Insights · November 2026 · 7 min

How AI agents change the bottom line

A plain model for what an agent-run function costs, what it saves, and where the hidden costs sit.

Short answer: AI agents improve the bottom line by lowering the cost of running a function, raising throughput without new hires, cutting error and rework, shortening cycle times, and lifting revenue where speed or consistency affects sales. The sound budget is the fully loaded cost of the current team, minus oversight, integration and compute, minus the savings margin you require.

Most conversations about AI agents start with capability and end with a vague promise of efficiency. A finance leader needs the opposite order. The first question is what the function costs today, line by line, and the second is which of those lines an agent can actually move. I find it helps to treat an agent-run function the way you would treat any operating decision, with a baseline, a set of drivers, and a ceiling on what you are willing to spend.

This piece walks through those drivers and then works a simple illustrative model. The figures are round examples, picked so each step is easy to check. Your own numbers will differ, and the method matters more than any single figure.

Where do AI agents show up on the P&L?

An agent that runs a whole function touches more than one line of the income statement. It helps to separate five effects, because they behave differently and some of them are easy to overstate.

Labor cost

This is the obvious one. Salaries, benefits, payroll taxes, recruiting, training, management time and the software seats each person needs all add up to a fully loaded cost per head. When an agent takes over the routine volume of a function, the team that remains can be smaller, or it can stay the same size and move to harder work. Both are legitimate choices, and only the first shows up as a direct saving.

Throughput

A human team scales in steps. When volume grows, you hire, and each hire takes months to become productive. An agent scales with volume at a marginal cost that is mostly compute. For a growing company the value here is often larger than the labor saving, because it is the cost of hires you never have to make.

Error and rework

Every function has a hidden cost of fixing its own mistakes: a refund issued twice, a purchase order with the wrong quantity, a campaign that ran with an expired offer. These rarely appear as their own line. They are spread across overtime, write-offs and customer churn. An agent that follows policy the same way every time, and that checks its own work against rules, can shrink this cost. It can also create new kinds of error, which is why measurement matters.

Speed

Faster cycle times have a cash value. A support ticket answered in minutes keeps a customer who might have left. A reorder placed the day a trend appears avoids a stockout. Speed is harder to price than labor, so I usually model it conservatively or leave it out of the base case entirely.

Revenue lift

Where a function touches customers, an agent can raise revenue as well as lower cost. Marketing agents that test and reallocate spend continuously, or service agents that resolve problems before a renewal date, are two examples. I treat revenue lift as upside. A business case that only works because of projected revenue is a fragile case.

How do you calculate the maximum budget for an AI agent?

The useful number for a buyer is a ceiling: the most the company can spend on an agent-run function each year and still come out ahead of the team it has today. Here is a worked example with round figures.

Say a company runs a support function with 20 people. Assume a fully loaded cost of $85,000 per person once salary, benefits, taxes, tools, training and a share of management are included. The function costs about $1.7 million a year.

Suppose the plan is for an agent to handle the routine volume while four people stay on to manage escalations, review quality and own the relationship with the agent. Those four people cost about $340,000. The rest of the baseline, roughly $1.36 million, is the pool the agent has to beat.

From that pool, subtract the costs that come with any agent deployment, whoever builds it:

Line (illustrative, per year)Amount
Current function, 20 people fully loaded$1,700,000
Retained team, 4 people−$340,000
Integration and maintenance of connections−$120,000
Compute and model usage−$90,000
Required savings margin, 25% of baseline−$425,000
Ceiling for the agent platform$725,000

The savings margin is the line people skip. If the agent costs exactly what the team cost, the company has taken on new risk for no return. I suggest deciding the margin before talking to any vendor, because it anchors the conversation in what the business needs rather than in what a supplier would like to charge. A margin between 20 and 40 percent of the baseline is a reasonable starting range for most companies, and it should be higher in the first year, when the system is least proven.

If the throughput and rework effects are real, they widen the gap further. In the example, if volume is expected to grow by a quarter next year, the human plan would need about five more hires. The agent plan needs more compute. That avoided cost belongs in a second, separate scenario, so the base case stands on labor alone.

What does compute actually cost?

Compute is the cost of the model calls, tool calls and storage the agent uses to do its work. It scales with volume and with how much reasoning each task needs. A password reset takes a few short calls. A disputed invoice that requires reading a contract, checking three systems and drafting a careful reply can take many more.

The practical way to estimate it is to sample real work. Take a few hundred representative cases, run them through a prototype, and measure the cost per resolved case. Multiply by annual volume and add a buffer for retries and long cases. Model prices have tended to fall over time, though a careful budget should not depend on that trend continuing.

What hidden costs should a business plan for?

The costs that sink agent projects are rarely the ones on the vendor invoice. Three deserve their own lines.

There is also the cost of being wrong in public. An agent that mishandles a sensitive complaint can do more damage than a slow human. Guardrails, approval thresholds and a clear path to a person are part of the price of running a function this way, and I would count them as cost, not as a feature.

Why measure a function instead of a task?

The economics above only work at the level of a whole function. An assistant that drafts emails for a support team saves minutes per person, yet the headcount, the management layer and the rework stay the same. The bigger savings appear when the agent owns the outcome, from intake to resolution, and is measured on it. I make that argument at more length in why you should automate a function rather than a task.

Owning the outcome is also what allows the agent to improve. When the system can see which of its actions led to a resolved case, a retained customer or a clean month-end, it can learn to prefer those actions. That loop is described in reinforcement learning for business agents, and it is the reason I expect the cost curve of a well-run agent function to bend down over time, while a human team's cost tends to rise with wages.

How should a CFO judge the business case?

I would ask four things of any proposal. First, is the baseline fully loaded and agreed with finance, rather than taken from a salary band? Second, is there a stated savings margin, set before the vendor conversation? Third, are oversight, integration and compute shown as separate lines with their own owners? Fourth, is there a measured pilot that tests cost per outcome on real work before the full handover?

If the answers are yes, the case can be judged like any other investment. If the numbers depend on revenue lift or on cutting the oversight team to nothing, the case needs more work. When you are ready to write the brief, the steps are set out in the buyer's guide to rolling out AI agents.

Questions people ask

How much can AI agents save a business?

It depends on the fully loaded cost of the function, how much of its volume is routine, and how many people stay on for oversight. A sound model subtracts retained staff, integration, compute and a required savings margin from today's cost before setting any budget.

What is the maximum I should spend on an AI agent?

Take the fully loaded annual cost of the current function, subtract the retained team, integration, compute and the savings margin you require. What remains is the ceiling. Spending above it means the company takes on new risk for little or no return.

What are the hidden costs of AI agents?

The main ones are human oversight, integration with existing systems and its upkeep, a transition period where old and new run in parallel, and the guardrails needed to prevent costly public mistakes.

Do AI agents increase revenue or only cut costs?

They can do both. Agents in customer-facing functions can lift revenue through faster response and continuous testing, but a careful business case treats revenue lift as upside and rests the base case on cost.

How do you estimate compute costs for an AI agent?

Run a sample of real cases through a prototype, measure the cost per resolved case, multiply by annual volume and add a buffer for retries and complex cases.

Keep reading