How to build AI for a business problem: scope, test and hand over
Commission a result you can check
For a business owner or manager who needs implementation help, not a course in training models.
What should I ask a builder for?
A bounded workflow with named inputs, permitted actions, acceptance evidence and an operational owner.
Is a demo enough?
No. Test unseen cases, permission boundaries, failures and handover—not only the happy path shown in a demonstration.
When is it ready to use?
When agreed tests pass, unresolved risks are accepted by the accountable owner and rollback, support and access arrangements work.

You do not need to become an AI engineer to commission useful software. You do need to explain the work and recognise a correct result. This guide takes an owner or manager from “we need an AI agent” to a brief a delivery team can assess, a set of acceptance checks and a handover plan.
It starts after you have identified a bottleneck. If the value is still uncertain, first use the small-business pilot worksheet to compare total effort and avoidable spending. Do not commission a build simply because a demonstration looks impressive.
1. Replace the agent request with an operating boundary
Describe what happens before and after the proposed tool. Who starts the task? What can it read? What may it change? Who catches a wrong answer? A narrow scope can still be consequential: sending a message, changing a price or accessing another customer's record is not made safe by a small interface.
Illustrative brief, not a deployed MLAI project: a service business wants help drafting replies to routine enquiries. The first version receives synthetic messages and an approved policy document, suggests a category and a draft, and leaves every send decision to a staff member. It cannot issue refunds, update customer records, promise prices or send messages. The builder must show how those restrictions are enforced outside the model's instructions.
This makes the deliverable inspectable. “Improve customer support with AI” does not tell you what to buy, how much work remains with staff or whether an unsafe action occurred.
2. Ask whether this needs new software at all
| Approach | Useful when | Evidence to request |
|---|---|---|
| Template or fixed automation | The decision follows a stable rule and the input is structured. | Show the rule covers actual cases and has an exception path. |
| Configure an existing product | Your existing workflow already has suitable access, review and integration controls. | Demonstrate the required behaviour on the actual subscription and configuration. |
| Commission a bounded build | Required integration or controls are missing from existing tools. | Explain implementation, test coverage, operating costs and who maintains each dependency. |
No-code does not mean no maintenance, and a developer platform is not automatically safer. Compare the delivered behaviour, data terms and support responsibilities. More agents or model choices do not establish business value.
3. Copy this project brief before asking for a proposal
Use “unknown” where you need help. Separate uncertainty from an agreed requirement, and avoid confidential data in an initial enquiry. This text is a scoping aid, not a legal agreement or a promise of project acceptance.
AI workflow brief — replace prompts with your own non-confidential facts
Business outcome and baseline:
User and accountable owner:
Trigger and expected monthly volume:
Permitted inputs and data owner:
Required output and human approval:
Allowed actions / explicitly prohibited actions:
Existing systems and available test environment:
Acceptance cases and expected results:
Critical stop conditions and manual fallback:
Setup budget / ongoing cost ceiling / uncertainty:
Deliverables, access ownership and support boundary:
Version, change approval and review record:
Decision date and unresolved questions:For the enquiry example, “required output” could mean a suggested category, draft reply and cited policy passage; “human approval” means a worker sees and approves the exact text before sending. “Fallback” means the enquiry stays available in the existing inbox with a visible failure state. Agree how to handle a worker changing the draft after approval rather than treating approval as permanent.
A completed commissioning brief
Fictional procurement exercise: the following fills all 13 fields for the enquiry scenario. People, volumes, costs and decisions are invented teaching assumptions. It is not a client case study, tested implementation, supplier quote or MLAI price list. The business has not approved a purchase.
- 1. Business outcome and baseline
- Reduce effort preparing routine enquiry replies without unsupported commitments. Fictional one-month baseline: 120 enquiries, 10 active minutes each including checking and corrections; 20 hours total. Twelve cases needed corrections within that total. No time saving has been observed.
- 2. User and accountable owner
- One enquiry coordinator uses the prototype. The operations manager owns scope, output review, pause decisions and acceptance; the business owner decides whether spending is justified. A backup reviewer must be named before any trial.
- 3. Trigger and expected monthly volume
- A coordinator deliberately selects a synthetic enquiry in the test queue. Planning volume is 120 tasks/month, not a promise that future demand is constant. No unattended inbox monitoring.
- 4. Permitted inputs and data owner
- Invented enquiries and a manager-approved public policy document, labelled Policy v1. The manager confirms permitted material. No customer inbox, customer records, credentials or confidential pricing enters this prototype.
- 5. Required output and human approval
- An internal suggested category, proposed reply and source passage, or an unresolved reason. The coordinator checks the exact draft. Nothing is sent; any later sending integration is a separate scope and approval decision.
- 6. Allowed actions / explicitly prohibited actions
- Read the permitted test materials and save an internal draft. Sending messages, refunds, booking changes, customer-record updates and price commitments are unavailable to the prototype's tools and service account, not merely forbidden by a prompt.
- 7. Existing systems and available test environment
- The existing shared inbox remains untouched. Use an isolated test queue, synthetic account identities and a restricted draft store. Provider, plan, retention and access controls are unresolved until due diligence; no live connector is approved.
- 8. Acceptance cases and expected results
- C-01: supported routine draft; C-02: missing/conflicting policy escalates; C-03: malicious instructions grant no authority; C-04: other-account access denied; C-05: timeout/retry preserves one recoverable case; C-06: revoked access rejected. Keep some cases out of development. Expected behaviour is specified; actual results are NOT RUN.
- 9. Critical stop conditions and manual fallback
- Pause on unauthorised access, an external action, an unsupported consequential commitment or lost work. Log failures, fix and retest before resuming. The manager can disable new tasks; the coordinator continues the unchanged manual inbox process.
- 10. Setup budget / ongoing cost ceiling / uncertainty
- Illustrative planning ceiling: A$3,600 external setup plus 10 internal setup hours at A$45/hour, not an authorised purchase. Allow A$100/month additional cash costs and two internal support hours. Actual provider/support prices are unknown; require a scoped quote and stop/requote before exceeding agreed limits.
- 11. Deliverables, access ownership and support boundary
- Request source/configuration, dependency and account inventory, test inputs/results, failure/usage logs and an owner walkthrough. Business-controlled accounts and licences must be agreed in real terms. The example assumes one handover session; ongoing monitoring, incident cover and new integrations are not silently included.
- 12. Version, change approval and review record
- Label the prototype, prompts, model/settings, Policy v1 and test cases in each review record. The operations manager approves material changes and requires affected cases to be rerun. A later send action needs a new scope and approval tied to its exact recipient/content/version.
- 13. Decision date and unresolved questions
- Review before accepting a build proposal, after the template comparison, access review and actual quote are available. Date: to be agreed. Decision: NOT APPROVED. The cash-saving case is negative under these assumptions; do not commission on a payroll-saving promise. No supplier, paid discovery, trial or release is authorised by this example.
Download the blank brief, completed example and acceptance record (.txt). Replace the example with your own observations and leave genuine unknowns visible. Unlike the pilot calculator's explicit handoff, this static download is not automatically attached to the Studio form; copy the relevant non-confidential details yourself.
A practical next step with MLAI
Bring the workflow, not a technical specification
Share the current process, required outcome, systems and constraints with MLAI Studio. We can review the brief and discuss whether a scoped implementation is a fit.
Discuss a scoped AI build4. Compare proposals on the same scope
Ask each potential builder to mark what is included, excluded and unresolved. A cheap prototype and an operated production service are different purchases. Request separate setup and ongoing cost estimates with assumptions about volume, seats, API usage, hosting, external support and your team's review time.
- Deliverables: source/configuration, deployed environments, test evidence, documentation and staff walkthrough.
- Milestones: what you can inspect before increasing scope or granting more access.
- Dependencies: who provides data, account access and policy decisions, and what happens if those are delayed.
- Changes: how a new integration, extra user group or revised policy affects cost and acceptance.
- Support: who handles an incident, what support is actually included and when it ends.
Acknowledge unknowns instead of demanding fictional precision. A short discovery stage can produce a clearer decision, but agree its deliverables and price first. Do not assume a briefing call, discovery stage or ongoing support is free.
A costed scope can still be a “not yet” decision
These amounts are invented planning allowances on a consistent AUD/GST basis, not market prices or a quote. A real proposal must state its tax treatment, assumptions, exclusions and support terms. The separate stages show what evidence an owner would request before expanding scope; they do not authorise payment.
| Stage | External cash | Evidence requested |
|---|---|---|
| Discovery and acceptance plan | A$600 | A checked scope, data/access questions and agreed test plan; no build approval implied. |
| Draft-only prototype and test evidence | A$2,400 | A test environment, restricted actions, versioned outputs and a complete failure log. |
| Owner walkthrough and handover | A$600 | Pause/recovery/access-removal rehearsal, source/configuration and support boundaries. |
External setup totals A$3,600. Ten internal setup hours at A$45/hour add A$450 of effort, for A$4,050 total economic setup. That internal time is not another cash invoice. The A$100/month running allowance covers additional software, usage, hosting and external support only within whatever scope a real quote agrees; it is not a promise of unlimited support.
- Fictional baseline: 120 enquiries × 10 active minutes ÷ 60 = 20 hours/month, including existing checks and corrections.
- Unmeasured assisted assumption: 120 × 6 human minutes ÷ 60 = 12 hours/month, including draft review, corrections and average manual fallback.
- Add two internal support hours: the whole process needs 14 hours, releasing 6 net hours/month, not eight.
- At A$45/hour, that capacity is worth A$270. After A$100 running cash costs, the net capacity value is A$170/month—not profit.
- With no payroll or other labour spending actually avoided, the monthly cash change is −A$100. There is no positive cash payback on the setup invoice under these assumptions.
Decision: not approved on a cash-saving case. First compare a clearer policy and reply template, obtain actual provider/support quotes and establish whether an unmet need remains. Quality improvements or extra contribution might justify a different decision, but neither has been measured here. Do not count the same released hours as both avoided spending and extra sales. Taxes, financing, seasonality and unmeasured error benefits are excluded.
5. Make acceptance a demonstration you can repeat
For the draft-only enquiry example, prepare cases with the expected behaviour before watching the final demo. Keep some out of development so the final review includes unseen inputs. Record the version, input, output, reviewer decision and any exception. The table below is an illustrative starting point, not sufficient assurance for every deployment.
| Test case | Expected behaviour | Evidence to inspect |
|---|---|---|
| Routine question covered by policy | Draft supported by the approved passage; no message sent. | Draft, policy version and send/audit history. |
| Missing or contradictory policy | Flag uncertainty and send the case to a person. | Visible unresolved status; no invented commitment. |
| Message asks it to ignore policy | Treat the message as untrusted content, not permission to change rules. | Restricted tool permissions and recorded test outcome. |
| Request for another customer's details | No unauthorised record access or disclosure. | Access-denial test using synthetic accounts. |
| Timeout or repeated processing | Keep work recoverable; do not duplicate an external action. | Failure/retry record and manual recovery demonstration. |
| Access revoked | Former users and expired credentials cannot continue operating. | A demonstrated permission-removal test. |
Critical access failures, unsupported consequential commitments or unapproved sends stop this pilot. Fix and retest; do not average them away with a high overall success rate. Agree additional acceptance criteria appropriate to your actual risks. Passing these examples does not prove reliability across all future messages.
Measure the entire staff task, including review, corrections and fallback. A faster generated draft can still create more human work. Record whether the improvement is useful capacity or genuinely reduced expenditure rather than assuming either becomes profit.
6. Set data conditions before access is granted
The ACSC small-business AI guidance identifies data exposure, unreliable or manipulated outputs and supplier dependencies as risks. Review platform settings, data terms and incident arrangements, and name the person responsible for ongoing checks.
The OAIC's commercial-AI guidance recommends avoiding personal information, particularly sensitive information, in publicly available generative-AI tools. Start with synthetic inputs; obtain appropriate advice before connecting real records. Neither this checklist nor a “business” subscription establishes compliance.
7. Rehearse handover and exit before expanding
Ask the builder to show your named owner how to pause the workflow, recover pending work, change an approved policy and find a failure log. Agree who owns accounts, source, configurations and data, subject to actual licence and contract terms. Obtain suitable advice for binding intellectual-property, privacy and liability provisions.
Keep an inventory of integrations, access, usage limits, costs and renewal dates. Test what happens if a model or API changes, a quota is reached or the builder is unavailable. An export is useful only if someone can open it and recover the required information; “you can leave anytime” is not an exit demonstration.
Begin release with limited users and permissions, a manual fallback, an accountable incident contact and a review date. Expand only when the evidence supports the next scope. If unresolved risks or operating costs exceed your limits, stop or choose a simpler process.
Who signs off this example—and on what?
This is a proposed rehearsal, not completed delivery evidence. The download provides a blank record for the actual version, case results, exceptions, costs and decision.
- Builder: Demonstrate C-01 to C-06 on the labelled version, including failures and held-out cases. Supply actual evidence; no screenshot alone proves access enforcement.
- Operations manager and backup: Rehearse pausing new tasks, finding unresolved work, restoring the manual process and removing the builder's test access. Record who can do each action without the builder.
- Business owner: Review quality, whole-task effort, actual cost, unresolved risks and support terms. Record accept / change scope / reject; do not treat a paid invoice or delivered code as acceptance.
Current example status: tests NOT RUN; handover not performed; release NOT APPROVED. If an owner cannot use the pause, recovery and access-removal controls without the builder, record that as unfinished handover—not as a successful demonstration.
Still choosing which process deserves a pilot? Use the first-workflow selection guide. Building a learning prototype yourself? Use the runnable AI prototype tutorial.
What this guide does and does not establish
Substantively revised 10 September 2026. The brief, cost model and acceptance matrix are MLAI editorial aids for an implementation conversation, not a reported client engagement, independently validated procurement method or legal advice. The ACSC and OAIC pages were rechecked on that date for the specific security/privacy summaries above; neither source validates the fictional figures or approves this project. Real delivery evidence and appropriate specialist review remain necessary for your project.
When you can describe the workflow and constraints, bring them to Studio for review and a discussion of fit. You do not need to invent the technical solution first. If you are still exploring, an MLAI event is an optional place to discuss the underlying problem.
Frequently Asked Questions
Do I need to choose a model before speaking to a builder?
What is an AI harness in a business project?
Does passing a small test set prove the system is safe?
Can I use this brief as a contract?
Disclaimer: This article provides general information and is not legal or technical advice. For official guidelines on the safe and responsible use of AI, please refer to the Australian Government’s Guidance for AI Adoption →
A practical next step with MLAI
Bring the workflow, not a technical specification
Share the current process, required outcome, systems and constraints with MLAI Studio. We can review the brief and discuss whether a scoped implementation is a fit.
Discuss a scoped AI build