of a full inbox. The operations manager can see which deliveries have been received but not yet invoiced, because the docket was read and matched to the order the day it came in. Nobody typed either one.
When it's worth getting help
You can trial a document-reading feature in software you already pay for without help, and for one supplier or one inbox that's often enough. Bring someone in when the task crosses systems, when your IT team needs a proper design to review, or when the process itself isn't settled. Automating a messy process just makes the mess faster; our business process improvement guide covers sorting that out first.
Our AI consulting work starts by finding the tasks where AI genuinely helps and saying plainly where it doesn't, and our AI workflow automation work builds the connected version, with the approval steps and records described above. If you want the whole picture mapped first, the $5,000 Process & Systems Audit maps how work and information move through your business and produces a written report on where automation and AI would and wouldn't pay off.
Frequently asked questions
Is it safe to put our supplier invoices and customer data into AI tools?
It can be, on a business plan with commercial terms that exclude your data from training, a clear answer on where data is stored, and access your IT team has reviewed. It isn't safe on staff members' personal accounts. Follow the OAIC's guidance and keep personal and sensitive information out of publicly available tools.
Will AI replace our accounts payable or admin staff?
It mostly changes the work. The typing and first-pass matching go; the checking, chasing and judgement calls stay, and they need experienced people.
How accurate is AI at reading invoices and delivery dockets?
Good on clear documents, weaker on poor scans and handwriting. That's why rules check what it read against the order and anything uncertain goes to a person. Test it on your own documents.
What should our IT team ask an AI vendor?
Where data is processed and stored, whether it's used for training, what access the tool needs to your systems, how sign-in and offboarding work, what's logged, and what happens to your data if you stop using the product.
If you'd like to talk through where AI would genuinely help in your operations, and where it wouldn't, book a free initial chat.