Singapore Budget 2026, delivered by Prime Minister and Finance Minister Lawrence Wong on 12 February 2026, made AI adoption cheaper for SMEs. The single new tax number you can actually bank on is specific and time-boxed: under the Enterprise Innovation Scheme, you can claim a 400% tax deduction on up to S$50,000 of qualifying AI expenditure per year, for Years of Assessment 2027 and 2028 only (Budget 2026 Statement, Section C; IRAS EIS page). At 400%, S$50,000 of qualifying spend produces a S$200,000 deduction.
Two things are worth saying plainly before you get excited. IRAS has not yet published what "qualifying AI expenditure" actually includes, and has committed to releasing details only by mid-2026. And this AI cap cannot be converted to cash, unlike the other Enterprise Innovation Scheme activities. The deduction is real and generous. It is also narrow, and it only holds up if you can prove what you spent on and what the AI did.
What Budget 2026 Actually Changed for SMEs
Section C of the Budget Statement is titled "Harness AI As A Strategic Advantage," and it moves on three fronts that matter to a small business: the tax deduction, a grant expansion, and a training push. Only one of them carries a firm new number.
The Enterprise Innovation Scheme (EIS) has existed since Budget 2023 and runs from YA 2024 to YA 2028. It already granted 400% deductions on qualifying spend across research and development, intellectual property registration and acquisition, and approved training. Budget 2026 added AI expenditure as a new qualifying activity, capped at S$50,000 per Year of Assessment, for YA 2027 and YA 2028 (Budget 2026 Statement, Section C, para 66a). The IRAS page carries this verbatim as a new item.
Keep one distinction clean. The S$50,000 AI cap is measured against its own separate ceiling. It is not pooled with the higher S$400,000 cap that applies to the main older activities such as research and development, IP, and approved training (IRAS EIS page). Do not blend the two. Your AI claim sits against its own S$50,000 line and nothing else.
The One Number You Can Bank On, and the Catch Underneath It
The 400% deduction on up to S$50,000 of qualifying AI spend is the anchor. Everything else in Budget 2026's AI push is either a standing scheme or a headcount target, not a new dollar figure.
Here is the catch most coverage skips. The older EIS activities carry a cash-payout option that lets a smaller or loss-making firm convert part of the deduction into cash. That option does not extend to AI expenditure (IRAS EIS page). So a pre-profit micro-firm gets no immediate cash from the AI cap this year, only a deduction it can set against future taxable profit. The incentive is worth the most to SMEs that are already profitable and paying tax.
That design has a quiet consequence worth naming. The firms the "AI for SMEs" story targets most loudly, the young and pre-profit ones, benefit least from the AI cap right now. And because IRAS has not yet defined qualifying AI spend, the temptation is to spend early to get costs on the books, then justify them later. That is exactly the pattern that fails an audit. Spend against a scope that does not exist yet, and you are guessing.
The Grant and Training Lines: What Is New and What Is Not
Budget 2026 also said it would "expand the PSG to support a wider range of digital and AI-enabled solutions" (Budget 2026 Statement, Section C, para 66d). That expansion is qualitative. The Budget attached no new rate or cap to it, and pushed the detail to the Committee of Supply.
The Productivity Solutions Grant terms you will see quoted, up to 50% of eligible costs and up to S$30,000, are the current standing terms published by EnterpriseSG, with SME eligibility set at group turnover of not more than S$100 million or group employment of not more than 200 (EnterpriseSG PSG page). Treat those as the current PSG figures, not as a Budget-2026 announcement.
On training, the number people will repeat is "100,000." It is a headcount target, not a grant. At the National AI Impact Programme on 2 March 2026, the Government committed to support 100,000 workers to become "AI Bilingual" and 10,000 enterprises over three years, and IMDA will expand its existing TeSA programme to non-tech workers, starting with the accountancy and legal professions (MDDI National AI Impact Programme factsheet). The Budget Statement also commits to redesigning the SkillsFuture website for clearer AI learning pathways and to six months of free access to premium AI tools for people who take selected AI training courses (Section C, paras 76-79). No dollar figure is attached to TeSA in any primary source. Do not expect a S$100,000 cheque.
What an SME Can Claim, at a Glance
| Scheme | What it covers | The cap or limit | The catch or precondition |
|---|---|---|---|
| Enterprise Innovation Scheme (AI enhancement) | Qualifying AI expenditure (scope pending IRAS guidance by mid-2026) | 400% deduction on up to S$50,000 per Year of Assessment, YA 2027 and YA 2028 only | No cash payout for AI spend, unlike the older EIS activities. Loss-making firms get no immediate benefit. Qualifying scope still pending IRAS guidance by mid-2026. |
| Productivity Solutions Grant | Pre-approved digital and AI-enabled solutions (Budget 2026 widened the range qualitatively) | Up to 50% of eligible cost, up to S$30,000 (current EnterpriseSG standing terms, not a Budget-2026 figure) | SME eligibility: group turnover not more than S$100M or group headcount not more than 200. AI-expansion detail deferred to Committee of Supply. |
| TeSA / AI training | AI capability building, starting with accountancy and legal; SkillsFuture pathways; 6 months free premium AI tool access with selected courses | Target of 100,000 workers "AI Bilingual" and 10,000 enterprises over 3 years (a headcount target, not a grant) | No dollar figure attached to TeSA in primary sources. This is capability funding, not a per-firm payout. |
Why a Governed Deployment Is the Precondition, Not an Afterthought
The tax break rewards AI spend, not AI results. So the burden of proof falls on your records. Because IRAS has not yet defined qualifying AI expenditure, and will only do so by mid-2026, the safe posture is to be able to show exactly what was bought and what the AI did. "We spent money on AI" is a sentence, not evidence.
This is where IMDA's Model AI Governance Framework for Agentic AI earns its place, though not in the way people assume. Announced by Minister Josephine Teo at the World Economic Forum in Davos on 22 January 2026, it is voluntary guidance that builds on the 2020 Model AI Governance Framework, and it states plainly that humans are ultimately accountable (MDDI announcement). Its four dimensions are: assess and bound the risks upfront; make humans meaningfully accountable by defining significant checkpoints at which human approval is required; implement technical controls and processes across the agent lifecycle; and enable end-user responsibility through transparency and training.
Be precise about the connection. The governance framework is not a legal precondition to claim the EIS deduction, and nobody should tell you it is. The link is evidentiary, not statutory. The artifacts that good agent governance produces, a log of what the agent did, which actions were bounded, and who approved consequential steps, are the same artifacts that make a tax claim defensible if it is examined. One caution from us: the biggest failure mode is over-trusting an agent that has behaved reliably before. That risk applies to the agent doing your work and to the finance lead assuming the spend will obviously qualify.
Where Origin Pi Stands
Our read is direct. Singapore has, in effect, put a price on governed AI adoption, and the deduction only survives contact with scrutiny if the deployment can prove itself. Budget 2026 pays you for AI spend, not AI outcomes.
That makes four things the precondition for claiming with confidence, not compliance theatre: documentation of what was bought and why, bounded permissions so an agent cannot act outside its remit, a human confirm-step before consequential actions such as a payment or a database change, and an immutable audit trail of who approved what and when. Those four map cleanly onto the governance framework's dimension two, human checkpoints for approval, and dimension three, technical controls across the lifecycle. The same confirm-step and audit log that make an agent trustworthy are what turn "we spent on AI" into "we can prove qualifying, governed, human-accountable AI spend."
We will not overstate the benefit. The AI cap runs two years, is deduction-only with no cash route, and its scope is still pending IRAS guidance. The honest move for an SME right now is to build the governance habit first, describe qualifying spend qualitatively, and wait for the mid-2026 detail rather than guess at inclusions. Govern first, and the tax break is yours to keep.
