How Much Does a Custom AI Agent Cost? What Actually Drives the Quote

The direct answer: price follows scope, and scope has five dials

A custom AI agent costs whatever its scope requires, and scope is set by five dials you control: how many systems the agent acts in, how messy its source data is, how much it may do without human approval, how many channels it works across, and how much volume it will carry once live.

That is not a dodge. An agent differs from a chatbot in one important way: a chatbot answers questions, while an agent takes actions. It reads an email, decides what the sender needs, looks something up in your systems, updates a record, drafts a reply, and escalates when it is unsure. Each of those actions is a connection to real software that has to be built, secured and tested. Two projects both described as “an AI agent for customer service” can differ enormously in effort, depending on whether one drafts replies for a person to send and the other issues refunds on its own inside your billing system.

For that reason EVOTECH IT LLC doesn’t publish a price list for agents. We hold a free consultation, look at the actual workflow and systems, and then provide a fixed-scope written quote. The rest of this page explains what goes into that quote, so you can judge ours or anyone else’s.

Three separate bills: building it, running it, keeping it healthy

Most confusion about agent pricing comes from quotes that show only one of these three bills. Ask for all of them.

The build: paid once

Discovery and workflow mapping; connecting the agent to each system it reads from or writes to; designing its instructions, tools and guardrails; assembling an evaluation set of real past cases; testing; deployment; and training the staff who will supervise it. This is the part a fixed-scope quote covers.

The run: paid monthly, scales with use

Model providers charge by usage, measured in tokens (roughly, the amount of text the model reads and writes). An agent that handles a few dozen requests a day uses far less than one handling thousands. Alongside model fees come hosting for the agent service, storage for logs, a search index if the agent retrieves from documents, and telephony and speech processing if it takes calls. These costs are variable, which is why an honest estimate is written as a range tied to volume assumptions.

The upkeep: paid over time

Someone must review flagged cases, update the knowledge the agent relies on when policies change, adjust behavior as edge cases appear, and re-test whenever a connected system changes its API. Model providers also retire older model versions on published schedules, and each retirement means moving to a newer model and re-running the evaluation set. Upkeep can be handled by your staff, by us under a support agreement, or a mix of both.

The seven cost drivers, roughly in order of impact

1. Integrations, and how clean they are

Every system the agent touches is a separate piece of engineering. A modern platform with a documented API is straightforward to connect. An older system with no API may need screen automation, which is slower to build and breaks whenever the screen layout changes. Writing to a system costs more than reading from it, because each write needs validation, error handling and a way to undo mistakes.

2. Autonomy

An agent that only suggests is the cheapest to build safely. One that drafts work for a person to approve sits in the middle. One that acts alone needs hard permission limits, spending or value thresholds, detailed audit logging, rollback paths and far more testing, because nobody is checking each action before it lands.

3. Data preparation

If the knowledge the agent needs lives in tidy, current documents, preparation is light. If it is scattered across scanned PDFs, old email threads and three versions of the same policy, someone has to decide which version is true, clean it up and structure it for retrieval. That work is often underestimated, and it also improves every other system that uses the same information.

4. Evaluation depth

Before launch, the agent should be tested against a set of real past cases where the correct outcome is known. A narrow, low-risk task needs a modest set. Work involving money, personal data or regulated decisions needs a larger one, including deliberately tricky inputs meant to push the agent into mistakes.

5. Channels

Email and web chat are the simplest. SMS adds carrier rules and consent handling. Workplace chat tools add permission models. Voice is the most involved: speech recognition, response speed, interruptions, call transfers and telephony fees all enter the picture.

6. Security and access control

An agent used by one team with one set of permissions is simple. One serving several departments, each allowed to see different records, needs per-user permissions, single sign-on and careful handling of personal information in prompts and logs.

7. Volume and speed

High volume turns model choice into a cost decision, so the build may include routing routine steps to smaller, faster models, caching repeated lookups and batching work that isn’t urgent. Real-time conversation adds pressure to respond quickly, which narrows the model choices.

The same agent, scoped lean and scoped wide

Every move toward the right-hand column adds build effort, and most also add running cost. A sensible first project sits close to the left column and moves right only after the agent has earned trust on real work. That staged approach is the backbone of our AI agent implementation roadmap.

Worked example: pricing an invoice-intake agent two ways

Take a hypothetical wholesale distributor whose accounts payable clerk spends much of each day processing supplier invoices that arrive as email attachments.

Version A. The agent reads digital PDF invoices from one inbox, extracts supplier, purchase order number and line totals, looks up the matching purchase order in the accounting system through its API, and prepares a draft entry. Anything that doesn’t match within a set tolerance is flagged with the reason. The clerk approves or corrects each draft. One entity, one inbox, read-mostly access, human approval on every posting.

Version B. Everything in A, plus scanned and photographed invoices, automatic posting of matched invoices below a value threshold, emails to suppliers asking about discrepancies, credit-memo handling, and three legal entities with separate ledgers.

What changes between them, driver by driver:

• Data preparation grows because scanned images need text recognition and more tolerance for poor quality.
• Autonomy jumps twice: auto-posting needs thresholds, audit trails and reversal, and emailing suppliers means the agent now speaks for the company.
• Integrations triple in configuration and testing across the three entities.
• Evaluation needs far more cases, including credit memos, duplicates and deliberately malformed invoices.
• Running cost rises because each page processed consumes model usage and text-recognition capacity.

In a case like this we would usually quote Version A as phase one, with each Version B capability priced as a separate optional phase in the same written quote. The distributor gets value from the first phase, sees real accuracy on its own invoices, and decides on the rest with evidence rather than hope.

How EVOTECH structures a custom AI agent quote

1. Free consultation. A phone or video conversation about the workflow, the systems involved and what a good outcome looks like. No charge, no obligation.
2. Discovery. We map the process step by step, review sample data with sensitive details handled appropriately, and confirm what each system’s API allows.
3. Fixed-scope written quote. Line items typically cover discovery and design, each integration named individually, agent logic and guardrails, evaluation set and acceptance criteria, deployment, and staff training. Ongoing support appears as a separate, optional line.
4. Running-cost estimate. We estimate model and hosting usage against your expected volume and state the assumptions. Where possible the model provider bills an account in your company’s name, so you see exactly what the agent consumes.
5. Acceptance criteria agreed up front. The quote defines what the agent must handle correctly on the agreed test cases before it goes live, so “done” means the same thing to both sides.
6. Changes quoted before work starts. If you want something outside the scope, you get a written change quote to approve first.

We don’t guarantee delivery dates before discovery, because an unknown legacy system can change the plan; once scope is fixed, the quote lays out the phases in order. Learn more about how we build on our AI agent development page.

Costs that surprise buyers after launch

• Usage creep. Once staff trust the agent, they route more work to it. That’s good news, and it also raises model fees. Set usage alerts on day one.
• Model retirements. Providers retire older model versions. Switching is usually simple, but the evaluation set must be re-run, and occasionally instructions need adjusting.
• Upstream changes. A SaaS product the agent connects to updates its API or login method, and the connection needs maintenance.
• Review time. Approve-before-send designs still cost staff minutes per item. Budget for that time until the agent earns more autonomy.
• Knowledge drift. Prices, policies and procedures change. An agent reading last year’s return policy will confidently apply it.
• The edge-case backlog. Real traffic surfaces situations the test set missed. Expect a steady stream of small adjustments during the first months.

Ways to lower the price without cutting safety

• Start with one workflow and one channel. Prove value, then expand.
• Begin with draft-and-approve. It’s cheaper to build and gives you the review data you need to justify more autonomy later.
• Clean up source documents yourselves. Your staff know which policy version is current. Handing over a curated set saves discovery time.
• Favor systems with real APIs. If a workflow depends on a legacy tool with no API, consider whether a step can move to a system that has one.
• Right-size the model. Routine extraction and classification often run well on smaller, cheaper models; reserve larger ones for steps that need real reasoning.
• Supply your own history. Past emails, tickets or invoices with known correct outcomes make evaluation faster and more convincing.

What we won’t cut to save money: action logging, permission limits and a clear path for handing a case to a person. Those are what make an agent safe to leave running. If you are unsure where to begin, an AI consulting session can narrow the first project before any build is quoted.

Red flags in an AI agent quote

• No mention of running costs, or a claim that there aren’t any.
• No evaluation set and no acceptance criteria, only a demo.
• “Fully autonomous from day one” for anything that touches money or customers.
• The vendor owns the model account, the prompts and the logs, with no way for you to export them.
• An all-inclusive monthly price with no explanation of what happens when your volume doubles.
• A promised accuracy percentage before anyone has looked at your data.
• Integrations described vaguely as “connects to your systems” instead of named one by one.

A quote that avoids these problems will look longer and more detailed. That is a good sign: detail is what lets you compare vendors fairly and hold whoever you choose to what was promised.

Related services

Frequently asked questions