When should we let an autonomous AI agent do the work?: Opinion

The rails for a more autonomous, agentic version of e-commerce are built. Artificial intelligence agents can now purchase products and services on our behalf. Current adoption levels are low, but swiftly rising.

The fundamental question the current AI excitement overlooks is: When is it worth handing work to an AI agent rather than a person?

Consider a scenario involving several purchases. An AI agent that pays as it goes runs up three kinds of costs:

• Inference cost accounts for the computing required for it to reason and act.
• Verification cost is involved in confirming the result is correct and the agent did what one asked.
• Settlement is the fee to make each related payment final, a fraction of a cent on the rails these agents use to pay, small but never quite zero.

To answer our question, we therefore need to understand how these three costs arise in more detail.

In the task-based model of Nobel laureate Daron Acemoglu and Pascual Restrepo, a task is worth automating once a machine can do it more cheaply than a person, which is usually taken to point towards ever more work being automatable as machines improve.

That comparison weighs the cost of doing the work, but it carries no cost for verifying it.

My new working paper, presented during the Quantitative Finance 2026 conference at the National University of Singapore in June, points to the verification cost as the one that largely decides where automation through agentic AI stops paying off once agents transact on our behalf, and eventually interact with other agents.

Christian Catalini of the Massachusetts Institute of Technology and his co-authors called it the “cost to verify” in their recent article.

Why verification is the catch

Picture an AI agent planning a complicated trip for you: multiple flights, trains, hotels, restaurants, a visa.

Each booking can be perfectly correct on its own, yet two (or more) of them can still clash: a flight that lands after the last train, a hotel booked for a night you are meant to be elsewhere.

To be sure the whole trip holds together, the agent cannot simply verify each booking on its own. It has to verify that they all fit together.

And that is the challenge. Each task one adds to a job can clash with every task already there, and with every pair of them. What has to be verified is every way individual tasks can interact with one another. Those combinations add up fast, and so do their costs.

Inference costs grow task by task, but verification costs grow combination by combination. Length alone is not the driver of a task’s verification cost. A long job of independent tasks stays cheap to verify. A shorter job whose tasks all depend on one another does not.

This is what carves out a narrow band where automation through agentic AI pays off. Very small tasks are not worth it, because their value may be smaller than the cost of settling related payments.

Tasks with densely interdependent actions are not worth it either, because verifying them costs more than asking a capable person to do the work. Autonomous AI agents win in the middle – a Goldilocks zone.

The hidden discount

There is a further catch, one that should worry anyone budgeting for an agentic future. AI today is sold well below what it costs to run: OpenAI is reported to be heading for losses of around US$14 billion this year.

Because verification is itself done by an AI, that discount makes both the work and the verification look cheaper than they really are.

Take the discount away, as it must eventually go if OpenAI and its rivals are to become sustainable businesses, and the viable band for automation shrinks from both ends.

A good part of what looks worth automating today may only add up while the technology is being sold at a loss.

Beyond financial services

On Jul 3, the Monetary Authority of Singapore, with a group of banks and fintechs, published a framework called Safeguards for Agentic Finance at Runtime (SAFR).

Its role is to verify and record what an AI agent proposes to do before it does it, so that while it is acting on our behalf, it aligns with our intent.

Finance is a natural place to start implementing guardrails. But the economics discussed here is not particular to financial services.

The same ceiling applies wherever an AI agent strings together many interdependent purchases and pays for each as it goes, as in our earlier trip example.

It is also why SAFR, sensibly, leans on explicitly-bounded mandates for AI agents: keeping the pieces small and independent is what keeps verification affordable.

Firms, and eventually individuals, from all different industries should test whether their sums still add up at full, unsubsidised prices before committing real money to autonomous AI agents.

A business that only works while AI is sold at a loss is quietly relying on someone else to foot the bill.

The agentic economy is real and worth building. The edge for Singapore will be, as usual, in asking the sharp question early: what to agentify and, perhaps more importantly, what not to.

The writer is the founder of Singapore-based Lightbulb Capital, an innovation research and advisory firm focused on financial services

Source: The Business Times © SPH Media Limited. Permission required for reproduction.

ADV: A Practical Workshop on Building & Construction Dispute Resolution Teen who allegedly licked iJooz straw has case adjourned as ICA checks his student pass status