Build, buy, integrate or stop: an AI investment memo for 2027

An investment committee should be able to approve an AI project without first accepting a view about the future of the entire technology sector. The proposal needs to explain one operating problem, the credible alternatives and the evidence that justifies the next commitment.

Write the memo around the business decision: which approach can improve this workflow at an acceptable cost, with an owner able to run it? Include configuring existing software, changing the process and stopping. A custom build deserves consideration when its additional benefit warrants the delivery and support obligation.

The framework below is for approving an individual workflow. It makes the assumptions visible enough for finance, operations and technology to disagree productively before money is committed.

Describe a completed unit of work

Begin with the process boundary and its current performance. “Automate finance” is too broad to cost or evaluate. “Prepare invoice exceptions for review, ending with an accepted or escalated case in the finance system” creates a measurable unit.

Record volume, staff effort, rework, elapsed time and the consequences of error. Name the source and measurement period for the baseline. Where data is incomplete, identify the uncertainty and fund a small measurement exercise before treating the estimate as a fact.

The Monetary Authority of Singapore (MAS) MindForge operationalisation initiative provides a useful reminder of the operating work surrounding AI: oversight, risk management, lifecycle management and supporting capabilities. Its March 2026 release supplies guidance and published implementation examples. It does not remove the need to establish the economics and acceptance criteria for this particular workflow.

Compare approaches on the same task

Ask each option to meet the same definition of completion and to expose the work that remains with the buyer. Otherwise a supplier's extraction price may be compared with an internal team's end-to-end processing cost.

Integration appears in several options. It is work to estimate rather than a separate promise that a purchased product will simply fit. Include identifiers, interfaces, permissions, testing environments, migration and the handover to operations.

Put the credible options next to one another using the same task, quality requirement and planning period. An option that fails a necessary requirement can be excluded with evidence. The remaining choices need comparable cost, retained effort, delivery timing and exit commitments. The options comparison record keeps missing evidence visible so the preferred solution does not receive a detailed case while alternatives are dismissed with adjectives.

Separate economic capacity from cash

Consider this illustrative case. A team processes 4,000 cases per month at eight minutes each, including current review and rework. At an allocated staff cost of S$45 an hour, the current effort is approximately 533 hours, valued at S$24,000 a month.

A configured solution is estimated to reduce routine handling to five minutes per case, while adding 20 hours of exception rework each month. Licence and support cost S$4,500 monthly; implementation costs S$27,000 once.

The scenario releases 180 staff hours per month, valued at S$8,100. After the additional licence and support cost, the economic difference is S$3,600 a month. Dividing S$27,000 by that difference gives a 7.5-month recovery period on the stated assumptions. This simplified calculation assumes immediate steady operation, unchanged volume and no other net cost changes; a real proposal must cost the transition as well.

That is not automatically cash payback. If the same staff remain on the same payroll, immediate spending increases by the software cost and implementation bill. The proposal must explain whether released time absorbs growth, avoids a planned hire, improves a service outcome or allows an actual cost reduction. Record that mechanism separately from the allocated cost comparison.

The five-minute assumption also matters. At six minutes per case, with the same extra rework and software cost, the proposed monthly cost rises to S$23,400. The economic difference falls to S$600, extending the illustrative recovery period to 45 months. A one-minute change can therefore alter the approval decision.

Make the next commitment conditional on evidence

The memo should request the smallest commitment that can resolve the important uncertainty. If the uncertainty is integration access, another model demonstration will not answer it. If the uncertainty is review effort, evaluate the proposed workflow with the people who will review it.

Use a defined acceptance set, including difficult cases and cases that should be escalated. Agree which failures are unacceptable and what happens when one occurs. The business owner should understand the result in terms of completed work, residual effort and consequences, not just a model score.

A supplier's promise to fix issues after launch needs a contractual and operational meaning. Identify who investigates incidents, pays for remediation, approves changes and maintains a fallback. Retain an exit path for the records and configurations required to keep the business running.

The memo to take into the approval meeting

Copy the following fields into the proposal. A short answer with a supporting record is more useful than an optimistic paragraph.

For the illustrative case, the most important next evidence is whether five-minute handling survives real review and whether the released capacity has a funded use. Approving a capped evaluation of those assumptions is a stronger decision than accepting the 7.5-month figure as a forecast. Take the memo to operations and finance together; ask them to identify the assumption most likely to change their answer.