Size the AI Market Around Customers Who Can Use It

On 12 October, Microsoft published Singapore findings from IDC research it commissioned. The page describes 200 local respondents and says 41% report AI-related revenue growth or new revenue streams. Its footnote identifies a July global survey of 5,204 respondents. These are reported outcomes, not audited profits or a measure of demand for a particular supplier. ¹

An older official comparator asks a different question. Singapore’s Ministry of Manpower reported on 30 April that 28.5% of private-sector establishments with at least ten employees had started adopting AI, including planning and piloting. Only 3.8% integrated it into core processes. ²

These figures must not be subtracted or read as a change over time. The populations, definitions and periods differ. The vendor release does not provide enough detail to treat its respondents as a representative count of every business. A market forecast needs the population behind the percentage, not just the most encouraging number.

Ask the commercial team to define what adoption means in its own pipeline. An employee experimenting with a chatbot, a department paying for software and a company renewing an embedded service are different signals. Record them separately. A free trial may establish curiosity; a paid renewal after implementation establishes something much closer to durable demand.

The World Bank’s October Latin America and Caribbean update argues that the ability to absorb AI matters more than access to advanced models alone. It examines constraints involving workers, firms and government. This is regional economic analysis, not a supplier sales forecast. ³

For a seller, the implication is practical. A customer may need cleaned records, a redesigned approval process, staff training and someone to resolve errors before a model becomes useful. That work can support a valuable service business. It can also consume the margin of a product sold as effortless software. The difference belongs in the price and delivery plan.

Consider an illustrative offer that helps firms prepare customer quotations. One buyer has standard product data and clear pricing authority. Another has inconsistent records and relies on an experienced employee to interpret every exception. They may share a sector and company size while requiring very different delivery effort. Do not give them the same implementation allowance by default.

Separate three groups: ready to implement, viable with paid support, and not yet economical to serve. These are proposed planning categories, not labels for a customer’s worth. A business may choose to invest in the third group for strategic reasons, but that investment should be visible rather than disguised as low-cost acquisition.

New research offers reasons to keep testing delivery costs. A study revised on 9 October compares six multi-agent workflows with a single-agent reference across ten benchmarks using the same model. One workflow improves overall accuracy; five do worse. Some configurations improve a particular task while consuming more tokens and performing worse overall. It is a preprint, not a study of customer retention. ⁴

The commercial lesson is not to avoid agents. It is to price the system that solves the customer’s actual job. A more elaborate architecture may be worthwhile for difficult cases and wasteful for routine ones. Keep the cost of review, support and failed attempts alongside model charges when deciding which customers and tasks to serve.

A separate new retail preprint explores how AI-generated product shortlists could change assortment decisions. It combines model rankings with synthetic customer preferences, so it does not measure real shopping revenue. It suggests a direction to watch: AI may change how customers discover offers as well as how suppliers deliver them. ⁵

The World Bank’s East Asia and Pacific update distinguishes the region’s role in AI-related trade from the wider spread of AI use. Infrastructure activity and application demand are not interchangeable measures. A company selling into a data-centre supply chain faces a different market from one selling workflow software to ordinary firms. ⁶

For North America and Europe, test the target sector’s procurement, security and integration requirements rather than assuming a mature technology market makes implementation easy. Across Asia-Pacific, Latin America, Africa and the Middle East, check local-language performance, support capacity, connectivity and payment arrangements for the specific buyer. These are diligence questions, not a ranking of regions.

The evidence reviewed does not provide comparable conversion or renewal rates across those markets. Keep that gap explicit. Start with a narrowly defined customer group in each proposed market and measure its economics. Expand because the service works for that group, not because a result from another country has been renamed a global opportunity.

Ask the sales, finance and delivery leads to rebuild one AI revenue assumption together. Name the buyer, the job, the purchase decision and the conditions required to go live. Show paid conversion, implementation effort, continuing support and renewal evidence separately. Where evidence is missing, label the assumption and decide how to test it.

Fund the support that makes a promising group commercially viable. Stop counting unqualified interest as committed demand. Watch whether lower delivery costs open a new customer group, and whether discovery shifts toward AI intermediaries. The prize is not the largest headline market. It is a repeatable offer that customers can use and your business can afford to deliver.