AI in Q1 2026: choosing a useful first investment

Retrospective edition covering information released from January to March 2026.

For a finance or operations leader deciding where to invest in AI, the opening months of 2026 supplied plenty of reasons to pay attention. Infrastructure suppliers reported substantial growth. New products offered to prepare financial analysis, work with spreadsheets and assemble presentations. Singapore placed AI transformation prominently in its Budget.

The decision still needed to begin inside the business. Which recurring task was expensive enough to improve, sufficiently understood to measure and contained enough to test safely? Q1 strengthened the case for a serious trial around that question. It offered much less evidence for approving a broad programme on the strength of industry momentum alone.

The quarter in one minute

AI demand was visible in reported supplier sales and ambitious infrastructure plans. These were different signals: a fiscal-quarter revenue figure described business already done, while a full-year capital expenditure forecast described management's intentions. Neither measured the return available to an individual buyer.

Office-oriented AI products made familiar finance and operations tasks worth testing. Their usefulness depended on the accuracy of the work, the checking it required and its fit with existing systems. A polished document was an incomplete measure of success.

Singapore's policy and governance announcements connected adoption with data, process redesign, staff capability and human accountability. For a regional business, the practical response was to select one valuable workflow, give it an owner and agree what would justify further investment before starting the trial.

Growth was clear; the measures needed care

On 25 February, NVIDIA reported US$62.3 billion of Data Center revenue for its fiscal quarter ending 25 January, up 75% from the corresponding prior-year period. Earlier that month, Alphabet said it anticipated US$175–185 billion of capital expenditure during 2026. The scale of both disclosures mattered. Their definitions mattered just as much.

These figures do not combine into a credible estimate of total AI-sector growth. Their periods differ, and infrastructure purchasing and supplier revenue can describe different sides of related transactions. NVIDIA's release also labels its condensed financial statements unaudited.

For buyers, the useful conclusion was that a significant commercial ecosystem was expanding around the technology. Procurement should therefore account for continuing product change. A multi-year agreement needed a practical answer to model replacement, data export and changing service requirements, alongside today's demonstration and price. Expanding upstream capacity could create more choices, but it did not guarantee that the complete workflow would become cheaper.

The work available to test became more recognisable

Anthropic's 5 February announcement of Claude Opus 4.6 described financial analysis, research, documents, spreadsheets and presentations. These are activities a business sponsor can recognise without understanding model architecture. The announcement also distinguished availability: Claude in PowerPoint was a research preview, so the associated features should not all have been treated as equally mature.

Consider the preparation of a monthly management pack. A trial might ask AI to assemble a first draft from approved reports, identify movements and attach supporting references. The finance team would then check the figures, challenge explanations and approve the commentary. That is a specific proposition with observable work at each stage.

Its evaluation should include awkward cases. Does the draft compare the correct periods? Can it distinguish a change in currency from a change in underlying performance? When two reports disagree, does it expose the discrepancy or choose one silently? A fluent explanation of an incorrect variance can create more work than an obvious missing answer.

This is why the trial needs the people who perform and review the task. They can identify errors that a generic product benchmark will miss and record whether correction takes longer than doing the work directly. Vendor announcements help identify capabilities to investigate. Only a trial on representative work can establish whether those capabilities help this particular team.

The production gap was an operating problem to investigate

A Deloitte report published on 21 January found that only 25% of respondents had moved at least 40% of their AI pilots into production. Its survey covered 3,235 business and IT leaders across 24 countries and six industries, with fieldwork in August–September 2025. It described that sample's prior experience rather than a Singapore adoption rate during Q1.

The finding did not establish why individual projects stalled. It did justify a more demanding pilot proposal. A team could produce a successful demonstration while leaving access permissions, support, exception handling or responsibility for the output unresolved. Those are useful questions to examine before funding a trial whose eventual destination is a live business process.

The sponsor should be able to describe the work after the experiment succeeds. Who uses the result? Where does it enter the system of record? Who maintains the connection when a source report changes? What happens when the service is unavailable? These answers determine the effort needed to turn a promising prototype into something staff can rely on.

Singapore connected ambition with implementation

Singapore's Model AI Governance Framework for Agentic AI, announced on 22 January, made human accountability explicit. It described approval checkpoints, baseline testing and controlled access to services. The framework provided guidance to organisations developing or buying agents; it was not a new universally mandatory AI statute.

For financial institutions, the Monetary Authority of Singapore's 20 March Project MindForge announcement added an AI risk-management toolkit, including an operationalisation handbook. Developed with 24 banks, insurers, capital-markets firms and partners, it offered practical implementation support aligned to proposed guidelines. It gave finance leaders a resource for structuring the work around AI, without certifying a particular deployment or establishing its return.

An agent is software that can use tools to perform steps in a task. That ability makes its permissions a business design choice. An assistant preparing a payment proposal needs different authority from a system releasing a payment. A trial can keep the former task useful while reserving the latter for an authorised person.

The 12 February Budget Statement set out AI Missions in advanced manufacturing, connectivity, finance and healthcare, and announced the Champions of AI programme. Its account of transformation included organising data, rebuilding systems, redesigning jobs and retraining workers. These were policy commitments and intended programmes, with implementation and results still to follow.

For a Singapore-based regional team, this direction favoured a proposal that explained the process change as clearly as the software purchase. It also supported involving domain experts early. Someone needs to know which records matter, which exceptions demand judgement and what competent work looks like.

Regional expansion introduces another test. A process described by headquarters as “the same across ASEAN” may use different source systems, approval rights, languages or local reporting practices. Map those differences before reusing an evaluation result from one country. Singapore guidance can inform the design, while applicable requirements and programme eligibility need separate confirmation in each market.

A worksheet for choosing the first trial

Complete one copy for each candidate workflow. Use observations from current work wherever possible. An unknown is a discovery task, not a reason to award an optimistic score.

Start with the workload estimate because it can expose a weak investment case early. In an illustrative example, 800 monthly cases taking 18 minutes each represent 240 hours of handling time. Saving five minutes on each case would release about 67 hours. At an assumed S$60 an hour, that is S$4,000 of monthly capacity value before additional technology and operating costs.

That amount is not automatically a cash saving. The business must be able to redeploy the capacity, avoid additional hiring or remove an expense. If reviewers absorb the five minutes elsewhere, the claimed benefit disappears. A low-volume task can still justify investment through better quality or reduced risk, but the proposal should name and measure that benefit directly.

Use the completed worksheets to choose one trial whose potential benefit, evidence and operating ownership are credible. Give its sponsor a fixed date to present both the result and the work required to run it. That decision would have been a sound response to Q1's growth signals, while preserving the ability to change course as the products and evidence developed.

To apply that decision today, use the workflow discovery pack. It provides a record for the process boundary, baseline, alternatives and next uncertainty to test. Complete it with the person who performs the work and the sponsor who can authorise a change.