Why the ranges you find online don't help

Search for what it costs to build an AI agent and most of the first page is agency marketing. The ranges start at a few thousand dollars and end somewhere in six figures, usually with a plus sign. Almost none of them say which facts about your company move you from one end of the range to the other. You can't defend a line in a 2027 budget with one, though, and a vendor that quotes a figure before looking at your systems is guessing.

The cost of guessing wrong is well documented. In June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, and listed escalating costs first among the reasons, ahead of unclear business value and inadequate risk controls. The analyst quoted in the release warned that hype leads organizations to underestimate what running agents at scale really takes.

Budgets are already slipping. Menlo Ventures' 2026 State of Enterprise AI report, published November 3, found that reported enterprise AI spending this year averaged 4.7 times last year's level, against a planned 1.9 times. The median overshot by less (2.2 times against 1.8), and Menlo attributes much of the gap to agents that run in parallel and carry on after people log off. Read it with its context: the survey covers nearly 400 US enterprise decision-makers, and Menlo is an investor in Anthropic, which leads the report's model rankings.

Cost is also starting to limit use. In McKinsey's State of AI 2026, published in August, about 20% of respondents said AI operating costs, token costs included, had constrained their AI use. McKinsey calls that meaningful but not yet widespread, and most respondents still plan to spend more.

So our answer to "how much does an agent cost" is that nobody can tell you honestly until they know the things below about your project. Once they do, the estimate can be fairly precise.

What drives the build cost

These are the drivers we look at first.

Driver What keeps cost down What pushes cost up
Scope and autonomy One workflow; the agent drafts and a person approves Several workflows; the agent writes to systems of record on its own
Integrations A few systems with clean APIs or existing MCP servers Legacy systems with no usable API; custom MCP servers to build and secure
Data readiness Digital, current documents with a clear owner Scans, duplicates, conflicting versions, nobody accountable
Evaluation A domain expert who can label real cases No agreed definition of a correct answer
Security and compliance Internal, non-sensitive data Personal or regulated data; a full audit trail required
Change and handoff One team and a named owner Several departments, new roles, training

Scope and autonomy. An agent that drafts a reply for a person to approve is a different project from one that issues refunds on its own. Each step up in autonomy adds guardrails, approval paths, rollback logic and test cases, because a wrong action costs more than a wrong suggestion. For a first deployment we usually recommend the lower end: the agent prepares the work and a person signs off.

Integrations. In our experience this driver moves build cost more than any other. Every system the agent reads from or writes to (CRM, ERP, document store, ticketing) needs authentication, permission scoping, error handling and tests. Where a maintained MCP server already exists for a system, connecting it can be a short task. Where it doesn't, someone has to build one, and older systems without a usable API take the longest.

Data readiness. Agents answer from what they can retrieve. If the source documents are scanned PDFs, out of date or contradictory, part of the project becomes cleanup, and that work is easy to underestimate because it rarely looks like engineering. Ask who owns each source and how often it changes.

Evaluation. This is the line item cheap quotes leave out. Before an agent goes live, you need a set of real cases from your domain with known correct answers, and a test suite that scores the agent against them on every change. Building that set takes time from someone who knows the work, not only from engineers. On TaxFigure, every answer has to cite the primary authority it relies on, which gives testing something concrete to check. Most domains need a definition like that written down before an accuracy number means anything.

Security, compliance and logging. Personal data, regulated decisions and audit requirements add work: access controls, redaction, retention rules, and logs that record what the agent saw and did. TaxFigure logs each query, source and saved position so firms have an audit-ready trail. Logging of that kind is cheap to design in at the start and expensive to retrofit.

Change management and handoff. Someone on your side will own the agent after launch. They need documentation, a runbook and a way to see what the agent is doing. If the agent changes how a team works, add time for training and for the first weeks of questions. Projects without a named owner tend to stall once the build team leaves.

What it costs to run: a worked example

Here is the token line for a hypothetical support agent at list prices, to show which choices move it.

Assume each ticket takes six model calls: classify, extract fields, retrieve and plan, call tools, draft, check. Each call reads about 8,000 input tokens, of which 6,000 are stable (system prompt, tool definitions, policies), and writes about 500. Volume is 1,000 tickets per working day, about 22,000 a month. Prices come from the Claude pricing page as of this week.

  • Everything on Claude Sonnet 5.5 ($2 input and $10 output per million tokens), no caching: about $0.021 per call, $0.126 per ticket, roughly $2,770 a month.
  • The same, with the stable 6,000 tokens served from the prompt cache ($0.10 per million on cache hits): about $0.058 per ticket, roughly $1,270 a month, ignoring the occasional cache write.
  • The same, with classification and extraction moved to Claude Haiku 5.5 ($0.10 input, $0.50 output): about $0.039 per ticket, roughly $870 a month, provided Haiku passes your eval on those two steps.
  • Add a refinement loop in which the check step rejects a third of drafts and the agent reruns three Sonnet calls: about $0.049 per ticket, close to $1,080 a month.

Vendor prices also move on their own schedule. Google's Gemini API pricing page lists Gemini 4 Argon at $2 input and $10 output per million tokens through November 30, 2026, then $4 and $20 from December 1. At the same token counts, the uncached version of this workload would go from about $2,770 to about $5,540 a month overnight. (Tokenizers differ between vendors, so treat that comparison as approximate.)

The rest of the run cost doesn't fit on a price sheet: hosting for the orchestration layer and any vector store, log storage, monitoring, eval runs on every prompt or model change (both vendors' batch APIs cut that by 50%), and the time of whoever triages failures. At this volume, those lines can exceed the token bill. We cover routing and refinement loops in more depth in Cheaper tokens, pricier agents.

Maintenance: the model you tested will be retired

Treat model migrations as a recurring cost. Anthropic's model deprecations page shows Claude Sonnet 4.5, whose model ID is dated September 29, 2025, deprecated on September 30, 2026 and retiring on November 30, 2026. Anthropic commits to at least 60 days' notice before retiring a publicly released model. Google's pricing page says that from December 1, 2026, requests to gemini-2.5-pro will be routed to Gemini 3.8 Flash and billed at its rates, so an agent can keep answering on a model nobody on your team evaluated.

Each migration means re-running the eval set, adjusting prompts and sometimes reassigning which model handles which step. Architecture decides how much that hurts. On Promptify, the execution engine we built sits in front of several model providers, so adding a provider is a backend change rather than a product change.

How to read a quote

A serious proposal for a discovery phase or a fixed-price build, ours included, should spell out:

  • The workflow in scope, written as steps, with what the agent may do on its own and what needs approval.
  • Every system to integrate, how access works, and whether an MCP server exists or must be built.
  • The data sources, and who on your side prepares them.
  • How the eval set gets built, who labels it, how many cases it holds, and the accuracy threshold that counts as done.
  • Security and logging requirements, and where data is processed.
  • A run-cost estimate with its assumptions visible: volume, calls per task, tokens per call, model per step.
  • What happens after launch: the support period, the handoff documents, and who pays for model migrations.
  • What is excluded.

A fixed price is only as reliable as the scope behind it. When those items are vague, the price has either been padded to absorb the unknowns or it will be renegotiated a few weeks in.

Questions to ask a vendor

  1. What is your run-cost estimate per task, and which assumption would double it?
  2. Which of our systems already have MCP servers or usable APIs, and which need building?
  3. Who builds the eval set, and do we keep it?
  4. What happens when the model you chose is deprecated?
  5. What do we receive at handoff, and could our own team run the agent without you?
  6. Which parts of the price are fixed, and what triggers a change request?

How we scope this at Matious

Our AI Deployment program runs about six weeks in five stages. Week 1 is a system audit: we inventory your tools, map the processes and produce a prioritized deployment strategy. MCP server development follows in week 2, skills and agent development in weeks 2 and 3, and integration and testing against real data in weeks 3 to 5, covering accuracy, latency, edge cases and fallbacks. Production deployment and handoff happen in week 6. Your team receives documentation, a runbook for every agent and a monitoring dashboard, and we provide 30 days of support after launch. After that, you own it.

The audit is where the drivers above get answered for your project, from the number of systems to what the eval set must contain. That is also why we don't publish a price range: without those answers, ours would be no more useful than the ones at the top of this page.

List prices are the numbers most likely to have changed by the time you sign, so re-run the arithmetic with the vendor's page open. If you are preparing a 2027 budget and want a scoped estimate for a specific workflow, tell us about it.

Sources