Agents scale on a decision. Guardrails scale on an integration project. Deloitte surveyed 3,235 technology and business leaders across 24 countries and found only 21% with a mature governance model for agentic AI. In the same research, 74% expect to be using agents at least moderately by 2027. - The gap is plumbing, not policy. What is missing is decision boundaries, real time monitoring that flags anomalies, and audit trails capturing the full chain of agent actions. - Retrofitting oversight is the slow route. Deloitte's finding is that skipping guardrails to move faster ends up costing more. - Winners start small on purpose. Organisations succeeding with agents begin on lower risk use cases and scale deliberately. - Governance is a group build. IT, legal, compliance and business unit leaders set policy, watch performance and own escalation together. Wire the oversight while the agent count is still countable.

Your agent holds tools and data. Your controls were written for a chatbot. Singapore's cyber agency has already said what it expects. On 17 June 2026 CSA published its addendum on securing agentic AI, designed to sit on top of the 2024 Guidelines and Companion Guide rather than replace them. It followed a public consultation that ran from 22 October 2025 to 31 December 2025. - Risk starts with the workflow map. CSA asks you to map the agentic workflow first and find where a threat actor could exploit the plan, the tools or the data access. - Three scenarios are worked through for you. App development and coding assistants, automated client onboarding and automated fraud detection each get an example at different levels of autonomy. - Controls span the lifecycle. The mitigations run across development, not as a checklist bolted on at launch. - Voluntary is not optional. The measures are voluntary and environment specific, but accountability sits with the system owner. Map the workflow before you widen what the agent can reach.

Four numbers, four decisions, none of which wait for you. This week ran from compliance timing through data readiness to the real cost of agents and the tools quietly leaving your stack. - Three days. Once you decide a breach is notifiable, the PDPC clock gives you three calendar days. The assessment that starts it is the part nobody owns. - Sixty percent. Gartner expects that share of AI projects to be abandoned where the data was never made AI ready. The failure is upstream of the integration. - Five times. Routing a task to an agentic reasoning model costs at least that much more than a chatbot reply, and total inference cost per workflow keeps climbing as prices fall. - Four dates. Imagen 4 endpoints, OpenAI o3 and the DALL·E GPT all retire this month, while the Sonnet 5 price rise was cancelled outright. Put all four on a calendar with a name against each one.

Three retirements, one cancelled price rise, all inside this month. Retirement has become a routine scheduled event rather than an exception, and August carries four dates that deserve a calendar entry and an owner. - Imagen 4 endpoints closed on 17 August. Google shut down the standard, ultra and fast generate endpoints and points callers at its Gemini image model instead. - OpenAI o3 leaves ChatGPT on 26 August. This is a ChatGPT surface change at the end of a 90 day sunset. The API is explicitly unaffected, so the blast radius is habits and saved workflows, not code. - The DALL·E GPT retires on 30 August. ChatGPT Images continues and user created GPTs with image generation stay live, so download anything you want to keep. - The Sonnet 5 price cliff was cancelled. Anthropic made the introductory $2 and $10 per million tokens permanent, so the planned rise to $3 and $15 on 1 September no longer applies. Check what you budgeted for September. One of these moved your way.

Cheaper tokens are not making agents cheaper. They fund bigger workflows. On 17 August Gartner named the trap directly. AI inference costs per agentic workflow will rise more than fivefold through 2028, even while model prices fall. Gartner calls it the Inference Paradox: better unit economics escalating total AI cost without a clear pathway to matching value. - An agent is not a chatbot with extra steps. Routing a task to an agentic reasoning model costs at least five times more than a basic chatbot interaction, and more as complexity grows. - Falling prices get spent, not banked. Every efficiency gain is absorbed by deploying more capable and more expensive models. - Tiering is the control that works. Route simple tasks to cheap models and reserve reasoning models for work that earns the premium. - Subsidy exists locally. Singapore's National AI Impact Programme will support 10,000 enterprises over three years, with pre approved solutions carrying grant support. Budget for the workflow, not the token.