How much does it cost to build an AI agent

The short answer: a single-workflow agent that reads from a few systems, writes to one, and has a human approval step usually costs $40,000 to $120,000 to build properly. A multi-system agent with custom MCP servers, granular permissions and a full evaluation suite usually runs $120,000 to $250,000. Anything quoted far below that range is a prototype. That can be the right thing to buy, as long as nobody calls it production.

The model call is the cheapest part of the build. The money goes into the tool layer, the permission model, the failure handling and the evaluation work that makes the agent safe to leave running.

The five drivers of agent build cost

Number and quality of integrations. Each system the agent touches needs a typed tool, auth, rate-limit handling and tests. A clean REST API with good docs might take two to four days. A legacy ERP with SOAP endpoints and no sandbox can take three weeks on its own.

Write access. Read-only agents are cheap. The moment an agent can send an email, issue a refund or update a record, you need idempotency keys, approval gates, rollback paths and an audit trail. Write access roughly doubles the engineering on any given tool.

Ambiguity of the task. Classifying a support ticket is narrow. Negotiating a schedule across three calendars and a CRM is not. The more judgement the task needs, the more scenarios the eval suite has to cover and the more iterations it takes to reach an acceptable success rate.

Compliance and data handling. PII redaction, data residency, zero-retention provider agreements, SOC 2 evidence and access reviews all add time. In regulated industries this can be a quarter of the budget.

Interface. An agent that runs in the background and posts to Slack is cheaper than one that needs its own web app, review queue and admin console.

Typical ranges by agent type

Internal assistant over company knowledge, read only, with retrieval and citations: $40,000 to $80,000.

Single back-office workflow, such as invoice intake or lead routing, reading from two or three systems and writing to one with human approval: $60,000 to $120,000.

Customer-facing agent with tool access, guardrails, escalation to humans and production monitoring: $100,000 to $200,000.

Multi-agent or multi-system platform with custom MCP servers, role-based permissions and an eval suite that runs in CI: $150,000 to $250,000 and up.

What it costs to run

Inference cost depends on task length and model choice. A short classification step on a small model costs a fraction of a cent. A long multi-step task on a frontier model with large context can cost tens of cents to a few dollars per run. Multiply by volume before you approve the design, not after.

Prompt caching, routing simple steps to cheaper models and capping tokens per task usually cut run cost substantially. Budget separately for observability tooling, hosting, and a few engineering days a month for model upgrades, eval maintenance and provider changes.

Where agent budgets get burned

Starting with autonomy. Teams that let the agent act unsupervised from day one spend months cleaning up. Shadow mode, then approval mode, then autonomy on the actions that earn it is faster overall.

Skipping evals. Without a scenario suite, every prompt change is tested by hand on a few examples. The project slows to a crawl around week eight because nobody can tell whether a fix broke something else.

Heavy frameworks. Some agent frameworks save a week up front and cost a month later when you need behaviour they don't support. Keep orchestration thin and in your own code.

How to get an accurate quote

Write down the job the agent does, every system it touches, which actions need human approval, and the volume you expect. Any serious AI development company can turn that into a fixed-scope estimate within a week. If a vendor quotes without asking about write access, approval gates or evals, the quote is for a demo.