What an AI agent really costs: the six budget lines a quote leaves out

What an AI Agent Really Costs: Six Lines Missing From Quotes

A quote for an AI agent is easy to read: a number, a timeline, a list of features. The bill a year later is a different document, and the gap between the two is not the vendor cheating or the scope creeping. It is six budget lines that were never in the first quote because nobody had seen them in an ordinary IT project. This article names them, prices each one against our own public pilot terms and a live deployment, and gives you the one question that exposes a quote done on paper.

A disclosure first: Grow2.ai builds these agents, so the reference numbers below are ours, from the pricing page and from a published e-commerce case. Where we do not have a number, we say so rather than borrow one.

Why does the first quote for an AI agent miss?

The first quote misses because the buyer prices what they have already seen in previous IT projects, a licence, integration hours, a project fee, and the vendor prices what is easiest to demo, the model answering questions. Neither side has a line for what happens after the first week of live traffic: the second system that needs wiring, the peak-day token bill, the log nobody reads, the price list that changed on Tuesday, the model version retired in month seven, the regulator's question about disclosure. Gartner warned in May 2024 that enterprise AI cost estimates can be off by 500–1,000%, and in June 2025 predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, mostly for unclear value and escalating costs. Both numbers describe the same failure: a quote that stops at the model. The fix is not a bigger buffer but a quote that names all six lines.

Which six budget lines does a quote leave out?

Six lines separate a paper estimate from a production bill, and the useful thing about them is that each can be priced before kickoff if someone asks. Integration is the first: the agent has to read and write your CRM, your helpdesk, your telephony and your documents, and every additional system is its own piece of work. Model usage in production is the second: the pilot month is not the rate, because peaks, retries and a supervisor model that checks every answer all add calls. Monitoring is the third: logs, an evaluation set, drift checks, and a person who reads them. Knowledge-base upkeep is the fourth: the agent answers from your price list and policies, and someone has to keep those current. Maintenance when the model is retired is the fifth: providers publish deprecation schedules, and every retirement means a migration and a re-run of the evaluations. Governance is the sixth: disclosure under the EU AI Act, GDPR, data retention, a human hand-off.

Integration. One CRM and one channel are in a Grow2.ai pilot; each extra system (ERP, telephony, a second messenger) is a scoped item, not a surprise. Ask the vendor to list the systems by name.
Model usage in production. Budget by volume: dialogues per month times the cost per dialogue. In our e-commerce deployment, 6,400 dialogues cost roughly $0.11 in LLM usage each, about €0.10, supervisor included. Voice is priced per minute and costs several times more than text.
Monitoring. In our monthly fee: logs, evaluation set, supervisor review, alerts. If you build it yourself, it is a separate tool plus the hours of whoever reads it.
Knowledge-base upkeep. Not a licence but a process owner's hours whenever the catalogue, prices or policies change. Ask who updates the sources and how fast the agent picks the change up.
Maintenance when the model is retired. OpenAI and Anthropic publish retirement schedules; every agent built on a model has a date. In our plans the migration is covered by the monthly fee. Ask who does it in a custom build, and at what price.
Governance. The agent must say it is an AI (EU AI Act, Article 50), data must live in an EU region with personal data redacted from logs and a DPA in place, and hard cases must reach a person. In regulated industries add legal review to the budget.

Where does the money actually go?

The money does not go where the negotiation goes. Buyers argue about the token price because it is the one number printed on the provider's website. By our own public terms it is not the largest line — and, as the numbers below show, not always the smallest. Take Grow2.ai's public numbers: on the Practice plan at €99 a month, five hundred dialogues cost about €50 in tokens, a third of the monthly bill; on Operator at €149, two thousand dialogues cost about €200 — more than the subscription itself, about 57% of the bill. The rest of the real cost is not on any invoice: the hours your process owner spends approving replies in shadow mode during the first days, the hours spent keeping the knowledge base current, and the integration work whenever a new system joins. That is why a quote that lists only "model usage" and a project fee is incomplete by design, and why the cheapest way to lower the total cost of ownership is to keep the process narrow, one channel and one CRM, until the KPI has moved.

Budget line	No-code platform	Agent studio (Grow2.ai)	Custom build with an agency
Integration	You, inside the platform's connectors	One CRM + one channel in the pilot; extra systems scoped	Project hours per system
Model usage	Platform credits	≈ €0.10 per dialogue on top of the plan	Your API account, peaks included
Monitoring	Platform dashboard	In the monthly fee	A separate tool plus someone reading it
Knowledge base	You	You, with the agent's update path defined	You
Model retirement	The vendor	Covered by the monthly fee	You or a retainer
Governance	Your responsibility	Disclosure, EU deployment, DPA included	Your responsibility

How do you check a quote in five minutes?

One question does most of the work: "What exactly is not included in this price?" A vendor who has run an agent in production answers with a list, and the list looks like the six lines above with numbers next to them; a vendor who has not answers with "it depends" or a silence. Then send them the checklist below and read the answers for names and numbers, not adjectives. Which systems, by name, are in the integration? What is the cost per dialogue, and did it include the supervisor model? Where do the logs go and who reads them? Who updates the knowledge base and how fast does the agent see the change? Who migrates the agent when the model is retired, and at what fee? How does the agent disclose that it is an AI, and where do the data live? Our own answers to the same six questions are on the pricing page, which is the point: the answers should exist before the contract, not after the first month.

Systems, by name. Integration scope is a list of systems, not a verb.
Cost per dialogue. A number, with the supervisor and retries inside it.
Logs and evals. Where they live, who reads them, how often.
Knowledge-base path. Who edits the sources, how the agent picks changes up.
Model retirement. Who migrates, at what price, within what time.
Disclosure and data. AI Act disclosure, EU region, DPA, retention period.

What does a three-year budget look like for a small business?

A three-year budget for one agent on one channel, built from Grow2.ai's public terms, is short enough to fit in a sentence: on the Starter plan, a €1,800 pilot, then thirty-six months at €49 plus about €50 of tokens for five hundred dialogues a month, roughly €5,400 in total for the agent itself. Add what is yours and not ours: a process owner's hours in the first weeks, when every reply is approved in shadow mode, and a smaller recurring slice whenever prices or policies change; the exact hours depend on your team, so put your own number there. Outside the budget by design: a second channel or a second system, voice minutes, and a larger volume, each of which is a scoped item with a price before it starts. A custom build from scratch works differently: tens of thousands of euros in the first year for a mid-market scope, then maintenance every year, with the retirement migration on you unless a retainer covers it. The full comparison of the three routes is in AI agents vs custom development.

For the arithmetic on your own volumes, the pricing page calculator runs leads, response gaps, conversion and deal value; the reasoning behind it is in How to calculate AI agent ROI. If the question is who should build the agent at all, start with how to choose an AI agent development company.

Sources

Frequently asked questions

Why does the real cost of an AI agent exceed the quote?

Because a first quote prices what the buyer has seen in ordinary IT projects, a licence and integration hours, and skips six lines that appear after the first week of live traffic: extra integrations, production model usage, monitoring, knowledge-base upkeep, maintenance when the model is retired, and governance. Gartner has warned that enterprise AI cost estimates can be off by 500–1,000%.

Is the model fee the most expensive part of an AI agent?

Not always. In Grow2.ai's public plans, five hundred dialogues a month cost about €50 in tokens against the Practice plan at €99 a month (pilot €3,600, 21 days), a third of the bill; at two thousand dialogues tokens are more than half. The larger cost is off the invoice: the process owner's hours in shadow mode and on the knowledge base, and integration work for each new system.

How much do tokens cost per dialogue?

In our published e-commerce deployment, 6,400 dialogues on Instagram, Viber and Telegram cost roughly $0.11 in LLM usage per dialogue, about €0.10, with the supervisor model included. Voice is priced per minute and costs several times more than text. Budget by volume: dialogues per month times the cost per dialogue.

What happens to the agent when the model is retired?

OpenAI and Anthropic publish deprecation schedules, so every agent built on a model has a date on it. When it comes, the agent is migrated to a newer model and the evaluation set is re-run to confirm nothing regressed. In Grow2.ai's plans that migration is covered by the monthly fee; in a custom build, ask who does it and at what price.

How do you check a quote without an in-house AI team?

Ask what exactly is not included in the price, then send the six-line checklist: systems by name, cost per dialogue with the supervisor included, where logs and evals live, who updates the knowledge base, who migrates the agent when the model is retired, and how the agent discloses that it is an AI. Answers should contain names and numbers, not adjectives.

What does a three-year budget look like for one agent?

On Grow2.ai's public terms: on the Starter plan, a €1,800 pilot, then thirty-six months at €49 plus about €50 of tokens for five hundred dialogues a month, roughly €5,400 for the agent over three years. Your own hours for approving replies in shadow mode and keeping the knowledge base current come on top, and a second channel, voice minutes or a larger volume are scoped separately.