The offer has been on the table since February. Four dollars off a company’s taxable income for every dollar it spends putting AI into its operations, capped at S$50,000 of spending a year. The same Budget created a National AI Council and put the Prime Minister, Lawrence Wong, in the chair. Five months on, the Inland Revenue Authority’s guidance on which dollars actually count is still pencilled in for the middle of this year, which is now.

That gap is a cashflow problem rather than a technology one. The deduction runs for Years of Assessment 2027 and 2028, and YA2027 assesses the financial year a December-year-end company is sitting in today. The spending has to happen in the months before the rulebook lands, or the first of the two years is gone.

The arithmetic is worth doing slowly. S$1,000 of qualifying spend produces S$4,000 of deduction; at Singapore’s 17 percent corporate rate that is S$680 back, and a net cost of S$320. Fill the S$50,000 cap and the deduction is S$200,000, worth roughly S$34,000 a year. The catch sits in the structure. Alone among the activities under the Enterprise Innovation Scheme, AI spending cannot be converted into a cash payout, so the benefit reaches only a company with enough taxable profit to absorb it. For a firm whose profit is smaller than its deduction, this is a question of timing rather than a cheque.

Which is why the sensible version of this is small and dull. Supplier follow-ups are the case in point, the task every operations manager does badly because it is boring: the chaser that should have gone out on day four and went out on day eleven. Zoho gates its Zia tooling behind the Enterprise tier of its CRM at US$40 per user per month on annual billing, and the Agent Studio that will build a follow-up-drafting agent behind Ultimate at US$52. Four seats of Ultimate runs about US$2,500 a year, a rounding error against a S$50,000 cap. The subscription is the cheap line. Scoping, integration, and training the staff who will use it are the expensive ones, and those are what the cap is sized for.

The operator who wires up one task, writes down what it was meant to do, and keeps the vendor invoices has a claim. The operator who expenses a seat licence has a hope.

They are also what makes a claim survive. The advisers reading the scheme keep arriving at the same warning: enhanced deductions are not automatic, and documentation is the main risk. A statement of work with named deliverables reads differently to an assessor than twelve months of subscription receipts with no project around them. The operator who wires up one task, writes down what it was meant to do, and keeps the vendor invoices has a claim. The operator who expenses a seat licence has a hope.

The guardrail arrived a month before the money. On 22 January, at Davos, Josephine Teo launched IMDA’s Model AI Governance Framework for Agentic AI, the first of its kind anywhere, built on the position that humans remain ultimately accountable for what an agent does and that an organisation must define the checkpoints at which human approval is required. It maps agent risk across five categories, unauthorised actions among them.1

The deduction covers the wiring. It does not cover the send.

So the incentive and the constraint point the same way, which is rare enough in policy to be worth noticing. The task worth wiring is the one where a person still presses send: the agent drafts the chaser, the human signs it. Zoho’s own walkthrough for building that agent says it plainly, that agent output should never reach a customer without a human in the loop. The deduction covers the wiring. It does not cover the send.

Footnotes

  1. The framework is non-binding, and IMDA describes it as a living document, open to feedback and case studies as the technology evolves. The deduction is administered by a tax authority. Singapore has made the incentive enforceable and the safeguard advisory, which is a fair guide to which of the two an operator reads first. ↩