The Hidden Costs of AI Agents Nobody Quotes You

The quote covers the build, the setup, and maybe the first month of API access. It does not cover what you pay from month two. That is where agent projects get expensive: the tokens, the data cleanup, the integrations that break, the logging you were told was included, and the person who still reviews what the agent cannot finish. Before you sign, you want a number for each of those, sitting next to the platform fee.

Gartner, as reported by Campus Technology in June 2026, expects half of generative AI projects to overrun their budgets by 2028. The overrun is not a surprise if you only priced the quote. The vendor comparison and the weekly owner report are where the product is described.

What follows is the cost math that belongs beside the proposal.

What does a vendor quote for an agent actually include?

Most quotes cover three things, and the three things are the easy part of the bill. A setup fee pays for configuring the logic, connecting the first tools, and writing the initial instructions. It is a one-time number, and it is usually the number that anchors the conversation, because it is the largest figure on the page and the only one with a clear end date. A monthly platform license pays for access to the software. It is fixed, it is predictable, and it is the line finance can put on a spreadsheet without asking a follow-up question. Sometimes there is an allowance for API calls. That allowance is the line that looks like a ceiling and behaves like a door.

Where the math breaks down.

The allowance, when one exists, is almost always sized for a pilot. A pilot might process 200 tasks a week. Production at the same company often runs 2,000 to 5,000. Cross the ceiling and the per-unit price turns on, and the monthly bill changes shape. The quote you are holding was built for the smaller number. The workflow you actually want is the larger one. Ask for the overage rate before you sign, in writing, at the volume you expect in month six, not the volume of the demo week.

The quote also skips the work to make your data usable, the cost of keeping integrations alive after the build, and the logging and review you will want once the agent is taking actions. Those lines add up faster than the platform fee. They are also the lines a vendor can leave off without lying about the three lines they did include. The setup, the license, and the allowance can all be accurate and still describe less than half of what you will pay.

Fill four lines before you compare quotes.

Before you compare quotes, fill four lines: platform fee per month, inference at production volume, data-prep hours times your labor rate, and maintenance per year. Add them and divide by 12. That is the monthly number the rest of this post is trying to estimate. If a vendor will not help you fill the inference line, that is information about the quote, not a reason to leave the line blank.

Why do token costs climb after the first month?

Every task sends text to a model and gets text back. The pricing unit is a token, roughly a word. You pay for the tokens in and the tokens out, on every step, whether the step produced something a customer saw or only something the next step consumed. A chatbot handles a short, predefined exchange. An agent reads a document, queries the CRM, drafts a reply, checks a flag, and writes a summary. Each of those is another pass through the model.

Gartner's March 2026 note on generative and agentic AI costs, "10 Best Practices for Optimizing Generative and Agentic AI Costs," the same reporting thread as the Campus Technology piece, found that agentic workflows consume 5 to 30 times more tokens per task than a simple chatbot exchange. At the same per-token rate, a task that costs $0.02 as a chatbot costs $0.10 to $0.60 when an agent runs it as several steps. The multiplier is the workflow, not a surprise rate change. You agreed to a per-token price that looked small against a chatbot, and then you asked the system to do five jobs.

Inference is the lifetime cost, not the training line.

The same Gartner reporting put inference at least 70% of a model's lifetime cost, well above the training spend buyers tend to stare at. You are not buying a model. You are running one, on every task, for as long as the workflow stays on. A pilot week hides this because the volume is small and the allowance covers it. Month two is when the same per-task cost meets the real volume, and month six is when a retry policy and a longer prompt have both quietly raised the tokens per task.

Retries multiply the bill without extra finished work.

When a call fails, a well-built agent retries. A timeout or a malformed response can become three calls where you intended one. On a busy workflow, retries alone add 20 to 40% to the inference bill without producing extra finished work. Ask whether retries count against the allowance, and ask what the agent does after the last retry. A retry that is free inside the allowance is a different contract from a retry that bills at the overage rate. Get that sentence into the order form.

How to estimate the month before you sign.

Take your expected weekly task volume. Multiply by the vendor's average tokens per task. Ask for that average from a pilot or from a deployment that looks like yours, not from a marketing example. Add 30% for retries and overhead. Multiply by 4.3 to land on a month. Compare the result with the allowance in the quote. If the estimate sits above the allowance, get the overage rate in writing before you sign. If the vendor cannot tell you the average tokens per task, you do not yet have an inference line. You have a guess, and the guess is the line that overruns.

BCG's 2026 analysis of governing agentic AI notes that a fully agentic workflow can require on the order of a million tokens a day. You may not be running that workflow. The figure is a ceiling to ask about, not a bill to assume. If your design calls the model once per customer message, your number will be smaller. If it calls the model once per document, once per CRM lookup, and once per draft, you are walking toward the larger end of Gartner's 5 to 30 times range.

What does data preparation cost before the agent can go live?

An agent is only as useful as the data it can read. If customer details live in three systems with three formats, if the knowledge base still describes last year's process, or if the product catalog has empty descriptions, the agent will produce weak output no matter how careful the build was. Preparation is the work of making that data usable: deduplication, one format, filled gaps, and articles that do not contradict each other. For most mid-market teams that work is internal, because the vendor has never seen your systems and should not be the one deciding which of two conflicting articles is current.

Agent-ready means a person could trust the fields.

An agent-ready CRM contact has a company, a current stage, a primary email, and a last-activity date that is actually current. A knowledge article is correct, current, and does not fight another article on the same topic. Complete fields, one format, and no contradiction is the whole definition. Most teams find, during this pass, that 20 to 40% of entries need some correction before the agent can use them. That range is the planning number. The invoice is the hours, not the percentage.

Estimate the hours before you believe the go-live date.

Budget preparation as hours times your internal labor rate. A clean CRM might be 20 hours. Three years of inconsistent entry across two migrations might be 200. The build does not start that clock for you. Put the hours on the plan as a gate: the agent does not go live on a knowledge base you have not finished correcting, because the first month of "the agent is wrong" is usually the data you skipped. A front-loaded cost still belongs in the monthly estimate. Divide the hours by 12 so a three-month cleanup does not disappear from the comparison against the platform fee.

One afternoon tells you whether the data is ready.

A quick test: if a new hire with read-only access could not summarize the last 30 customer interactions in an afternoon, the data is not ready, and you will pay the hours before the output is reliable. Run that test before you accept a go-live date. If the new hire has to open three systems, reconcile two spellings of the same account, and guess which note is current, the agent will do the same thing faster and with more confidence.

How much does integration maintenance add?

The agent does not sit alone. It connects to the CRM, email, the calendar, the accounting tool, or whatever the workflow touches. Every connection is an integration, and every integration breaks. The break comes when an API changes, when someone renames a field, when a platform moves an endpoint, or when a token expires. None of those events are on your launch calendar. All of them require someone to find the break and fix it before the agent can run again. While it is broken, the workflow is not "degraded." It is stopped, or it is guessing.

Integrations go stale inside a year.

BCG's 2026 analysis found that agent tools and frameworks become outdated in less than a year, which forces unplanned rework. That churn applies to the connections too. A webhook that worked in January may not work in September, when the platform vendor ships an update. Put that rework in the annual budget, not in the one-time setup. A setup fee that "includes integrations" includes the January version.

Ask who pays after the first 90 days.

Read the contract on this point and no other point first. Most agreements cover the initial connection. Maintenance after the first 90 days is often time and materials, billed hourly, or a flat retainer that was not in the quote you have. Either way, the cost is not on the page in front of you unless you ask for it. Ask which of those two shapes they sell, what the hourly rate is if it is time and materials, and what the retainer does not cover. A retainer that excludes "platform changes initiated by third parties" excludes the thing that actually breaks.

A year-one maintenance budget you can defend.

For three connections (CRM, email, calendar), budget 4 to 8 hours a month in year one. That assumes one small break per connection per quarter, about two hours to diagnose and fix. If your tools update often, or your team renames fields as a habit, use the high end. Add a fourth connection, such as accounting, and add the hours rather than hoping the three-connection number stretches. This is internal time or vendor time. It is not free because the setup fee felt large.

What oversight fees get left off the price sheet?

An agent that sends mail, updates contacts, routes leads, and opens tickets needs a way to see what it did and a way to stop a bad action. Governance is that pair of controls: the log of actions, and the stop. Deloitte's State of AI in the Enterprise 2026 surveyed 3,235 business and IT leaders across 24 countries and six industries between August and September 2025. Only 21% reported mature agent governance. The other 79% are running agents without the controls they will want, and most of them notice the gap when something has already gone wrong. The same report found that 25% have moved 40% or more of their pilots into production, and 75% plan to deploy agentic AI within two years. The governance work is arriving whether the quote mentioned it or not.

Logging, alerts, and a list a person actually reads.

Split the oversight stack into three pieces so the quote cannot blur them. Logging is a trail of every action the agent took. Alerting is a notification when the agent hits an error or a condition you did not expect. The review list is the set of actions the agent flagged as uncertain, waiting for a person. Basic logging is sometimes in the platform fee. Alerts and a human review list almost never are. You will build them or buy them. Many small teams park that list in the project tool they already pay for, which keeps the software line small. The hours are still real.

Most companies build the monitoring after the agent is live.

BCG found that only 15% of companies have the backbone in place to monitor agentic AI at scale. Everyone else builds it after deployment, which costs more and leaves a blind spot while the agent is still learning your environment and is most likely to do something you did not intend. Building the log and the alert before the first customer-facing week is cheaper than reconstructing a bad week from whatever the platform happened to store.

Budget the review hours, not a software line.

For one agent, budget one to three hours a week of human review in the first 90 days, then one to two hours a week as the error rate falls. Add the cost of any logging or alerting tool you have to buy separately. Do not undercount the hours because the tool was cheap. The hours are the real cost, and they belong on the same monthly total as the platform fee.

When does an agent create headcount cost instead of removing it?

The pitch is that the agent reduces headcount. Sometimes, later, it does. In year one it more often changes what the existing team does. MIT Sloan Management Review and BCG, "The Emerging Agentic Enterprise" (2026), found that 76% of organizations describe agentic AI as closer to a coworker than to a tool. In the same study, 35% are deploying it now, and 44% expect to deploy soon. The same report's uncomfortable pairing is that organizations are adopting quickly, well before the operating rules are in place. The coworker framing is the cost: the agent takes a set of tasks, and someone still oversees it, reviews the output, and handles what it escalates.

The work shifts before the headcount does.

If the agent handles 300 inquiries a week, someone still reviews the 30 to 50 it cannot resolve and updates the knowledge base when a product changes. Those are new tasks, not deleted tasks. They usually absorb 5 to 10 hours a week for a person who used to do the work the agent now handles. You did not remove the job. You changed the job, and the new job is exception handling plus upkeep. Budget the 5 to 10 hours as occupied time. If you also expected that person to take on a different revenue task with the "freed" hours, you have assigned the same hours twice.

Training is part of year one.

The team has to learn how to work next to the agent, how to read what it produced, and how to correct it when it is wrong. Budget 4 to 8 hours per person for the first orientation, plus coaching as the behavior changes. A launch email is not that training. The coaching shows up when the agent starts handling a case the orientation did not cover, which is the second week, not the second year.

Savings, when they arrive, arrive in year two.

Headcount savings, when they show up, usually show up in year two, after the agent has been tuned to your workflows and the review habit is real. Budget year one as a transition year with a full headcount, not as a savings year. If the business case only works because year one removes a salary, the business case is using a year you do not have yet. The one-time setup side is in what an agent costs to build. The staffing choice is in building in house or hiring it out.

How do you build the full estimate before you sign?

Take the quote and add five lines. That total is the estimate you can defend. Each line maps to a section above, so a blank line means a section you skipped.

1. Platform fee. Copy the monthly number from the quote. This is the only line already in your hands.
2. Inference. Weekly volume times the token cost per task, plus 30% for retries, times 4.3, minus whatever the allowance already covers. The remainder is the overage you should price before you sign.
3. Data preparation. Hours times your labor rate, divided by 12 so a front-loaded cleanup still shows up as a monthly equivalent. It is real even if it stops after month three.
4. Maintenance. Four to eight hours a month per integration, at your rate or the vendor's rate if you will pay them to keep the connections alive.
5. Oversight. The weekly review hours, the exception handling, and the knowledge-base updates, times your labor rate.

What the five-line total tells you.

Add the five lines. If the total is more than twice the platform fee, the lines that were missing from the quote are driving the budget. That does not mean the agent is a bad buy. It means the return has to be calculated on the real cost, not on the number that anchored the meeting. A return computed on the platform fee alone will look better than the business can deliver, and the disappointment will arrive in the same month the allowance runs out. Get your free Systems Plan. We will price the workflow you have now, before anyone recommends an agent.

An illustrative month, not a client result.

The month that follows is illustrative, not a client result. A business processing 1,000 tasks a week at $0.04 per task, plus 30% for retries, spends about $228 a month on inference. Add $60 for maintenance (4 hours at $15), $25 a month for data preparation spread across the year, and $120 for 8 hours of oversight. Total: $433 a month before the platform fee. Swap in your volume, your token price, and your labor rate. The structure is what matters: four lines the quote skipped can exceed the line the quote featured.

After you have the estimate, ask the vendor two questions. What is the average token count per task in deployments like yours? What is the most common maintenance job in year one? Both answers tell you something about the vendor as well as about the estimate. A vendor who can answer both has operated the thing they are selling. A vendor who can answer neither is selling the setup fee.

How we know this.

Gartner's March 2026 guidance on generative and agentic AI costs, reported by Campus Technology, is the source for the budget-overrun expectation, the inference share of lifetime cost, and the 5 to 30 times token multiplier. Deloitte's State of AI in the Enterprise 2026 (n=3,235, August to September 2025, 24 countries, six industries) is the source for the 21% governance figure, the 25% production figure, and the 75% deployment-plan figure. BCG's 2026 four-pillars analysis is the source for tool obsolescence inside a year, the 15% governance-backbone figure, and the million-tokens-a-day note. MIT Sloan Management Review and BCG, The Emerging Agentic Enterprise, is the source for the 76%, 35%, and 44% figures. The $228 and $433 example is round arithmetic, labeled illustrative, and not from a client. The maintenance hours and the oversight hours are ranges drawn from the cited research and from the mid-market deployments those sources describe, so you can replace them with your own rates without changing the shape of the estimate.