The EDGE Grant and AI projects: what it covers, what it doesn't, and when to apply

As of 30 October 2026. Grant terms change; EnterpriseSG's EDGE Grant page is the authority, not this post.

Who this is for

You run a Singapore SME, and someone has asked you to "put AI on" a document-heavy process: supplier invoices, insurance claims, loan applications, complaints. The next question is usually whether a grant will pay for part of it. The answer depends less on the technology than on the order you do things in.

What EDGE is, and the two numbers that matter

EDGE replaced three schemes. In EnterpriseSG's words, it "streamlines EDG, MRA and PSG into a single grant". Those schemes ceased on 29 September 2026, and new applications go to EDGE from 30 September.

The cap. A company can receive up to S$100,000 a year across all EDGE activities, refreshed on 1 April. Within that, at most S$30,000 can go to single-function digital solutions, integrated enterprise systems and pre-scoped equipment.

The support level. EDGE's landing page says "up to 70% for SMEs". The FAQ is more precise:

• In general: up to 50% for SMEs and up to 30% for non-SMEs.
• Internationalisation and sustainability activities: up to 70% for SMEs and up to 50% for non-SMEs.

The activity whose conditions name AI/ML, Large-scale Digital Transformation, states up to 50% for SMEs. Plan on 50%.

Who applies. The applicant must be registered in Singapore with at least 30% local ownership. The company applies and pays, not the vendor. Submissions "must not be made by vendors, service providers, or any other third party" on the applicant's behalf. A vendor who offers to "handle the grant for you" is offering something the rules don't allow.

The sequencing rule most companies get wrong

EDGE does not support retrospective applications, and the FAQ defines retrospective plainly. An application is retrospective if the applicant "has started work or made any payment or deposit to a supplier, vendor, or third party prior to application submission". A deposit paid to hold a vendor's slot is enough to lose the grant.

Payment made after you submit but before approval may be supported if the application is approved, at the company's own risk. In practice: submit before you sign the build contract, start the build or pay anything towards it.

For an AI project that starts with an assessment, that sets the order:

1. The assessment is paid directly and sits outside the grant. It is a separate engagement that measures the process on your own past invoices or claims. It is not the build, and it is not claimed.
2. The build is scoped from the assessment. The assessment gives you the numbers the application needs.
3. You apply before the build contract. Only after submitting do you sign and pay for the build.

If you are unsure whether earlier scoping work counts as "started work", ask EnterpriseSG before you apply, not after.

What assessors look at, and what a strong application contains

The Large-scale Digital Transformation activity sets the conditions. The project must:

• "demonstrate measurable impact on company growth and efficiency";
• use technologies such as "AI/ML, RPA, or workflow automation";
• involve integration across systems.

It excludes "automation of isolated tasks without broader business transformation". A tool that digitises a paper form and connects to nothing is out.

The application asks for:

• a project proposal;
• vendor proposals and quotations;
• your latest financial statements;
• a three-year projection of revenue growth and cost savings.

At the end, you submit a project report with "qualitative and quantitative benefits". For digital solutions, the FAQ adds that supportability is assessed on "business use and reasonableness".

Vendors are assessed by activity.

• Activities with a pre-approved vendor list: the FAQ says to choose from it, because those vendors "have been assessed for their credibility".
• Activities without a list: you may engage a vendor of your choice and are expected to do your own due diligence.
• Consultancy costs: supported only if the consultant holds SS 680 certification and is "deemed suitable for the project".

So a strong application for an AI project contains four things:

• A measured baseline. What each invoice or claim costs you to handle today, from your own records rather than an estimate.
• A defined scope. One process, its inputs and channels, and the systems it touches.
• Success numbers. The share handled without a person, the error rate on those, and the cost of handling each one afterwards.
• A named vendor. With a proposal and quotation that match the scope.

What EDGE doesn't do

A grant changes who pays, not whether the project works. A project that can't name its baseline before it starts can't show its impact at the end, grant or not.

No grant covers a pilot that never reaches production. Most pilots stall on what they skip, not on the model:

• real inputs;
• the surrounding systems;
• the languages documents arrive in;
• who reviews the uncertain ones.

There's more in why most AI pilots never reach production. Half the cost of a demo is still the cost of a demo.

If an off-the-shelf tool would do, there are faster routes:

• SMEs Go Digital. IMDA's programme lists pre-approved solutions with up to 50% grant support.
• EDGE's single-function solutions. The list includes AI accounting automation.
• IMDA's GenAI Sandbox. In its 2024 second round, it funded up to 50% of a three-month trial of curated tools in web design, customer engagement and HR. Check IMDA for current rounds.

Three questions to answer before applying

These come from our fit test. They decide whether there is a project to apply for at all.

1. Does the process end in a decision from a fixed list of outcomes? Approve, hold, reject; route to a team; code to an account. If every one ends in one of a known set, an agent can learn to pick one, and hand over when it can't.
2. Do you have past invoices, claims or applications, with how each was decided? That is your baseline. Without recorded outcomes there is nothing to measure against and nothing to put in the application.
3. Is there someone who can say what a correct decision is, and spend a few hours a week saying it? During the parallel run, someone has to compare the agent's decisions with what should have happened. Without that person, nobody can say whether it works, and the project report has no numbers in it.

How we work with it

We start with Discovery: three weeks that end in an assessment of one process, measured on your own past records. Discovery is paid directly; we never make it conditional on a grant. If the numbers work and you want to apply, the assessment supplies the baseline, scope and success numbers. We help you prepare the application, and you submit it from your own account before the build contract is signed. The build then runs in parallel with your team until you trust it, and is switched on one decision type at a time. If you have one process in mind, book a 30-minute call about it.