Where AI actually pays for itself in an Australian business, and where it does not
An even-handed look at the tasks AI does well today, the ones it does badly, and the four questions to answer before you spend anything.
Ask what AI can actually do for a small Australian business and you get two unhelpful answers. One says everything, from a vendor. The other says nothing, from someone who tried a chatbot once and got a confident wrong answer.
The useful answer is narrower and more boring. There is a short list of tasks where current AI earns its cost reliably, a longer list where it does not, and a set of preconditions that decide which list your work falls into. The preconditions matter more than the technology choice, and almost nobody selling to you will raise them first.
What follows is both lists, plus the four questions — data quality, review, cost, and where the data goes — that decide whether any of it works in your business. The Australian angle is not decoration: where data is processed and stored is a live question for buyers here, particularly in regulated work.
We build AI into systems Australian businesses already run, so we are not neutral. We do have to keep things working after they go live, which makes optimism expensive.
Where it pays: reading documents and pulling out fields
The strongest case, by a wide margin. Supplier invoices, remittance advices, delivery dockets, purchase orders, insurance certificates, plant certifications, signed site paperwork that comes back as a photograph. Anything that arrives as a document and gets retyped into Xero, MYOB or an ERP.
It works because the task is bounded — you know which fields you want — the output is checkable in seconds by the person who used to type it, and the comparison is against a human doing repetitive transcription late on a Friday. It also measures cleanly: you can test extracted fields against documents your team already keyed, before committing.
It is worth doing when the volume is there, dozens of documents a day rather than a handful a week, and it holds up when formats are inconsistent, which is where rules-based scanning falls over. More on that on our document and knowledge intelligence page.
Where it pays: classifying and routing what arrives
Support email, warranty claims, incoming site paperwork, supplier correspondence, job enquiries. Anything that arrives as a stream and has to be sorted before anyone can act on it.
Sorting into a small number of well-defined categories is among the most reliable things current models do. It pays not because classification is hard, but because the sorting currently happens in one person's head, one item at a time, first thing in the morning, delaying everything behind it.
Keep the routing rules yours, visible and editable. The model suggests a category; the logic deciding what happens next should stay in plain configuration your team can change. Where the categories follow a fixed rule with no judgement in them, skip the model and use workflow automation — cheaper to run, easier to audit.
Where it pays: first drafts of repetitive correspondence
Quote covering letters, standard customer replies, incident summaries, shift handover notes, supplier chasers, tender boilerplate. Work that has a house style, gets written many times a week, and is edited before it leaves.
The value is specific: never starting from a blank page. A person still reads and sends. If nobody is reading and sending, you have not automated correspondence, you have automated the risk of writing something wrong to a customer under your own letterhead.
Where it pays: summarising long records
Reading a 40-page contract and pulling out the terms that matter, condensing a year of maintenance history before a shutdown, summarising a long email thread for someone joining it late, digesting a claim file. Useful when a person will act on the summary and can open the source if something looks wrong. Insist that summaries cite the passage they came from — an uncited summary is unverifiable, and unverifiable output is where trouble starts.
Where it pays: flagging anomalies for a human
Pointing at what looks unusual: a price that moved, a stocktake variance outside the normal range, a timesheet with impossible hours, an invoice that does not match the purchase order, a supplier that suddenly changed bank details.
Note the framing. Flagging, not deciding. The output is a queue for a person, and the measure of success is whether the queue is short enough that someone actually works through it. A flagging system that raises 200 items a day gets ignored within a fortnight, which is worse than not having it.
Where it does not pay, or is not safe yet
Be as specific about this half.
- Anything final that touches money, safety, employment or a legal commitment. Not because the model is always wrong, but because the cost of an occasional confident error is asymmetric. Pricing, pay calculations, safety sign-offs and contract decisions need a person in the loop, and building that review step is part of the job.
- Exact arithmetic over many records. Models are unreliable calculators. Totals, reconciliations and margin calculations belong in code or a database query; use the model only to get data into a structured state first.
- Low-volume, high-variation tasks. If something happens twice a month and looks different each time, there is nothing to learn from and the build cost will not be recovered.
- Anything where you cannot describe what "right" looks like. If you cannot produce fifty past examples of the correct answer, you cannot measure the system — and an unmeasured AI system is a hopeful one.
- Judgement your customers are paying for. In professional services the advice is the product. Drafting support is fine; advice without review is a liability question, not an efficiency one.
- Situations where a wrong answer is invisible. If nobody would notice an error for six months, do not put a model there.
The precondition: your data has to be good enough
This is the step that quietly kills more AI projects than anything technical. If the same customer exists three times under three spellings, a model treats them as three customers. If two systems disagree about stock on hand, no model resolves the disagreement — it just produces confident output from whichever one it was pointed at.
You do not need a full data clean-up first, and anyone insisting on one is selling you a longer project. You need the specific fields the AI touches to be consistent, and you need to know which system is authoritative when two disagree. That is usually a few weeks of targeted work, not a year. Where the answer is that nothing is authoritative, the honest next step is integration between the systems rather than AI on top of them.
The precondition: who reviews, and what happens to overrides
Decide before you build: which outputs a person checks, who that person is, what they see when they check, and what happens when they disagree.
The last part is the one people skip. Overrides are the most valuable data the system produces. If your team corrects the same field 30 per cent of the time, that is a defect report, and it should be visible as a number someone looks at monthly. A system with no override tracking cannot improve, and you will not know it has drifted until a customer tells you.
What it costs, honestly
Three costs, and vendors usually quote one.
Build. Getting it working against your real data and your real systems, not a demo. Usually the largest single number.
Running it. Model usage is metered, so per-document or per-request costs are real and they scale with volume. They are typically small relative to labour, but get the number before you commit, and get it at your actual volume rather than a sample.
Keeping it working. Suppliers change document formats. Models get retired and replaced by versions that behave slightly differently. Your process changes. Someone has to notice accuracy sliding and fix it. A proposal with no line for this understates the cost — it is much of why Drawn AI bills a monthly subscription rather than a project fee.
Against all that, price the alternative properly. The comparison is not with zero. It is with the hours currently spent, the errors currently made and the decisions currently delayed.
Where the data goes, and why it matters here
For Australian buyers this is often the deciding question, and it deserves a direct answer rather than a page of assurances.
Ask any vendor four things. Which country the data is processed and stored in. Whether your data is used to train models available to anyone else. Who at the vendor can see it. What happens to it when you stop paying.
Anything Drawn AI hosts for clients sits in Australian regions — Azure Melbourne and AWS Bedrock Melbourne. That matters where contractual data-residency obligations apply, which is common in mining services and resources supply chains and in professional services handling client records, and it removes a procurement conversation that can otherwise stall a project for months.
Worth knowing too: staff are probably already pasting company information into consumer chatbots because it is faster than the official process. That happens whether or not you have an AI strategy, and it is a stronger argument for providing a sanctioned tool than any efficiency case.
How to decide, in one paragraph
Pick one task. It should happen many times a week, have a checkable answer, cost something real when it goes wrong, and have fifty past examples you can test against. Measure the current cost in hours and errors. Build the smallest version, measure it against those examples before go-live, and put a person on the output. If it does not clear that bar, do not start with AI — the honest answer is often a report, a rule or an integration, and those are cheaper.
Frequently asked questions
Usually document extraction — invoices, remittances or dockets going into Xero, MYOB or an ERP without retyping. It has clear volume, a checkable output, a measurable baseline, and it does not require anyone to change how they work. It also tests your data quality honestly before you commit to anything larger.
It depends on the task. Sorting into a handful of clear categories is usually strong. Pulling fields from inconsistent documents is good but never perfect. Free-text generation is the least predictable. The only number worth trusting is one measured against your own historical decisions before go-live, so ask for that rather than a vendor benchmark.
Drawn AI works on a monthly subscription — three-month minimum term, billed monthly, 30 days' notice to cancel — covering build, integration, hosting and ongoing tuning, so no separate maintenance bill arrives later. The figure depends on scope and how many systems are involved. Whatever the provider, insist on seeing metered running costs at your real volume.
Partly, and less than you would fear. The fields the AI touches need to be consistent, and you need to know which system wins when two disagree. That is normally weeks of targeted work on specific fields, not a full data programme. Anyone requiring a complete clean-up first is proposing a much longer project than you asked for.
Ask every vendor directly and get it in writing. Anything Drawn AI hosts sits in Australian regions, and client data is not used to train shared models. If a provider cannot tell you the processing region, the retention period and what happens at cancellation in plain terms, treat that as the answer.
For the work described here, the realistic outcome is that repetitive transcription and sorting shrink while review, exception handling and judgement grow. That changes what a role contains rather than removing it. Anyone promising headcount reduction as a guaranteed outcome is describing a sales pitch, not a system.
Want to talk it through before committing to anything?
A first conversation costs nothing and usually ends with a clearer idea of what is worth building — sometimes that answer is “not yet”, and we will say so.
- No obligation and no sales sequence
- Built around your existing systems
- Australian-based, Australian-hosted data
- 3-month minimum, then 30 days’ notice
No lock-in templates. Built around you.