Singapore IMDA Defines What Qualifies as Agentic AI; Dayos Featured as Reference Case Study for Architecture Replacing $1M+/Year ERP Support Operations

SINGAPORE — May 22, 2026 — The Infocomm Media Development Authority (IMDA) of Singapore on 20 May 2026 published Version 1.5 of its Model AI Governance Framework for Agentic AI, defining in a government-issued standard what qualifies an enterprise system as truly agentic and contrasting it with traditional workflow automation. Dayos Pte. Ltd. is one of 14 industry case studies featured in the framework. The framework was unveiled at ATxSummit during opening remarks by Singapore Minister for Digital Development and Information, Mrs Josephine Teo.

For Chief Information Officers and engineering leaders, the framework provides clear technical criteria. An agentic system is built from eight identified components: a reasoning model, instructions, memory, planning and reasoning, tools, communication protocols, controls, and logging and monitoring. The framework also draws a sharp distinction between two dimensions of agent design, action-space and autonomy, and requires both to be scoped, bounded, and accountable.

Most platforms marketed as agentic today were not designed against this architecture.

Traditional IT Service Management platforms focus on deflection and routing. They are workflow automation systems built on deterministic rules. They do not perform multi-step reasoning across enterprise systems, do not autonomously invoke tools to execute changes in financial or HR systems, and are not designed to satisfy the framework's definition of agentic AI.

This matters because the operational cost it leaves unsolved is significant. A typical enterprise spends more than $1 million per year on Oracle or Workday managed services, on top of an original implementation that cost between $5 million and $40 million. Across the installed base of more than 16,000 Oracle, Workday, SAP, and ServiceNow customers, cumulative implementation sunk cost totals approximately $320 billion. That figure is the reason migrations almost never happen, and the reason the same labor pool that built those implementations is still billing the tickets that follow. Public-record per-ticket pricing from local government Oracle Cloud customers shows a range of $1,744 to $4,722 per support ticket, with an industry average around $3,233. Approximately 30 percent of every ERP runs on manual workarounds that cannot be self-served. The labor supply problem behind this number is structural: two million computer science students graduate globally each year, and almost none of them are trained on Oracle, Workday, or SAP.

This is the layer where genuine agentic AI is most economically valuable. Multi-step reasoning across enterprise systems is exactly what large language models with planning, tool use, and structured controls were designed to do. ITSM platforms were not.

The Dayos case study, featured in the framework's "Assess and bound the risks upfront" dimension, documents the retirement of the company's own ServiceNow instance in 45 days, reducing annual licensing costs by $121,000. The replacement is an AI-powered ticketing agent built on Hero that uses three distinct reasoning strategies depending on ticket risk tier:

• Simple Feedback (Tier 1, 60% of tickets): A propose-confirm loop for password resets, access requests, and status enquiries. Every action produces a reasoning chain and a confidence score, audited biweekly.
• ReAct (Tier 2, 30% of tickets): A multi-step diagnostic loop for chart of accounts updates, integration mapping corrections, and failed API connection diagnosis. The agent checks logs, queries connected systems, identifies probable root causes, and writes a diagnostic summary with a proposed fix. A qualified engineer signs off before execution.
• No autonomous agent action (Tier 3, 10% of tickets): Production deployments, security changes, and permission modifications are excluded from autonomous handling.

This tiered architecture matches the framework's recommendation to bound agent autonomy by severity of impact, reversibility, and feasibility of human oversight.

Hero implements deterministic controls at the system level, not at the prompt layer, a distinction the framework specifically highlights, noting that prompt-layer safeguards "may be bypassed or 'forgotten'." Access controls at the tool layer prevent unauthorised tool calls regardless of agent reasoning. Each agent has a unique, catalogued credential. Every action produces an immutable, auditable reasoning chain.

Dayos holds SOC 2 Type 2 and ISO 42001:2023 certifications and is auditable against the new IMDA framework end-to-end.

The same architecture extends across both operational surfaces. Hero Enterprise applies the same controls, identity model, and reasoning strategies to net-new Oracle or Workday implementations. The case study in the IMDA framework documents the ticket-side proof; the Enterprise tier handles the implementation surface that produced the labor problem in the first place.

"The framework is the part most enterprise AI vendors will not engage with," said Brad McElhannon, Founder and Managing Director of Dayos. "Governance, evaluation, human oversight, and accountability for agents that take real actions inside production financial and HR systems. Without that layer, what gets called agentic AI is a demo. With it, agentic AI becomes operational infrastructure at software economics, replacing both the implementation cost and the ongoing support cost that have defined enterprise ERP for thirty years."

For CIOs, the framework provides procurement criteria that did not exist 30 days ago. Vendors that cannot map their product to the eight components, that cannot describe their action-space and autonomy in technical terms, that cannot demonstrate immutable logging and scoped identity, are not selling agentic AI. They are selling workflow automation with a chat interface.