How to tell whether an AI agent is paying for itself

Vendors quote savings. Your accounts show costs. A simple method for measuring what an agent really returns, including the hidden hours that most calculations leave out.

An AI agent pays for itself when the value it adds is greater than everything it costs to run, check and fix. That sounds obvious, but most early calculations only count half of each side. They count the vendor's bill and ignore the hours spent checking. Or they count the hours saved and ignore the work that moved to someone else. Getting this right decides whether you expand an agent, change it or switch it off.

The stakes are real. In June 2025 Gartner predicted that over 40 per cent of agentic AI projects would be cancelled by the end of 2027, naming escalating costs, unclear business value and inadequate risk controls as the reasons. Unclear business value usually means nobody measured the starting point. McKinsey's 2026 State of AI report, subtitled On the road to ROI, found that about two in ten respondents had scaled AI agents across their organisation. Proving value is what separates the projects that grow from the ones that stall.

Start before the agent starts. Pick the job and measure it as it is today for at least a few weeks: how many cases come in, how long each takes, who does them, how many errors occur and what each error costs to fix. Record the cost of the people involved using their full cost, not just salary. Without this baseline, any later number is a guess.

Next, list every cost of the agent. There is the vendor's charge, which may be per token, per outcome, per conversation, per credit or per seat; read the bill, not the price list. There is set-up time: describing the job, connecting systems, writing rules and testing. There is checking time: the minutes a person spends reviewing the agent's work, which is usually high at first and should fall. There is exception handling: the cases the agent passes back, which can be harder than average because the easy ones were taken. And there is management time: someone reviewing logs, updating instructions and dealing with complaints.

Then measure the value. The obvious value is hours saved, but only count hours that were actually redeployed or not hired. If the agent saves a team member two hours a day and those hours are spent on more valuable work, that counts. If they disappear into the week, the saving exists on paper only. Other value is often larger: faster response times that win sales, invoices processed in time to avoid late fees, fewer errors, work done outside office hours, or growth handled without adding staff.

Watch quality as closely as volume. An agent that handles more cases but gets more of them wrong can cost more than it saves. Klarna said in February 2024 that its AI assistant did the equivalent work of 700 full-time agents. In May 2025 Computing reported that the company was hiring human service staff again after its chief executive said quality had fallen. Track error rates, customer complaints and rework alongside the volume handled.

Look for ways to lower the running cost before deciding an agent does not pay. Model makers price their models very differently, and a cheaper model may be good enough for simple steps. Anthropic, OpenAI and Google all offered a 50 per cent discount for batch processing at the time of writing (November 2026), according to their pricing pages, for work that does not need an instant answer. Caching repeated instructions can also cut costs. Outcome-priced tools, such as Intercom's Fin, which was priced per outcome at the time of writing, move some of the risk to the vendor.

Finally, set a review date and a decision rule. Agree in advance what result would justify expanding the agent, what would lead to changes, and what would end the trial. Review monthly at first, then quarterly. Agents and their prices change often, so a decision that was right six months ago may not be right now.

What to do on Monday

1. Choose one agent or planned agent and record the baseline this week: cases per week, minutes per case, error rate and the full cost of the people involved.
2. Ask the people who check the agent's work to log their checking time for two weeks.
3. Pull the last three vendor bills and compare the actual cost per case with the quote.
4. Write down what result would justify expanding the agent and what would end it, and put a review date in the diary.

More on this topic: AI agent costs

Sources

• Gartner: Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027
• McKinsey: The state of AI in 2026: On the road to ROI
• Klarna: Klarna AI assistant handles two-thirds of customer service chats in its first month
• Computing: Meat back on the menu as Klarna rows back on AI
• Anthropic: Pricing
• OpenAI: Batch API
• Google: Gemini Developer API pricing
• Intercom: Intercom Pricing | Plans for every team size