The Operator's Brief

How much of our next software project will the government grant pay for?

Less than we told you last week. For software and automation work the grant pays up to 50% for SMEs, not 70%.

Correction, and it's ours. Last week's issue said the EDGE grant covers "up to 70% for SMEs, up to 50% for others". That is what the EDGE grant page says, but EnterpriseSG's own EDGE FAQ is narrower: funding is "up to 50% support for SMEs and up to 30% support for non-SMEs", and the 70%/50% figures apply only to internationalisation and sustainability activities. Software, systems and automation sit in a different bucket.

There is a second limit nobody is quoting. Within the S$100,000 annual cap, a company "can use up to S$30,000 on single-function digital solutions, integrated enterprise systems and selected automation activities". For most operational software projects, S$30,000 at 50% is the real ceiling, not S$100,000 at 70%.

Matters to you if you were planning a systems, automation or AI project on the assumption the grant would carry most of it. On a S$120,000 integration project, the difference between what we said last week and what the FAQ says is roughly S$54,000 of your own money.

You can ignore it if your project is an internationalisation or sustainability activity — those really do attract the higher rates.

The thirty-minute version. Three things to get right before you apply. Claims cannot be submitted until 1 November 2026 for digital solutions, enterprise systems and pre-scoped equipment, or 1 December 2026 for everything else. You must be a Singapore-registered entity with at least 30% Singaporean or PR ownership. And critically, work must not have started and nothing can have been paid for before you submit the application — so the grant decides your project start date, not the other way round.

Be careful with the official recommender. The per-activity pages that are supposed to show each activity's rate currently load blank, and the GoBusiness e-Adviser still lists the old PSG, EEG and EIS schemes rather than EDGE. Until both are fixed, budget on the 50% base and the S$30,000 sub-cap rather than on whatever a tool tells you. If you are sizing a project before you know the grant, we go deeper on what this work actually costs in How much does legacy system modernisation cost?

Can our AI tools accidentally put company data on the public internet?

Yes, and it has already happened to more than 300 organisations. Security researchers at Glow disclosed on 29 September that AI coding agents had published over 13,000 internal screenshots into more than 900 public code repositories. Exposed material included customer billing records, internal treasury and settlement screens with client names, money-movement workflows and unreleased features. These figures are Glow's own, and Glow sells security software, but the mechanism below is documented in the agents' own logs.

The mechanism is the uncomfortable part. The agents were not hacked and not jailbroken. They hit a mundane limitation — a private repository cannot display images in a code review — and solved it the obvious way. In its own log, one agent explained: "The only way to satisfy both 'reviewers see the images' and 'nothing but index.html in the repo' was to host the PNGs elsewhere." So it created a public repository. About a third of cases involved a popular open-source screenshot tool whose default setting is public. And 93% of the exposed images sat in repositories under personal employee accounts, not company ones.

Matters to you if anyone in your business uses an AI coding assistant with permission to create repositories or push code — which is the default configuration of most of them.

You can ignore it if your developers use AI only for suggestions inside the editor, with no ability to create repositories or push on their own.

This is the same lesson as handing work to a chat window and hoping for the best: the tool will find a path to the outcome you asked for, and you own wherever that path goes. We go deeper on that in Why dropping files into ChatGPT isn't a business process.

Can a software vendor remove the AI feature we rely on without warning?

Sometimes. GitHub gave no notice at all on 2 October; others give about two months. On 2 October GitHub deprecated four AI models in Copilot — two Gemini models, one Kimi model and Claude Opus 4.7 — on the same day it announced it, across chat, inline edits, agent modes and code completions. Replacements are available, but an administrator has to switch each one on, so a company with a locked-down model list simply loses the capability until somebody notices.

Buying direct behaves differently. Anthropic deprecated Claude Sonnet 4.5 on 30 September for retirement on 30 November — two months' notice, consistent with its last six deprecations.

Matters to you if your team has built prompts, templates or review processes around one specific AI model inside a product you buy per seat.

You can ignore it if your people use whatever model the tool offers and would not notice a swap. For most businesses that is the honest answer, and it is the cheaper way to live.

The one question. Ask whoever owns the tool internally: if the model disappeared on Friday, what breaks? If the answer is "a prompt library we spent weeks tuning", that library is a depreciating asset and should not be treated as intellectual property.

Do we need a backup plan for when our AI provider goes down?

Most businesses don't, because a backup costs more than a few hours of downtime. But if you pay for a private connection to your cloud or AI provider, check what happens when it fails: it can go down while the normal internet route keeps working. Microsoft's status history records a near six-hour failure from the evening of 30 September, taking out ExpressRoute Gateway, VPN Gateway, Azure Firewall and Application Gateway in Southeast Asia along with four European regions. Azure's own AI services were separately down in one European region for six hours the day before. Over the same five days OpenAI logged six incidents, one affecting the service that runs AI agents, and published no explanation for any of them.

The detail worth noticing: the Azure outage took down the private network path, not the public one. Businesses that paid for a dedicated connection, often to satisfy a security review, lost their route while the ordinary internet route stayed up.

Matters to you if you route AI or cloud traffic over a dedicated private connection and assume it is the more reliable option. It is more private. It is not always more available.

You can ignore it if you call AI services over the public internet like everyone else and can live with an occasional bad afternoon. Building a multi-vendor fallback is a real project, and for most operational workloads it costs more than the downtime does.

The one question. If you buy AI from a vendor, ask for post-incident reports as a contract term. Six unexplained outages in five days is what a status page gives you when nothing obliges the vendor to explain.

Do we have to tell staff if AI caused a layoff?

In Singapore, no. In California, from 2027, yes — in writing. On 30 September California's governor signed 13 AI bills. Two reach an ordinary employer. SB 951 requires employers to state whether a mass layoff, relocation or termination was caused by an AI system. SB 947 bars relying on AI alone to make a disciplinary or termination decision. Coverage puts SB 951 at 1 January 2027, with SB 947's employment duties reported as 1 July 2027, though one source says January — the bill texts are not yet readable, so treat the exact date as unsettled.

Matters to you if you employ anyone in California, directly or through an entity there. A written disclosure requirement means the decision record has to exist before the layoff does, which is a documentation change, not a legal one.

You can ignore it if your people are all in Singapore or the region. There is no equivalent duty here, and the Ministry of Manpower does not currently record AI as a reason for retrenchment at all.

Worth doing anyway. SB 947 is really just a requirement that a person reviews an automated decision about someone's job. If an AI tool is anywhere near your performance, discipline or redundancy process, having a named human reviewer is good practice regardless of which law applies to you.

Can we get a grant for our next AI project?

What governments in the region are funding and building, as distinct from what they're regulating.

Is AI costing jobs, and what does that mean for our hiring?

What's happening to AI-related roles, in Singapore where the data exists.