Where AI agents earn their keep

AI agents can save hours of work, but a single expensive mistake can wipe out those savings. The business case depends as much on the cost of getting things wrong as it does on the work they get right.

Imagine two AI agents, both right 98% of the time. One labels customer inquiries. The other changes prices on your website. You might happily renew the first subscription while spending Sunday explaining the second to your CFO.

The accuracy is identical. The remaining 2% has very different consequences.

An agent has a much easier path to paying for itself when mistakes are cheap to catch and fix. Yet a business case can calculate the value of every minute saved while leaving the cost of getting things wrong completely blank.

There is already plenty of cleanup happening. In Nylas’s September 2026 survey, 54% of respondents reported at least one agent-caused error at their organization in the previous 90 days. The most common problem was a workflow that somebody had to reverse or correct.¹

That statistic cannot tell you whether those agents were worth having. For that, you need the repair bill.

The missing line in the spreadsheet

Take a hypothetical business handling 10,000 requests a month. Its current process costs $5 per request, including the cost of its existing mistakes. Monthly bill: $50,000.

An agent brings the operating cost down to $2 per request. That includes software, routine human review and an allowance for setup and maintenance. Before paying for its mistakes, the business spends $20,000.

So far, everyone likes the presentation.

Now suppose 2% of requests still go wrong after routine review. That is 200 failures. If each costs an average of $10 to investigate and put right, that leaves a potential monthly benefit of $28,000. Lovely.

If each failure costs $150 on average, the entire saving disappears.²

Nothing changed about the agent’s accuracy. We changed the price of being wrong.

That price includes more than the person fixing the record. A customer might receive compensation. A shipment might need replacing. A missed deadline might lose an order. The error can also sit quietly for weeks, collecting interest in the form of further bad decisions.

And saved time needs somewhere useful to go. If the same people remain on the payroll, the immediate gain is capacity. It becomes a financial return when they handle more business, avoid overtime or spare you another hire. A spreadsheet cannot spend a saved hour.

Give the agent a sensible job

If a fixed rule already handles the job reliably, keep it. Agents earn their extra complexity when they must interpret messy information or work through changing situations. Replacing a working rule with a model creates another bill without necessarily adding any value.

There is plenty of work with manageable consequences inside businesses where the final decisions are expensive.

In October 2026, Barclays and Anthropic described a system that helps classify and route roughly 120,000 emails a day in the bank’s Global Markets business. It prepares information for operations colleagues to act on, within a deployment that includes human oversight.³

The announcement gives no error rate or net savings. Still, the division of work is useful. The intended benefit is straightforward: less preparation for the people making decisions.

You can apply that idea to smaller businesses. An agent could gather the documents needed for a refund and prepare a recommendation. A person approves the payout. It could investigate an overdue invoice and draft the follow-up, with somebody checking before a valued customer receives an enthusiastic threat.

The attraction is that much of the tedious work gets done before the business makes a commitment. Reviewing a prepared case can be worthwhile if it takes substantially less effort than assembling it.

Check that assumption carefully. In an August 2026 analysis, McKinsey put human oversight at 70% to 75% of variable running costs for banking customer-service agents.⁴ That figure is specific to those workflows, but it shows how much of the running cost can sit outside the model bill.

Make the mistakes smaller

Product decisions can change this arithmetic without waiting for a better model.

In April 2026, hosting provider Railway described an agent deleting a customer’s production database using credentials with broad access. Railway recovered the data, then extended a 48-hour undo window to deletions through the software interface the agent had used.⁵

That is a useful change to the economics of a mistake. Recovery becomes an option built into the action. It still requires somebody to notice in time.

Similar choices are available in ordinary business workflows. Let an agent propose a price change before publishing it. Limit the value of refunds it can approve. Start with a small batch of records and check the results before releasing the rest.

A refund ceiling alone is insufficient if the agent can issue thousands of refunds. Put a limit on the total as well. A mistake repeated at machine speed can turn a modest allowance into a very generous promotion.

For a pilot, keep a record of completed work alongside the minutes people spend checking and repairing it. Include customer complaints and cases that eventually return to a human. Compare the total with the process you actually run today, complete with its delays and errors.

Agree what counts as finished before the pilot starts. A closed ticket can hide an unresolved problem; a completed form can contain unusable information. Check a sample with the people receiving the work. Otherwise, the agent can look wonderfully productive while handing its unfinished jobs to somebody else.

Test a bad day, too. What happens if a supplier sends an unfamiliar invoice format? Or a policy changes and the agent keeps using the old one? A quiet week tells you little about a mistake that could affect an entire batch.

The best starting jobs have enough volume to matter and errors you can spot before they become expensive. You should be able to describe the likely failure, name the person who will catch it and estimate the cost of putting it right.

Once those numbers work, give the agent more responsibility in measured steps. There is no prize for granting it every permission on day one.

Before approving the next agent proposal, add a line for mistakes. If nobody can fill it in, the return is still a guess.

---