MLAI AI use-case prioritisation record — 2026-09-11 Article: https://mlai.au/articles/featured/what-is-artificial-intelligence-used-for-in-everyday-work-and-life Fictional retailer, invented inputs and planning decisions. Not a supplier quote, AI benchmark, client result, financial forecast or compliance approval. Independent owner/technical review pending. All proposed business tests are NOT RUN. No live trial or purchase is approved. THREE TASKS — SAME MONTHLY BASIS Active baseline and assisted case minutes include case review, correction and fallback. Rework cases are a subset, not extra time to add again. Extra oversight/maintenance is separate. A$60/hour is an invented internal capacity value. All cash figures need a consistent cost/tax basis in real quotes. Existing payroll is not avoided; no additional revenue is evidenced. Extract supplier-invoice fields Monthly workload and rework: 40 cases x 8 minutes = 320 baseline minutes; 6 cases need correction within that total. Planned assisted work: 40 x 4 = 160 case minutes, plus 60 additional oversight minutes. Potential capacity release: 100 minutes/month. Data availability: Invented invoice samples can be used for a demonstration. Live supplier records have not been approved for connection; the accounts lead must confirm permission and source access. Human oversight: Accounts lead checks every proposed field against the original. No posting to accounts, bank-detail changes or payments are in scope. Cost of error: A wrong total, duplicate or misread credit creates reconciliation work. Payment must remain unavailable. The cost and frequency of escaped errors are unknown, not zero. Setup cash: A$300; internal setup: 4 hours x A$60 = A$240; combined setup economic value: A$540. Incremental monthly cash: subscription A$15, usage A$5, external support A$10, extra labour cash A$0; total A$30. Monthly cash change with no avoidable payroll: −A$30. No positive cash payback under these assumptions. Capacity is not profit. Non-AI alternative: Compare manual entry, supplier CSV exports and existing accounting rules before a model. Decision: Investigate invoice extraction. Prepare a synthetic demonstration brief with the accounts lead, then check document-level quality and review effort. This is not permission for a live pilot or a purchase. Draft routine customer replies Monthly workload and rework: 60 cases x 5 minutes = 300 baseline minutes; 9 cases need correction within that total. Planned assisted work: 60 x 3 = 180 case minutes, plus 90 additional oversight minutes. Potential capacity release: 30 minutes/month. Data availability: The fictional returns policy has conflicting versions. No approved answer source is ready; reviewer time has not been committed. Human oversight: The operations lead must approve the policy and assign a reviewer. Any future system would draft only; sending and commitments remain separate. Cost of error: An unsupported refund, price or delivery promise can create a customer dispute. Consequential losses are unpriced; a fluent answer is not a control. Setup cash: A$150; internal setup: 3 hours x A$60 = A$180; combined setup economic value: A$330. Incremental monthly cash: subscription A$15, usage A$5, external support A$5, extra labour cash A$0; total A$25. Monthly cash change with no avoidable payroll: −A$25. No positive cash payback under these assumptions. Capacity is not profit. Non-AI alternative: Repair the policy and saved replies first. Decision: Fix the source first. Do not select this AI trial while policy versions conflict or reviewer capacity is unknown, even if the arithmetic appears favourable. Suggest compatible catalogue items Monthly workload and rework: 120 cases x 3 minutes = 360 baseline minutes; 12 cases need correction within that total. Planned assisted work: 120 x 2 = 240 case minutes, plus 60 additional oversight minutes. Potential capacity release: 60 minutes/month. Data availability: The fictional catalogue has approved compatibility fields and a catalogue owner. Customer profiling data is unnecessary for the initial comparison. Human oversight: Staff confirm suitability and stock before recommending an item. A ranking does not approve a sale or determine customer eligibility. Cost of error: Wrong compatibility can cause returns, replacement work and trust damage. These losses are not measured or priced into this planning example. Setup cash: A$300; internal setup: 6 hours x A$60 = A$360; combined setup economic value: A$660. Incremental monthly cash: subscription A$10, usage A$5, external support A$10, extra labour cash A$0; total A$25. Monthly cash change with no avoidable payroll: −A$25. No positive cash payback under these assumptions. Capacity is not profit. Non-AI alternative: An exact compatibility filter using the existing catalogue is the first comparison. Decision: Try the catalogue filter first. The highest task volume is not a reason to add AI. Establish what the simpler filter cannot do before considering generated recommendations. SENSITIVITY — NOT A NEW OBSERVED RESULT If invoice review/correction/fallback requires seven minutes per case: 40 x (8 - 7) - 60 = -20 net minutes/month. It uses 20 more minutes, not a saving. The monthly cash change remains −A$30. Setup, subscriptions, usage, external support, internal maintenance and per-case review are recorded separately. Error losses, revenue effects, taxes, financing, seasonality and actual supplier/support terms are not established. These exclusions prevent a complete ROI claim; obtain relevant evidence before spending. COMPLETED CAPABILITY-TO-TASK REQUEST — INVESTIGATE ONLY 1. My business task and current method: Copy supplier-invoice fields into a review sheet. Fictional baseline: 40 documents/month, 8 active minutes each including existing checks and rework; six cases need correction. No actual baseline was collected. 2. Capability needed: draft, extract, recommend, detect or other: Extraction: proposed supplier, date, document type and total alongside the source. It is not accounting advice, autonomous bookkeeping or payment approval. 3. Permitted input and source owner: Invented documents initially; accounts lead owns the source/checking rules. Live-record permission and vendor data handling remain unresolved. 4. Exact output and who checks it: Proposed fields or an unresolved reason, with source location and version. The accounts lead reviews each complete document, not just an overall field-accuracy percentage. 5. What the output must NOT trigger automatically: No accounting write, payment, supplier bank-detail change or outbound commitment. The implementer must demonstrate these boundaries; they are not implemented by this worksheet. 6. Normal, missing, misleading and duplicate test cases: Clear invoice; unreadable total; credit note; repeated document; embedded hostile instructions. Actual outcomes: NOT RUN. Retain failures and review effort. 7. Evidence needed to judge accuracy and review effort: Original input, expected and actual fields, whole-document acceptance, all attempted cases, errors and review/correction/fallback time on the identified version. Seven review minutes per document would erase the planned capacity benefit. 8. Non-AI alternative to compare: Manual entry, a supplier-provided CSV and existing accounting rules. Stop if the simpler option meets the task without an AI build. 9. Systems involved, budget constraints and unresolved questions: Unconnected test document folder and review sheet. Illustrative A$300 setup cash plus four internal hours; A$30/month extra cash and one hour/month extra oversight. Obtain actual quotes and a spend cap. No purchase, live trial, independent review or positive cash payback is established. BLANK PRIORITISATION WORKSHEET — COPY FOR EACH OF YOUR THREE TASKS Business task / capability / current workaround / process owner: Representative period / task count / active time including current checks and rework: Rework case count (subset, not extra minutes) / input availability and permission: Required output / reviewer and capacity / forbidden automatic action: Error consequence / recovery work / unresolved cost or obligation: Non-AI alternative and evidence of an unmet need: Assisted case time including review, correction and fallback: Extra maintenance/oversight time / internal hourly capacity value: External setup cash / internal setup time / training and handover: Monthly subscriptions / API use / external support / extra labour cash: Spending demonstrably avoidable / extra contribution evidenced (do not double count): Unknown costs / required quotes / tax basis / spend cap: Investigate, fix prerequisites, prefer simpler option or reject — and why: Evidence that would change the choice / owner / next review date: BLANK CAPABILITY REQUEST My business task and current method: Capability needed: draft, extract, recommend, detect or other: Permitted input and source owner: Exact output and who checks it: What the output must NOT trigger automatically: Normal, missing, misleading and duplicate test cases: Evidence needed to judge accuracy and review effort: Non-AI alternative to compare: Systems involved, budget constraints and unresolved questions: ACTUAL TEST RECORD — DO NOT REPLACE UNKNOWN WITH ZERO Case ID / permitted source and version / normal, missing, conflicting or duplicate: Expected output / actual output / whole-document or whole-task decision: Reviewer / date / failure and recovery / handling minutes / actual costs: Pass, fail, unresolved or NOT RUN / decision and next action: Optional enquiry: https://mlai.au/mlai-studio/start-project Bring a permitted summary of the selected task, current work and acceptance criteria. The file is not automatically attached; acceptance, availability or savings are not guaranteed. SOURCE ROLES — checked 11 September 2026 https://business.gov.au/online-and-digital/artificial-intelligence — possible business capabilities and responsible adoption; not support for fictional prices or savings. https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products — commercial-AI privacy due diligence, not compliance approval of this example. Use non-confidential facts only. Do not insert customer documents, credentials or private business details into a public article form.