AI in Q2 2026: what the purchase commits you to

Retrospective edition covering information released from April to June 2026.

An AI proposal can look attractive while leaving most of the operating commitment outside the price. The licence buys access to a capability. A useful service also needs authorised data, a place in the workflow, staff who know how to use it and someone responsible when the result is wrong.

Q2 made that distinction increasingly relevant to enterprise buyers. Supplier revenue continued to expand, research put numbers around the preceding investment boom, and new agent platforms advertised the machinery required to manage access and control execution. Singapore's refreshed AI strategy put organisation-wide change on the agenda. The buyer's decision was how to connect those capabilities to a process the business could operate and afford.

Executive summary

Commercial disclosures released during Q2 showed substantial demand for AI infrastructure and services. NVIDIA reported further Data Center growth, while Microsoft described a US$37 billion annual AI revenue run-rate. These measures established supplier scale; they did not establish customer profitability.

The adoption picture was more varied than a single headline suggested. Stanford's AI Index reported widespread organisational use in its underlying surveys, alongside much less mature agent deployment. Using an assistant and authorising software to execute a workflow remained different management decisions.

Before adding a platform, buyers should compare configuration of existing software, purchase of a specialist product and custom delivery against the same acceptance criteria. Include integration, review, maintenance and an exit path. Assign an operating owner before procurement closes, while responsibility is still possible to negotiate.

Three growth signals, three different meanings

NVIDIA's 20 May release reported US$75.2 billion in Data Center revenue, up 92% year on year. Its fiscal quarter ended on 26 April, so the figure covered parts of calendar Q1 and Q2. It was company-reported Data Center revenue, including accelerated computing and AI.

Microsoft's 29 April earnings release said its AI business had surpassed an annual revenue run-rate of US$37 billion, up 123% year on year. The release accompanied results for the quarter ending 31 March. An annualised run-rate expresses the current pace of business over a year; it is not US$37 billion earned during that quarter or a completed financial year.

Stanford's 2026 AI Index, released in April, supplied a wider view. It reported global corporate AI investment of US$581.7 billion in 2025, up 130% from 2024. The measure includes acquisitions, minority stakes, private funding and public offerings. It is not an estimate of enterprise AI operating budgets. This was research about the preceding year, published during Q2.

The figures support confidence that AI was becoming a substantial purchasing category. They cannot be added together or turned into a single Q2 market growth rate. A buyer still needed to inspect the economics and continuity of the particular supplier it was considering. Large investment flows can support new capability without making every vendor durable or every product suitable for a multi-year dependency.

Adoption needed a more useful definition

Drawing on McKinsey's 2025 survey, the AI Index economy chapter reported that 88% of respondents said their organisations used AI in at least one business function. Scaled agent use remained in single digits across nearly all functions. Those findings describe the surveyed populations and underlying definitions, rather than all businesses or an ASEAN adoption rate.

The gap matters when an executive asks how widely AI is being used internally. A team might count purchased accounts, employees occasionally drafting text and a production system that prepares customer communications under the same heading. Those uses have different costs, risks and evidence of value.

A useful inventory therefore records the unit of adoption. For individual assistance, examine active use and the work users report completing, then validate the claimed benefit on a sample. For an embedded workflow, examine completion quality and the work left for staff. Where an agent can act through other systems, add the scope of its authority and the record of its actions.

This distinction also prevents unnecessary procurement. Existing workplace software may already cover a proposed drafting task, while a workflow requiring specialist data, approvals or transaction handling may justify a separate product. The inventory should make that difference visible before another team signs a contract.

Agent platforms brought operating questions into procurement

Google's April introduction of Gemini Enterprise Agent Platform included Agent Identity, Agent Registry and Agent Gateway alongside tools for building and running agents. These were vendor-announced capabilities, with availability and suitability to be checked for each deployment.

The terminology points to practical questions. Identity concerns which software actor is making a request. A registry helps an organisation know which agents exist. A gateway can provide a controlled route to tools and services. Buyers need to establish how those mechanisms work in their environment and what evidence they retain.

Imagine an invoice workflow in which an agent reads a document, checks supplier information and prepares an exception for review. The procurement specification should say which supplier fields it can see, whether it can alter them and who can approve the next step. It should also cover duplicate invoices, unavailable systems and a document containing misleading instructions. A successful demonstration of ordinary invoices will not answer those questions.

The value of the platform's controls depends on their configuration and use. Someone must maintain permissions, investigate failures and decide when a changed model or tool connection needs retesting. Asking for those responsibilities in the proposal makes the operating commitment visible before a contract turns assumptions into dependencies.

Compare the routes on the same work

The cheapest route can be the one that uses software the business already owns. It can also fail if configuring that software creates a fragile collection of exceptions that nobody can support. A specialist product may bring useful domain features, while custom development may be warranted where the workflow creates a material advantage or requires unusual integration.

Compare these routes using the same representative cases and output requirements. Otherwise, one supplier can appear faster because its demonstration excludes difficult documents, while another appears expensive because its proposal includes the handling of exceptions and ongoing support.

For each route, ask for a clear account of delivery and operation. Record who owns the integration, what changes require specialist help and which data or configurations can be exported. A custom system needs maintenance capacity. A managed product needs a credible service arrangement. Configuration inside an existing platform needs an owner too.

Singapore's refreshed National AI Strategy priorities, published on 20 May, reinforced this wider view. They included sectoral AI Missions, domain experts with AI capability, job redesign and governed access to data. These were announced priorities and intended actions, rather than measured productivity gains.

Regional buyers should apply the same discipline country by country. A Singapore contract may serve a regional function, but the underlying data, approval authority and system integrations can differ. Specify which entity owns the process in Malaysia, Indonesia or another operating market. Confirm local requirements separately and include the extra delivery effort in the comparison. The policy direction in Singapore does not make a regional rollout uniform.

Procurement worksheet: the first year and the handover

Use the table below as the common brief for an internal configuration team, a product vendor and a delivery partner. Each should answer for the same workflow scope. Record evidence beside the answer so that a polished proposal does not receive the same weight as a demonstrated capability.

An illustrative first-year calculation shows why the common brief matters. Suppose a proposal includes S$24,000 for setup, S$18,000 for annual licences, S$12,000 for internal integration and S$9,000 for review and support. Its first-year cost is S$63,000. Against a measurable annual benefit of S$48,000, the first-year balance is negative S$15,000.

If that benefit and the recurring S$27,000 of licence, review and support costs remain unchanged, the following year's balance would be positive S$21,000 before any new work. These are invented amounts for illustration, not market prices or expected returns. They show why a sponsor should see both the first-year commitment and recurring economics, with assumptions exposed. Quality or risk benefits may strengthen the case, but they need their own evidence rather than an unexplained addition to the savings figure.

Before the procurement decision, bring the proposed operating owner into the comparison. Ask that person to accept the handover obligations, the expected workload and the fallback arrangement. If nobody can accept them, the proposal needs further work even when the software performs well. Q2's expanding product choice made it easier to buy capability; the worksheet makes clear what the organisation agrees to run.

For an upcoming purchase, use the investment and options comparison record. Put the credible routes side by side using the same task, completion standard and planning period, then record the smallest commitment that resolves the remaining uncertainty.