Quarter-to-date review covering information released from 1 July to 28 September 2026. This edition does not cover the final two days of Q3.
A useful AI workflow creates an attractive invitation to expand: more staff, higher volumes, another business unit or another country. Each change can alter the result that justified the original deployment. Review queues grow. New documents appear. A system that prepared an answer may now be asked to execute an action.
The evidence available through 28 September supported continued investment in AI, alongside closer examination of those expansion decisions. Supplier sales and paid software adoption remained substantial. A large-enterprise survey reported more attention to costs. Singapore's financial regulator discussed how agent actions could be authorised and recorded, while Malaysia advanced its national AI institutions and governance work.
For a regional CFO or COO, the useful question was what the business could demonstrate before authorising the next stage.
Executive summary
NVIDIA's August results and Microsoft's July paid-seat disclosure showed continuing commercial demand. They described supplier performance and purchasing scale, with reporting periods that preceded much of Q3. They offered no direct measure of the value obtained by a particular customer.
KPMG's September survey of large U.S. businesses found more organisations including cost reviews in AI approvals. It also recorded attention to limits on autonomous decisions. These findings provide a comparison for management practice, rather than a measure of ASEAN adoption or productivity.
The next funding decision should use evidence from the existing workflow, plus an assessment of what expansion changes. Prepare a short memo covering realised value, performance on the new work, operating capacity, permitted actions and local ownership. Choose an explicit course: expand within stated limits, hold while a named issue is resolved, or stop further spending.
Commercial demand continued to grow
On 26 August, NVIDIA reported US$89 billion of Data Center revenue, up 117% year on year. The result covered its fiscal quarter ending 26 July. Data Center includes accelerated computing and AI, and the release labels the condensed statements unaudited. The disclosure was evidence of substantial realised supplier demand, rather than a total for the AI sector during calendar Q3.
Microsoft's 29 July earnings release said Microsoft 365 Copilot had exceeded 30 million paid seats. The statement accompanied June-ended reporting. A paid seat records purchased access. Daily use, reliable completion of work and financial benefit are separate measures that the buyer needs to examine.
| Evidence available during Q3 | What was measured | Useful management response |
|---|---|---|
| NVIDIA Data Center revenue of US$89bn, up 117% year on year | Company-reported revenue for the fiscal quarter ending 26 July | Check capacity commitments and service economics against the organisation's demand. |
| More than 30 million Microsoft 365 Copilot paid seats | Management's adoption disclosure accompanying June-ended results | Examine use and outcomes from licences already purchased before expanding access. |
| 74% of organisations in KPMG's survey included cost reviews in AI approvals | Self-reported practice among 314 U.S. leaders at businesses with at least US$1bn annual revenue | Give the next approval a documented cost review with explicit assumptions. |
The third signal came from KPMG's AI Pulse release on 24 September. The cost-review share had risen from 61% in the previous quarterly survey. Fieldwork ran from 24 July to 25 August, and more than a third of respondents represented businesses with annual revenue of at least US$10 billion.
The survey population is important. These are reports from large U.S. organisations, published by a services provider. They show how that sample described its practices, without establishing that cost review caused better returns or that Singapore businesses behaved similarly. The relevant question for a regional leadership team is whether its own approval process can explain what has been spent and what the proposed expansion will require.
Expansion changes the operating calculation
KPMG also reported that 62% of respondents were building, deploying or developing AI agents, compared with 53% in the preceding survey. The combined category includes work in development. Reading it as a production-deployment rate would overstate what the survey establishes.
The distinction has a close equivalent inside a company. A dashboard can show many initiatives while obscuring how many have an operating owner, routine users and a measured outcome. Before discussing expansion, identify which stage each proposed investment has reached and attach the relevant evidence.
For a deployed workflow, examine the full path through the process. Faster preparation may be useful, but a growing approval queue can delay completion. A result accepted in one team's documents may fail on another team's formats. Aggregate success rates can hide a category of cases that should remain outside automated processing.
An illustrative volume calculation makes the staffing issue clear. Suppose a workflow handles 2,000 cases a month and sends 10% to a reviewer, who takes three minutes per case. That requires ten review hours. At 10,000 cases and the same exception rate, it requires 50 hours. The percentage has not worsened, but the operating commitment has grown by 40 hours a month. Different case types or longer investigations would change it again.
The expansion proposal should therefore identify the source of additional review capacity and the response when that capacity is exhausted. A system can defer cases, narrow its scope or return work to the existing process. The chosen arrangement affects service commitments and needs a named owner.
Authority should expand deliberately
The same KPMG survey reported that 49% of respondents had defined high-risk use cases in which autonomous agent decisions were not allowed. This describes a stated boundary, not evidence that the boundary was effectively enforced.
The Monetary Authority of Singapore (MAS) addressed related questions in its 5 August parliamentary reply on agentic AI in financial services. It discussed a potential approach to authorising agent actions, activating human oversight and recording consequential decisions. The reply said proposed AI risk-management guidelines would be finalised soon. That dated statement should not be read as a final mandatory standard.
The operational lesson is useful even before a proposal becomes a requirement. Define the action the system may take and the conditions attached to it. Reading an account record, recommending a change and committing that change are separate permissions. An expansion into a new task should not acquire the last permission simply because the first two worked well.
For each consequential action, the operating team should be able to reconstruct the input, the authorisation and the outcome. It also needs a practical way to intervene. A policy document saying “human oversight” is incomplete if the person receives the alert after the action or cannot suspend further processing. Check the sequence using a realistic failure case before widening the authority.
Regional policy created opportunities and local work
On 31 August, MAS announced a S$220 million, three-year commitment under FSTI 4.0. Its six tracks covered several areas of fintech innovation. AI Pathfinder included support for eligible market-ready AI fintech solutions listed on PathFin.ai.
This was a multi-year programme commitment across more than AI. It should not be treated as Q3 expenditure or automatic funding for a proposed consulting engagement. An institution considering support needs to check the relevant track, eligible solution and project conditions alongside its own business case. Funding changes the financing of a project; the workflow still needs to work.
Malaysia's 28 July announcement of AI Malaysia introduced AI Malaysia Berhad, the National AI Action Plan 2026–2030 and the Malaysia AI Safety Institute. A 10 July ministry announcement described an AI Governance Bill being drafted with a risk-based approach. A draft bill and an institution launch are developments to monitor, not proof of enacted obligations or enterprise productivity gains.
For a Singapore headquarters extending a workflow into Malaysia, or any other ASEAN market, local ownership belongs in the expansion plan. Record who can confirm the applicable requirements, approve access to the local data and accept responsibility for the process. Reusable technology and controls can support several markets, while the legal entities, records and approval arrangements remain specific to each deployment.
A one-page decision memo for the next funding meeting
Complete the following memo for one proposed expansion. State the current scope and the proposed change at the top. Attach the evidence, then identify the assumption most likely to invalidate the recommendation.
| Memo field | Required entry | Decision implication |
|---|---|---|
| Value delivered | Measured benefit to date, period covered and how it was calculated. Separate released capacity from cash savings. | If benefit remains unmeasured, fund a bounded measurement exercise before wider deployment. |
| Performance on new work | Results for the additional case types, teams or languages, including material errors. | Exclude failing categories or resolve the failure before expanding scope. |
| Financial commitment | Additional setup and recurring cost, usage assumptions, spending limit and budget owner. | Set an affordable commitment and the trigger for reconsidering it. |
| Operating capacity | Review workload, support arrangements and fallback during failure or overload. | Expand only as fast as the operating team can support. |
| Permitted actions | Read, recommend and execute permissions, approval points and intervention mechanism. | Approve authority explicitly rather than inheriting it from the pilot. |
| Country ownership | Entity, accountable process owner and confirmation of local requirements. | Resolve country-specific gaps before including that market. |
| Recommendation | Expand, hold or stop, with reasons, an owner and a dated follow-up. | Make the next commitment and the evidence required for reconsideration explicit. |
“Hold” should name a solvable issue and a decision date. “Stop” should address existing users, retained records and any contractual commitments. “Expand” should state its limits, including volumes or categories that remain outside scope.
Take this memo to the next budget or renewal meeting for one existing deployment. Ask the business owner to recommend a course and the finance and technology leads to challenge its assumptions. Release the next commitment against the agreed evidence, with a date to review the result. That creates a useful record of why the organisation expanded, paused or stopped while the wider market continued to grow.
Use the release and operating pack to turn an approved expansion into named responsibilities, explicit permissions and a rehearsed recovery procedure. For another country or entity, add the ASEAN rollout record so the new scope receives its own acceptance decision.
Read our editorial principles for evidence, contribution and correction standards.