AI Development in Singapore

About AI Development

Most AI projects fail in the same place: a demo works on ten hand-picked examples, then falls apart on real user input, real data volume, and a real invoice at the end of the month. We build the unglamorous parts that decide whether an AI feature survives contact with production: retrieval that returns the right context, evaluation you can run before every deploy, guardrails on what the model is allowed to do, and cost tracking per request so the bill is a number you chose rather than one you discover.

What the engagement includes

• Retrieval-augmented generation over your own documents and databases
• AI agents with tool access, scoped so they can only do what you allow
• Model Context Protocol (MCP) servers that connect assistants to your systems
• LLM integration into existing products: OpenAI, Anthropic, and open models
• Document extraction and classification for invoices, contracts and forms
• Evaluation suites so a prompt change cannot quietly break output quality
• Cost and token controls with per-request tracking and hard budget limits
• Chatbots and support assistants grounded in your own content, not the open web

The AI tax deduction, the data-analysis copyright exception and MAS's AI proposals in a Singapore build

Budget 2026 changed how commissioning AI development is taxed in Singapore, and the invoice is where that shows. IRAS now lists adoption of AI as a qualifying activity under the Enterprise Innovation Scheme for the years of assessment 2027 and 2028, with a 400 per cent tax deduction on the first S$50,000 of qualifying AI expenditure each year for a business carrying on active operations in Singapore. Qualifying AI expenditure covers subscribing to or licensing an AI system from another person and acquiring a qualifying AI business service, which IRAS defines as a service supporting the adoption or development of AI in the business and relating to the development, deployment, operation or maintenance of an AI system, expressly including system development, consultancy, data and analytics, and system engineering and compliance services. Three exclusions shape the paperwork. Physical infrastructure and hardware, such as servers, storage and computing equipment, are out, so a project that bundles a GPU machine with the build prices them separately. Spending subsidised by a government grant is out, so the same cost cannot sit in a grant claim and in this deduction. And where one payment mixes AI and other work, only the AI portion qualifies, apportioned on a reasonable basis if the evidence does not separate it. That last rule is the practical one, because an AI feature usually arrives inside a larger web or mobile build. So our statements of work and invoices split the AI system's development, evaluation and deployment from the surrounding application, each line described in IRAS's own terms, rather than leaving your accountant to reconstruct the split from a single project fee at year end. The cash payout option some other EIS activities offer does not apply to this one. YA 2027 is assessed on the financial year ending in 2026, so for a company with a December year end, AI work invoiced this year may already be in scope. Whether and how much you can claim is for your accountant.

Singapore's Copyright Act gives training a model on other people's material an express statutory footing, with conditions specific enough to design a data pipeline around. Section 244 of the Copyright Act 2021 makes it a permitted use to copy a work for computational data analysis, which section 243 defines to include using a work as an example to improve how a computer program handles that type of information, with training a program to recognise images as the Act's own illustration. The conditions are that the copy is made for that analysis or to prepare for it, is not used for anything else, is not supplied to anyone except to verify the results or for collaborative research on the same purpose, and that you had lawful access to the source. The Act's illustrations say access is not lawful if it came from circumventing a paywall, or in breach of a database's terms of use, ignoring any terms that section 187 makes void. Section 187 voids any contract term that tries to exclude or restrict this permitted use, for contracts made before, on or after 21 November 2021, and section 188 voids a foreign choice of law clause imposed mainly to evade it. In the pipeline that becomes a provenance record for every source in the training and retrieval sets: where it came from, how it was accessed, under what terms and when it was copied, kept with the dataset. Scrapers stop at logins and paywalls rather than working round them. The training copy lives in a store used only for that purpose, never reused as a content library or served back to users, and shared only for the verification or research the section allows. The exception covers making copies, not what the model later outputs. It is also Singapore law, so where any copying is done by engineers outside Singapore, which law governs that copy is a point to settle first. That, and whether a given source qualifies, is for your own counsel.

For a bank, insurer, payment firm or other MAS-regulated institution, the AI system we build will be judged against expectations MAS has set out in draft. MAS consulted from 13 November 2025 to 31 January 2026 on Guidelines on Artificial Intelligence Risk Management that would apply to all financial institutions in proportion to their use of AI, and proposed a transition period of 12 months after the final guidelines are issued; whether they have been issued, and when your clock starts, is for your compliance team to confirm. The proposals are concrete about what a build must leave behind. The institution is to identify AI use across the firm, whether developed in-house or by third parties, and keep an accurate inventory recording, among other things, the data used, dependencies, lifecycle status, the risk materiality rating, validation status, owners and developers, and links to documentation. Materiality is assessed on impact, complexity and reliance, and life cycle controls, from data management, fairness and explainability to human oversight, evaluation, reproducibility, monitoring and change management, are applied in proportion to it. Two points land on an outsourced build. Third-party AI is defined to include systems, models and the data used for AI, so a hosted model API, an open-source model and an outside dataset each become entries with their own due diligence, and the institution is expected to test them in its own use cases and arrange to be told when they change. And the development process is to be documented well enough for an independent reviewer or auditor to understand and potentially replicate the system and its results. So the handover is an inventory entry per model and dataset with versions pinned, the evaluation sets and their results, the prompts and configuration live for each release, and a change log tied to deployments, rather than a repository and a demo. Which parts apply to your institution, and on what timetable, is for your compliance team.

How we work with Singapore clients

Singapore is the one market we serve where the working day is simply shared. We are two and a half hours behind you and neither country moves its clocks, so our morning starts partway through yours and the whole of your afternoon is live time for both sides. That changes the shape of an engagement more than anything technical does: a call is booked for the same afternoon rather than the same week, a question raised before your lunch is answered after it, and the written-first method we run for North American clients is a courtesy here rather than a necessity. We keep the written record anyway, with a default chosen on every open decision, because it is the thing an auditor or a new hire can read later and a call is not.

Invoicing is worth stating plainly because the GST position confuses buyers on both sides. We are outside Singapore, so no GST appears on our invoice. Whether you account for it yourself is the reverse charge question, and IRAS's own guidance is that a GST-registered business making only taxable supplies is not affected, while a partially exempt one accounts for it in its own return. We quote in Singapore dollars, the figure holds for the project whatever the rupee does, and payment by bank transfer or Wise avoids the card fees a five-figure invoice does not need. The contract can sit under Singapore law with Singapore courts, and unusually among our overseas markets that clause has teeth: Singapore is a reciprocating territory under section 44A of India's Civil Procedure Code, so a judgment of the Singapore High Court can be executed in India directly rather than re-litigated.

Indicative pricing in SGD

Converted from our published Indian rates and rounded. Exchange rates move, so the figure on your quote is calculated on the day it is issued.

Questions Singapore clients ask

AI Development: common questions

Other markets we work in