The AI Capability Ledger — what AI can and can’t do in the SME back office

AI learns how transactions have been coded before and suggests the account and GST code for each new bank line. For routine, repeating transactions — subscriptions, fuel, regular suppliers — suggestions are consistently usable and improve with more data.

Constraint: Reliable for the routine majority only; a person confirms, and unusual transactions must route to review.

AI extracts supplier, date, amount and GST component from photographed receipts and emailed invoices into structured ledger data. One of the most mature applications — it removes most manual data entry from day-to-day bookkeeping.

Constraint: Extracted data still enters the review layer; the extraction is mature, the classification decisions downstream are not always.

AI suggests matches between bank transactions and the invoices, bills and expense claims in the system, including harder cases like partial payments and combined deposits. A person confirms the match rather than hunting for it.

Constraint: Confirmation stays human; unmatched exceptions still need a person to resolve.

Having learnt the normal pattern of a ledger, AI flags what breaks it: duplicate invoices, unusually large payments, transactions coded inconsistently with their history. It surfaces questions a busy person would miss.

Constraint: A flag is a question, not a finding — a person decides what each flag means.

For common, well-understood purchases AI can suggest the GST treatment as it codes the transaction, which materially speeds up BAS preparation. It is a first draft for review, not a final answer.

Constraint: TPB(GS) 55/2026 requires AI output to be assessed and supplemented by professional judgement before being relied on.

Mixed supplies, GST-free food line items, exports, insurance settlements, second-hand goods — these treatments depend on facts the software cannot see, and AI trained on routine transactions will confidently misclassify the unusual ones.

Constraint: Wrong classifications present with the same confidence as right ones; the TPB notes AI models may hallucinate or produce inaccurate information.

Whether money drawn from a business is a wage, a loan, a repayment or a distribution is a judgement call with real tax consequences. AI sees a bank transfer; it cannot know the intent behind it or the structure it sits inside.

Constraint: These transactions need a person every time; the correct treatment depends on documents, intent and structure the ledger does not hold.

AI can guess from the description and amount whether a purchase is an operating cost or a capital asset, but the correct answer depends on what the item is for and how the business uses it — context that lives outside the ledger.

Constraint: Treatment changes tax outcomes; the deciding context is not in the transaction data.

Anyone who provides BAS services for a fee or other reward must be registered with the Tax Practitioners Board under the Tax Agent Services Act 2009. No AI tool holds a registration; a person or firm does, and answers to the ATO for the work.

Constraint: TPB registration is legally required for paid BAS services; TPB(GS) 55/2026 holds practitioners ultimately responsible for AI-assisted work.

For routine, well-understood transactions — the bulk of most businesses' activity — AI suggests the GST treatment as each transaction is coded, consistently and without fatigue, keeping treatments uniform across the period.

Constraint: Suggestions require review before reliance; TPB(GS) 55/2026 requires professional judgement to be applied to AI output.

AI keeps bank feeds matched to invoices and bills continuously, so pre-BAS reconciliation becomes confirmation of an already-current ledger rather than a quarterly archaeology exercise.

Constraint: Only as good as the bookkeeping discipline feeding it; exceptions still route to a person.

Before figures go near a statement, AI flags duplicated invoices, transactions coded inconsistently with their history, unusually large claims and GST codes that break the established pattern — a systematic version of the experienced pre-lodgement scan.

Constraint: Flags need a person to interpret; absence of flags is not evidence of correctness.

Pulling a period's coded, reconciled data into draft activity statement figures is deterministic, rule-following work AI and software automation do well. The draft arrives ready for review rather than being built by hand.

Constraint: The draft is an input to review, never the lodged statement.

Good AI-assisted workflows record what was classified, when, on what basis and who reviewed it — documentation that lines up with what the ATO expects of records and what the TPB expects of practitioners using AI.

Constraint: The ATO requires most business records to be kept for five years; records must be complete, retrievable and explainable regardless of the technology.

Mixed supplies, GST-free food rules, exports, insurance recoveries, second-hand goods, property transactions — GST's hard cases turn on facts and definitions, not patterns, and AI processes the exceptional transaction with the same confidence as the routine one.

Constraint: A wrong GST code in an AI-assembled draft looks identical to a right one; only review separates them.

Deciding whether a change belongs in this BAS as an adjustment or requires correcting an earlier statement is judgement about events, timing and thresholds — context that does not live in the transaction data.

Constraint: Where done for a fee this is registered-agent territory under TASA; either way the call is human.

Movements between entities, or between an owner and their company, carry GST and tax consequences that depend on structure and intent. The ledger shows a transfer; the correct BAS treatment depends on what the transfer actually is.

Constraint: A person has to establish the character of the transaction before any treatment is defensible.

Providing BAS services for a fee or other reward requires registration with the Tax Practitioners Board. Whoever reviews and lodges the BAS stands behind it; AI never absorbs that role, it only changes how much preparation sits underneath it.

Constraint: TPB registration required under TASA 2009; TPB(GS) 55/2026 requires practitioners to assess AI output with professional judgement before relying on it.

First drafts of monthly packs — statements, commentary skeleton, key movements — assembled from live data rather than copy-pasted from spreadsheets. The reviewing human edits and challenges instead of compiling.

Constraint: The pack cannot defend itself across a table; the presenting human's credibility is the asset.

Machines are better than people at noticing that a number moved: margin creep, cost drift, a customer segment quietly shrinking — patterns spread across hundreds of transactions that no one reviewing totals would see.

Constraint: Detection is machine work; deciding whether the movement matters and what to do is not.

AI compares actuals against budget and prior periods, isolates which lines drove the difference and explains it in plain English — the monthly 'why are we off?' conversation drafted before the meeting.

Constraint: AI can produce a fluent, well-formatted analysis containing a fabricated ratio or a misread account, and nothing in its tone will tell you; an experienced reviewer is the defence.

Price rise, new hire, equipment purchase, losing a key customer — AI drafts multiple financial scenarios in minutes, each with its cash and margin implications laid out, so more scenarios actually get tested before a decision.

Constraint: The assumptions going in and the choice coming out are human; the model only reruns the arithmetic.

Instead of discovering the cash position at month end, AI watches it daily against commitments — payroll, super, tax, suppliers — and raises early flags when the runway shortens.

Constraint: Monitoring keeps the picture current; it does not discharge anyone's duty to act on it.

Questions that once meant an analyst and a day — software spend last year, which clients are less profitable than they look — now take a query against the ledger.

Constraint: Someone must sanity-check the answer before acting on it; fluency is not accuracy.

Pricing decisions, when to raise capital, whether to take the contract with the difficult client — real CFO calls weigh incomplete information against competing goods. AI can lay out the options; it cannot weigh what matters most to this business and this owner.

Constraint: Strategic trade-offs are values questions wearing financial clothes; a model completes the pattern it was shown.

Lenders lend to people. Negotiating facilities, maintaining credibility through a rough quarter, making the case for headroom before you need it — relationship work no tool performs.

Constraint: Lenders and investors back people they can question across a table.

Software can keep the financial picture current, but it cannot carry the duty. The human at the top of the finance function is how directors stay properly informed — and defensibly so.

Constraint: ASIC expects directors to be constantly aware of the company's financial position and to prevent an insolvent company trading (s588G Corporations Act); that duty lands on humans.

Every request gets an immediate, useful reply at any hour, including a clear answer on what happens next. Nothing sits unread over a weekend, and urgent items are flagged to a person straight away.

Constraint: First response is not resolution; escalation paths to humans must actually exist and be staffed.

AI reads the request, classifies it, attaches the relevant history and sends it to the right queue. Escalations arrive with context instead of a one-line subject, which shortens the human part of the fix.

Constraint: Misrouted edge cases still occur; a person owns the queue.

One of the most common ticket categories, handled through verified self-service flows with identity checks built in — the AI guides the process; the security controls decide who gets back in.

Constraint: The identity verification is done by the security controls, not by the AI's judgement.

Questions about software you already run — sharing a calendar, setting an out-of-office, connecting a printer — answered from your environment's documentation rather than generic web advice.

Constraint: Quality is bounded by the documentation; undocumented environments get generic answers.

Where an issue has a documented, low-risk resolution — clearing a stuck update, re-mapping a drive, reinstalling an application — AI can execute the runbook and record exactly what it did.

Constraint: Documented and low-risk only; anything novel or destructive is out of scope by design.

AI notices when the same fault keeps recurring across different people — a failing update, a flaky switch, a licence about to lapse — and turns repeat symptoms into one root-cause job for an engineer.

Constraint: The root-cause work itself is engineering, not pattern matching.

AI resolves what has been seen and documented before. The outage nobody has seen — the strange interaction between an update and your line-of-business software — needs a person who can reason from first principles.

Constraint: Pattern matching fails precisely when the incident has no pattern.

Deciding whether an alert is noise or the start of an incident, whether to isolate a machine, whether anyone needs to be notified — judgement calls with real consequences. AI can surface the signals; a human must make the call.

Constraint: ACSC guidance for small business assumes human oversight of AI use; incident decisions carry consequences no tool answers for.

Changing bank details, granting new access, elevating permissions, processing an urgent payment request from the director — these are precisely the requests attackers forge.

Constraint: Always require human verification, no matter how confident the AI sounds.

Cabling, hardware failures, Wi-Fi dead spots, a new office fit-out — AI cannot hold a screwdriver. A service that leans entirely on AI has quietly excluded a large part of what IT support actually is.

Constraint: Physical work is simply outside the medium; this is a hard boundary, not a maturity gap.

After a pay run is calculated, AI can compare outcomes against currently published minimums and flag anything sitting below or oddly against them. Checking a specific number against a specific source is the shape of task AI does reliably.

Constraint: It is generating rates from memory that fails, not checking against a supplied source; the Fair Work Ombudsman's Pay and Conditions Tool remains the authoritative reference.

Award minimums move after every Annual Wage Review — from 1 July 2026, award minimum rates rose 4.75%. AI monitoring is well suited to noticing that a rate in payroll settings no longer matches a published one, before an underpayment compounds.

Constraint: Monitoring detects the mismatch; applying the correct new rate and back-checking affected periods is human work.

AI is genuinely good at turning a clause into a readable explanation — provided the actual clause is supplied and cited, so a person can verify against the source rather than trusting a paraphrase drawn from training data.

Constraint: Without the source clause attached, a summary is an unverifiable paraphrase and should not be relied on.

Patterns a person checks occasionally, AI checks every pay cycle: unusual overtime, allowances that quietly stopped appearing, penalty-rate hours that do not line up with rostered times.

Constraint: It raises the flag; a person decides what the flag means.

When something genuinely needs an adviser — or the Fair Work Ombudsman — AI can assemble the picture first: the classification in use, the hours, the clauses in play, the discrepancy that triggered the review. Better questions in, faster and cheaper answers back.

Constraint: Assembly of the facts, not resolution of the question.

Ask a general chatbot a specific award question and it will answer confidently and often wrongly. Superseded rates, invented clause references and mixed-up award versions all present identically to the right answer.

Constraint: Models are trained on the past and award rates change every July; fluency is not accuracy.

Award questions are rarely one clause. Casual loading, penalty rates, overtime, allowances and enterprise-agreement overlays interact, and the order of operations changes the answer. General AI does not reliably execute that interaction; it approximates it, plausibly.

Constraint: A plausible approximation of a pay calculation is an underpayment risk, not an answer.

Whether someone is a Level 2 or a Level 3 depends on the duties actually performed, not the title in the contract. That is a judgement call with legal consequences, made by a person who understands the work.

Constraint: An AI has only the words it was given; the duties actually performed live outside the text.

If AI gets payroll wrong, the employer wears the underpayment, the back-pay and the compliance consequences. No AI vendor stands behind an award interpretation the way an accountable payroll provider must.

Constraint: Fair Work record-keeping (7 years) and payslip (1 working day) obligations sit with the employer, not the tool; the Fair Work Commission has stated it will not use generative AI to make decisions under the Fair Work Act.