One week, four shifts, all pointing at the same gap. Attackers upgraded first: CSA logged a 142% rise in infected infrastructure and warned agentic AI compresses intrusions into hours. Custom connectors are legacy: MCP is now the default integration standard, cutting development overhead by around 30%. Tax is a budget lever: Budget 2026 gives a 400% deduction on up to S$50,000 of qualifying AI spend. Cheap models got good: Gemini 3.6 Flash and Kimi K3 shipped at a fraction of frontier pricing. Secure it, standardise it, subsidise it, then right size the model.

Frontier capability is no longer where the interesting pricing sits. Gemini 3.6 Flash launched 21 July at $1.50 and $7.50 per million tokens. Kimi K3 landed 17 July at $3 and $15 per million tokens, with open weights promised. Flash tier got cheaper: Gemini 3.6 Flash cut output pricing roughly 17% while keeping a 1M token context window. Open weights are viable: Kimi K3 and DeepSeek V4 Flash give SG firms self hosting options for sensitive workloads. Match the model to the task: High volume support and summarising rarely need a frontier model at frontier prices. Audit your default model: Most teams route every request to one expensive model by habit, not by design. Route by task, not by brand, and your AI bill falls without touching quality.

The integration standard war is over. Most SME roadmaps have not noticed. 41% of software organisations now run MCP servers in limited or broad production. Forrester links standardised context protocols to a 30% cut in development overhead. One server, every client: Build one MCP server per system and any compliant AI tool can use it. Connector maths finally inverts: Connector count stops multiplying every time you add a model or a tool. Vendor lock in weakens: MCP is model agnostic, so switching providers no longer means rewriting integrations. Governance is the real blocker: Permission scoping and audit logs must be designed in, never retrofitted. Before commissioning another custom API, check whether an MCP server already exists.

Your AI business case likely ignores the largest subsidy available to you. The Enterprise Innovation Scheme gives a 400% tax deduction on up to S$50,000 of qualifying AI spend. It applies to YA2027 and YA2028, with no cash payout option for AI expenditure. Model the after tax cost: S$50,000 of qualifying spend deducts S$200,000 from taxable income for a profitable firm. Cash payout is excluded: Loss making firms get no liquidity benefit, so timing spend against profitability matters. Training and consultancy qualify: Qualifying AI training, consultancy and software development all sit inside the scheme. Stack it with PSG: PSG still cofunds up to 50% of approved AI tools, capped at S$30,000 per year. An AI business case without the tax line understates your real return.

Attackers now have agentic AI. Most Singapore SMEs still have a spreadsheet. Infected infrastructure in Singapore rose 142% to 284,300 detections in 2025 and ransomware cases rose from 159 to 165, with SMEs disproportionately affected. 1. Agentic AI compresses attacks: CSA warns autonomous systems can automate the kill chain and shrink intrusions from days into hours. 2. SMEs remain the soft target: Lower cyber maturity and thin resources keep smaller firms overexposed to ransomware. 3. Certification now covers AI: Cyber Essentials and Cyber Trust were expanded in 2025 to include mandatory cloud and AI security requirements. 4. Funding covers advisory work: Eligible SMEs get up to 70% funding support under CSA's CISO as a Service programme. Your AI security posture is now a procurement question, not an IT one.