Half of Singapore enterprises now use agentic AI. One in ten has changed how work moves because of it. That gap is the whole story of AI adoption in 2026, and it is where the forecast return on investment quietly disappears.
The figures come from the ServiceNow Enterprise AI Maturity Index, published on 17 August 2026. It surveyed 4,500 senior leaders across 19 countries, including 200 in Singapore. Agentic AI adoption rose from 22% to 51% in a single year. Only 10% had redesigned processes so that AI completes multi-step business tasks end to end.
What this is and why it matters now
Agentic AI describes systems that take a sequence of actions without a human approving each step. In human resources that means an agent which monitors a source, interprets what it finds, maps it against existing records and raises a specific task. It is not a chatbot that answers a question when asked.
The distinction matters commercially because the two are bought the same way and return value differently. Adding an assistant to an existing process makes the existing process faster. Redesigning the process around what an agent can do changes the process. ServiceNow reports that enterprises which redesigned work entirely around AI were nearly four times more likely to report significant productivity and efficiency gains. The comparison is against those that layered AI on top of existing ways of working.
That fourfold difference is the reason the distinction is worth labouring, because the money has already moved. Singapore AI budgets rose 108% year on year and now account for 15.4% of IT spending. That money meets a mixed maturity picture. The largest group, 33%, uses AI to help individual employees work faster. A further 18% report no progress on advanced adoption at all, above the global average of 11%.
This piece is for HR and operations leaders who have bought AI capability and are now asked what it returned. It is also for the compliance partners who must govern whatever gets built next.
The measurement gap sits underneath the redesign gap
Before arguing about which processes to redesign, there is a more pressing problem that most organisations cannot tell which processes are working. The SHRM State of AI in HR 2026 report was fielded in December 2025, with 1,722 HR professionals responding. It found 56% of HR functions do not formally measure the success of their AI investments. Only 16% use a return-on-investment metric.
An organisation that does not measure cannot distinguish a tool that saved time from one that moved effort somewhere less visible. It also cannot build a case for the redesign work, because redesign is expensive and its justification is comparative. Without a baseline, “we should restructure this process around the agent” is an assertion.
The sequence therefore runs backwards from how most programmes are planned. Measurement is not the reporting layer applied after adoption. It is the precondition for deciding what to redesign.
Why do so many AI investments go unmeasured?
Because the easiest thing to measure is usage, and usage is not a result. Counting prompts, seats or tool purchases produces a number that rises reliably and says nothing about whether the work got better. A useful measure names the outcome the process exists to produce, and tracks it before and after.
What redesign looks like in an HR function
Measurement tells you where to look. The redesign itself is more specific than the word suggests, and smaller.
Smaller, because redesign happens one process at a time. Take multi-jurisdiction compliance monitoring. The AI-enabled version gives the compliance analyst a summarisation tool, so reading legislative updates takes less time. The analyst still decides what to read, still tracks which jurisdictions were covered this week, and still maps findings against policy by hand.
The AI-native version inverts it. An agent scans the sources continuously, interprets what changed, maps it against the live policy library and flags only what touches an existing policy. The analyst’s role moves from finding to judging. The queue is no longer “everything published this month”. It is “these four things touch our policies”.
Moving the analyst from finding to judging is the whole redesign, and the work that makes it real is not model selection. It is deciding what the agent owns, what the human owns, what gets recorded, and what the escalation path is when the agent is wrong. Aicadium People Operations runs this pattern across seven jurisdictions, and the governance design took longer than the technical build.
Why HR is usually absent from the decision
Redesign in HR requires HR in the room, and frequently HR is not. SHRM found 52% of organisations do not involve HR in AI strategy or vision. Leadership of AI initiatives typically sits with IT, legal or cross-functional teams.
This produces a predictable failure. AI gets deployed into people processes by functions that do not own the fairness, explainability or employee-relations consequences. The tool arrives, the workflow does not change, and the governance question surfaces after the first contested decision.
That failure is not for want of capability. Singapore is ahead on the structural foundations that would support redesign. ServiceNow found 28% of Singapore enterprises have formal processes for testing, auditing and managing AI risk, against 20% globally. A further 23% have replaced legacy technology, against 16% globally. The foundations are further along than the operating model built on them.
Is HR involvement a governance requirement or a performance one?
Both, and the performance argument is the one that gets heard. A process redesigned without the function that owns it produces a workflow people route around. The governance exposure follows from the same cause, which is why treating HR involvement purely as a compliance control understates it.
The honest version of the adoption number
Adoption figures are useful for market timing and misleading as a measure of maturity. An organisation that has bought agentic tooling, deployed it to individual employees and changed no process is counted in the 51%. So is an organisation that rebuilt a function around it. The number does not distinguish them.
Treating those two examples as equivalent is what makes the headline dangerous for planning. The more useful question for a 2027 plan is not what percentage of peers have adopted AI. It is which of your processes has a named owner, a measured baseline, and a documented decision about what the agent does and what the human does. Most organisations today fall short on this measure.
