Understanding agent economics

The issue is becoming more urgent as agentic AI moves from experiments into day-to-day work. KPMG’s Q2 2026 AI Pulse found that 55% of Asia-Pacific companies had delayed or scaled back AI-agent rollouts after expected operating costs began to exceed the value generated. Only 5% had demonstrated a return on investment. Yet 70% still plan to spend more than US$50m on AI over the next 12 months.

More than tokens

Token consumption alone, however, is the wrong yardstick. ‘One of the most important indicators is whether an AI solution materially improves an end-to-end workflow rather than simply automating an isolated task,’ says Kelvin Wong, chief technology officer at KPMG in Singapore.

‘If adoption is strong and the solution significantly enhances how work is performed, we scale it. Where utilisation or outcomes fall short of expectations, we redesign the use case to align it more closely with business needs,’ he adds.

Measure the workflow

AI agents are already embedded in professional-services work. At KPMG in Singapore, generative AI supports audit tasks including document extraction, flowchart generation and navigation of accounting and auditing knowledge. Analytical AI helps direct attention to higher-risk areas, including transaction scoring. In tax, research and advisory agents help professionals navigate legislation, guidance, firm methodologies and engagement documents.

‘Overall, the strongest business case has been in reducing the time spent on information discovery and synthesis, even as professionals continue to apply their professional judgment,’ says Lyon Poh, partner, head of corporate transformation and AI CoE lead at KPMG in Singapore.

Deloitte’s Omnia audit and assurance platform includes agents that help professionals identify risks, analyse evidence, draft documentation and preliminary conclusions for review, and support compliance evaluations. Crucially, the agents sit inside the audit workflow rather than in a separate tool.

That integration shapes how the firm measures them. For each workflow, Deloitte tracks changes in human effort, cycle times, rework, quality and risk coverage, as well as the overall cost of completing the work, and whether the capacity freed up is reinvested in client service or higher-value work.

‘In terms of technology, we monitor factors such as model consumption, infrastructure costs and deployment economics. However, we do not make investment decisions based solely on metrics such as cost per token,’ says Kenneth Tay, technology and transformation partner at Deloitte Southeast Asia.

Where the costs hide

The model itself is often the least of it. ‘One of the biggest misconceptions about AI is that the primary cost lies in the model itself,’ Tay says. ‘In reality, the more significant costs often emerge when organisations move from pilot projects to enterprise-scale deployment.’ A pilot has clean data and a handful of users; production demands integration, security reviews, compliance controls, monitoring and real-world exceptions.

Wong sees the same pattern at KPMG. ‘The biggest cost challenge often pertains to the complexity that arises from scaling many agentic use cases across various teams and functions,’ he says. The firm’s answer is to treat agentic AI as ‘an enterprise capability, rather than a collection of standalone tools’, with reuse, trusted AI controls and clearly defined business outcomes agreed before anything is scaled.

Match tools to tasks

Another common source of waste is a mismatch between the task and the model used to perform it.

‘A lower-cost model may reduce inference costs, but if it generates outputs that require substantially more professional review and rework, the apparent savings quickly disappear,’ Tay says. ‘Equally, using a large frontier model for a routine classification or extraction task may not be economically justified if a smaller model can achieve the same outcome with comparable quality.’

Deloitte brings these controls together under a single governance model. Teams across the firm are expected to reuse existing agents, connectors and design patterns rather than build their own. Each workload is tracked against outcomes, routed to the cheapest model that can do the job and held to cost thresholds. And over the life of each tool, it monitors adoption, quality, risk and realised value. Agents that perform are scaled up; those that fall short are redesigned or retired.

Trust is part of the value

Cost optimisation ultimately cannot be separated from trust. ‘Trust has to remain part of the economics,’ Tay says. ‘In professional services, an outcome is only valuable if it is secure, explainable, compliant and supported by accountable professional judgment.’

Poh points to human oversight, access controls, data security and rigorous quality standards, backed by audit trails that identify accountable owners and show how conclusions were reached. Tay adds that agents should operate within defined authority boundaries and least-privilege access, subject to the same independence and conflict requirements as a professional.

The maker-checker principle still applies. ‘Although sometimes both the maker and checker functions may be AI agents, a robust governance framework is nonetheless essential for ensuring responsibility and accountability,’ Poh says.

If agents take on the research and drafting junior staff once did, Poh argues, managers will need deeper critical thinking – and pay may shift towards outcomes instead of time spent. For Tay, that is the measure that matters. ‘Success will increasingly be determined by the cost and quality of outcomes,’ he says, ‘rather than the cost of individual technology components.’