The Agent Is Cheap. The Dependency May Not Be.

The Agent Is Cheap. The Dependency May Not Be.

What happens when we replace a wage bill before understanding the full cost of automated work?

A business replaces part of a software team with AI agents. Output rises. The salary bill falls. The subscription looks modest beside the cost of employing people.

On the spreadsheet, the decision appears obvious.

But are we comparing the full cost of one system with the introductory price of another?

An employee’s cost is relatively visible: salary, employer contributions, equipment, training and management. An agent’s subscription is an access price. Behind it sits a chain of model providers, cloud platforms, datacentres, chips, memory, electricity and capital.

Some of those costs are included in the price. Others may be funded by investors, absorbed elsewhere in a profitable business or supported by expectations of future revenue.

That does not make the technology uneconomic. It means today’s price alone cannot tell us whether the economics are sustainable.

The infrastructure bill belongs to someone

AI can feel almost weightless at the point of use. We type a request and receive an answer.

The machinery behind that exchange is anything but weightless.

There are assets to build, equipment to replace, power to secure and financing obligations to meet. Alphabet has explicitly described rising infrastructure investment as a source of higher depreciation and datacentre operating costs. Alphabet’s earnings discussion

Those obligations can be distributed across several businesses. The organisation selling access to an agent may own little of the infrastructure on which it depends.

We should be precise here. Outsourced infrastructure is not necessarily “off balance sheet,” and leasing does not make obligations disappear. Microsoft, for example, discloses datacentre leases and associated commitments. Microsoft’s annual report

The economic question is simpler: who ultimately pays enough to sustain the whole chain?

Losses at one layer do not establish that every layer is unprofitable. Equally, profitability at a cloud provider does not establish that every agent business built on top of it has a viable model.

Cheaper tokens do not guarantee cheaper work

It is tempting to conclude that prices must rise when investors demand returns. That is possible, but it is not inevitable. Better hardware, more efficient models and competition could continue reducing the price of individual operations.

The total bill can still rise.

An organisation may automate more tasks, run more agents, retain more context and introduce additional checking because the consequences of errors have grown. A cheap attempt can become an expensive outcome after retries, verification and recovery.

The useful measure is therefore the cost of a completed, accepted piece of work, including the effort needed when something goes wrong.

That is the comparison a wage bill deserves.

We may also be spending our understanding

There is another cost that is harder to put on a spreadsheet.

A team can generate software faster than it can develop a shared understanding of that software. It may know what a feature does without understanding why it was built that way, which assumptions it depends on or how it will behave under unfamiliar conditions.

This risk existed before AI. Agents could allow it to accumulate much faster.

Tests and documentation help. Neither automatically gives an organisation the ability to diagnose a difficult failure or judge the consequences of a major change.

If experienced people leave while the codebase expands, the business may gradually lose the ability to challenge the system it operates. Every unfamiliar fault then requires more assistance from the same tools on which it already depends.

The software remains in the repository. The practical ability to maintain it may have moved elsewhere.

That is a form of dependency worth measuring.

The price matters differently once you cannot leave

A price increase is manageable when there are credible alternatives.

It becomes more consequential when workflows, knowledge and operational capacity are deeply tied to a supplier—and the people who could rebuild the alternative are gone.

Switching then involves more than choosing another model. It can mean reconstructing knowledge, validating behaviour, retraining staff and accepting disruption.

An organisation could reduce its cost per task while becoming less able to control its future costs.

Before removing a team, it should ask:

• Can we move this work to another provider?
• Can we maintain the resulting systems ourselves?
• What happens if access is interrupted or the commercial terms change?
• Have we retained enough expertise to recognise when the output is wrong?

These questions belong in the business case, alongside the expected savings.

A saving for the business is not the whole economic result

There is also a public question.

Employment supports tax receipts through wages and employer contributions. If work moves from employees to automated services, the location and form of taxable activity can change.

It would be inaccurate to say that AI pays no tax. Companies, infrastructure owners and their employees can all contribute tax revenue. The question is whether the new activity replaces the revenue associated with the employment it displaces—and in which country that revenue arises.

Productivity gains could create new businesses, employment and taxable profits. They could also become concentrated among a relatively small number of suppliers and owners.

The distribution matters. A private saving does not, by itself, establish a public gain.

Count the costs before removing the alternatives

I am interested in agents because they can do useful work. That makes their economics worth examining carefully.

A credible comparison needs to include infrastructure, supervision, verification, failures, maintenance and switching costs. It also needs to account for the knowledge an organisation retains—or loses—as its operating model changes.

Some deployments will deliver substantial savings after all of that is counted. Others may prove to have moved costs elsewhere or deferred them.

We should be able to distinguish the two.

Before we dismantle our ability to do the work, we should understand its sustainable automated price—and what it would cost to change our minds.

Also published on Medium ↗.

This article reflects the evidence and development status at its original publication date.