Choose a route after approving the use case
You have a defined business problem, an accountable owner and enough evidence to justify a pilot. Now the market offers a crowded choice of software subscriptions, configurable platforms, integration partners and custom development. A polished demo can make that choice look simpler than it is.
The sourcing decision follows the use-case decision. If the opportunity hasn't passed its value, readiness and risk gates, start with the business AI use-case prioritisation framework. Once it has, compare routes by the fit the business needs, the control it must retain, the surrounding systems and the capacity to own the result.
Custom AI can mean a purpose-built workflow, interface, integration or application layer that uses an existing model or service. Model training is only one possibility. Buying software still leaves the business responsible for data, oversight, evaluation and change.
Five routes can serve one approved capability
Build versus buy is a useful search phrase, but it hides the middle of the decision. Most growing businesses are choosing among five practical routes.
- Buy a product when the capability is common, the standard workflow is acceptable and the product meets the evidence and control requirements.
- Configure a product when built-in rules, permissions, prompts, fields and interfaces can create enough business fit without changing its core.
- Integrate existing products, APIs and business systems when the capability is available but information and actions must move through a deeper workflow.
- Custom-build the workflow, interface, integration or application logic when distinctive behaviour or stronger ownership justifies continuing responsibility.
- Defer when the process, knowledge, permissions, evidence, budget or accountable ownership isn't ready.
A partner is a delivery owner across these routes. They can help evaluate, configure, integrate or build. The business still needs to know what it owns, what the partner owns, what the supplier controls and how another team could operate the system later.
Start with the decision requirements
Write the requirements before comparing products or proposals. Otherwise, a vendor's strongest features can quietly become your selection criteria.
First, name the required business-specific fit. A standard drafting assistant may need little adaptation. A qualification workflow that must combine service rules, website inputs, CRM state and a salesperson's judgement needs more. If the process and handoffs are still unclear, use the workflow automation readiness guide before selecting a route.
Then define the ownership and control requirement. Record which data can enter the service, where it can be stored, who can change rules, how outputs are reviewed, what logs are available and how the business can stop or switch. A licence works best with clear sources, permissions and review responsibility. Resolve those knowledge-readiness questions first.
Time matters, but “fastest to launch” is too narrow. Compare time to trustworthy evidence, including setup, integration, representative testing, staff preparation and stakeholder review. Also record the capability required after launch. Someone will need to handle changes, exceptions, supplier updates, incidents and evaluation across every route.
The Australian Government Guidance for AI Adoption describes accountability, impact and risk assessment, information management, testing, monitoring and transparency as foundations for responsible adoption. The voluntary NIST AI RMF Core similarly organises lifecycle work across govern, map, measure and manage, including risks involving third parties. Together, they show why accountability remains with the adopting organisation even when a supplier provides the technology. The route itself remains a business decision.
Choose the route from fit and ownership needs
The matrix makes the first route discussion easier because it separates two questions that product comparisons often blend together. How specifically must the capability fit this business, and how much of its logic, data path and change process must the business control?
Fit and ownership needs determine the delivery route
Ownership and control increase
Integrate
Standard capability needs deeper connection to business systems.
Custom-build
Distinctive capability needs stronger control and owned logic.
Buy or configure
Standard capability can use proven product patterns.
Defer
Required fit exceeds the business's present ownership capacity.
Business-specific fit increases
Buying or configuring is the default area for standard needs with modest ownership requirements. Integration becomes more suitable when proven capability must connect deeply with business data or systems and the organisation needs control over that connection. Custom delivery becomes credible when distinctive behaviour and stronger ownership are both commercially important.
Defer is the honest answer when the required fit is high but the business can't yet own the rules, sources, review or maintenance. It creates space to repair the foundations or narrow the requirement. The matrix gives you a starting area rather than a verdict. Due diligence, ownership cost and pilot evidence can still change the route.
Run due diligence on evidence
Ask every shortlisted supplier or delivery team to demonstrate fitness for the intended use rather than broad AI capability. Provide representative cases, expected answers, unacceptable behaviours and the human decision boundary. Record which claims came from a contract, technical document, test result or sales statement.
The OAIC's commercial AI product guidance asks organisations to consider the intended use, necessity, accuracy, human oversight, provider access, security, transparency and ongoing privacy fitness. For any personal information, clarify what enters the service, what the provider retains or uses, where it is processed, who can access it and how inaccurate information can be corrected. Privacy questions should receive suitable specialist review where the proposed use or contract creates material uncertainty.
For generative AI, the NIST Generative AI Profile identifies third-party intellectual property, privacy, information-security and supply-chain risks. Its suggested actions include using procurement requests and contracts to seek evidence such as service-level agreements, software bills of materials and assurance reports where appropriate. A proportionate growing-business review can ask for the following evidence:
- test results for the intended task and relevant failure cases
- data use, retention, location, access and deletion terms
- security controls, incident process and relevant assurance reports
- intellectual property terms for inputs, outputs and configured material
- model, subcontractor and material-service dependencies
- uptime, support, change-notice and service-level commitments
- logging, human review, correction and monitoring capabilities
- export formats, termination support and deletion evidence
Connect the contract review to the business's AI governance controls. Legal, security and contractual specialists may need to review high-consequence uses, unusual data flows, unclear liability or terms the team cannot evaluate confidently.
Compare three-year ownership cost
Price comparisons need the same boundary. A subscription can appear cheap when integration and review sit elsewhere in the budget. A custom proposal can appear expensive when it includes work that a product still requires internally.
For each viable route, estimate the same three-year period. The categories below are Off Piste's synthesis of lifecycle work reflected in the DTA guidance for moving from proof of concept to scale, the Australian adoption foundations and the NIST lifecycle framework. Use them as a reader-input worksheet rather than as source-stated costs or market benchmarks.
Record a low, expected and high case when volume or change is uncertain. Include staff time rather than treating it as free. Then separate one-off investment from continuing cost and identify the assumptions most likely to move the result, such as API volume, review rate, integration maintenance or supplier pricing.
The worksheet won't produce a perfectly forecast number. It will expose where one route transfers cost into human review, technical dependence or exit. That's more useful than a single purchase price detached from ownership.
Prove the route before committing
Begin with a market scan against the written requirements. Request evidence from the most credible options, then use a controlled sandbox with representative cases and approved data. Review the result with the people who own the workflow, information, risk and customer outcome before funding a bounded pilot.
The Australian Government's proof-of-concept-to-scale guidance treats scaling as both technical and organisational work and includes evaluation and procurement planning. For a growing business, the practical lesson is to define acceptance evidence before a pilot becomes a quiet production dependency.
Acceptance evidence should cover task quality, reliability, exception handling, human effort, customer or staff impact, control performance and actual operating cost. Write the baseline, target range, test cases, prohibited data, review point, owner and stop condition into the pilot brief. Then decide whether to scale, fix, stop or reconsider the delivery route.
Estimated ownership cost belongs in the sourcing decision. Once the pilot is live, use actual outcomes, review effort, failures and cost in the AI workflow ROI and reliability scorecard.
A hybrid route in practice
Consider an illustrative lead-qualification workflow for a service business. A bought form or conversational component collects structured information. Configured rules check service area and minimum requirements. An integration passes the record to the CRM. A model summarises free text, and a salesperson makes the commercial decision with a clear handoff for uncertain cases.
That route mixes purchased capability, configuration, custom integration, business-owned data and human judgement. The AI lead qualification guide shows how to bound each role. If the underlying intake, forms, content or lead journey need redesign, website design becomes part of the delivery problem rather than an incidental connection.
Hybrid delivery can also apply to internal and content workflows. The business may buy model access while retaining its source library, review rules, interface and publication decision. Our guide to an AI-assisted content workflow shows why business expertise and approval remain owned even when a tool accelerates drafting.
Make ownership the final decision
Finish with a short decision record. Name the approved use case, selected route, evidence reviewed, material assumptions, accountable owner, review date, acceptance range, stop condition and exit plan. Record which parts are controlled by the business, supplier and delivery partner.
If the choice remains unclear, test the unresolved requirement through an AI systems discovery, focused vendor evaluation, integration design exercise or bounded custom pilot. The route is ready when the evidence supports the intended use and an accountable person can own its operation, change and exit.