In short
Choose AI initiatives by business value, feasibility and risk, then fund the next useful test rather than an entire rollout. Compare AI with simpler process and system changes, and do not count the same benefit in several proposals.
In this guide
Start with a decision the business already needs to make
A finance team wants faster reporting. Customer service wants help answering enquiries. Operations wants fewer manual updates. These may all be worthwhile, but they compete for the same integration capacity, subject-matter experts and management attention. An AI strategy needs a way to choose between them.
Ask each sponsor to name the business problem, the person doing the work and the decision that would improve. "Introduce an AI assistant" does not tell you whether the proposal reduces a backlog, improves a customer response or simply changes how a report is written. The existing AI strategy guide covers the overall structure; this article focuses on allocating investment within it.
Use the same questions for every proposal
A short comparison is more useful than separate presentations with different assumptions. Treat each estimate as provisional until the relevant workflow has been observed.
| Decision input | What to establish | Reason to pause |
|---|---|---|
| Business value | Current volume, delay, effort, quality and consequence | The sponsor cannot identify a measurable problem |
| Alternative | What simplification or existing software could achieve | The proposal assumes AI before comparing options |
| Delivery feasibility | Accessible data, integration permissions and available team time | The demonstration depends on data unavailable in production |
| Risk and recovery | Who can be affected, approvals and a workable fallback | An incorrect output could act without an authorised check |
| Full cost | Setup, licences, support, review and ongoing changes | The business case excludes the people checking the output |
A high score should not override a failed safety or access condition. An initiative can be attractive but not ready. Keep it in the portfolio with a specific dependency instead of forcing it into the delivery queue.
Look for shared work and duplicated benefits
Reporting automation and a management assistant may both depend on the same account mapping. Fixing that mapping once could benefit both. However, each proposal must not independently claim all the reporting time saved. Assign benefits to the process change that produces them and record shared costs explicitly.
Also check who will review the work. Five small pilots requiring the same finance manager are not five independent projects. A realistic sequence accounts for close periods, payroll deadlines and service peaks, not only software delivery estimates.
Approve a bounded next test
The first commitment can be a read-only reconciliation, a sample of support cases or a supervised reporting cycle. Define the accepted output, the quality threshold, the maximum spend and the condition that stops the test. An approved investigation is not approval to connect production write access.
For example, a proposed board-report assistant could first assemble an approved source pack and flag missing information. Management can then compare preparation and checking effort with the current process. Only after that comparison should it consider generated commentary or additional data access. This is a proposed test design, not a reported client result.
Give leadership a portfolio it can act on
Report each initiative as investigating, testing, operating, paused or stopped. Show what changed since the last review, what remains uncertain and the decision required. A stopped experiment can be a sensible outcome if it prevents a larger, unsupported commitment.
The AICD and UTS HTI guidance connects AI investment with strategy, resourcing, governance and measurable use cases. The practical implication is to review the business case and operating responsibility together, rather than presenting adoption counts as the return.
Advery can help turn a list of ideas into a sequenced opportunity portfolio and define the first test through its technology and AI work.