What an AI Agent Costs to Build in 2026, and to Run

$5,000 to $400,000. That is the real range for AI agent development cost in 2026, and it is not a vendor being vague. The same 4 words, build us an agent, describe a $6,000 email triage bot and a $300,000 claims processing system. Until the scope is pinned down, every quote you get is a guess about a different product.

We have shipped 315 products since 2017. This is the version of the answer we give founders on the first call.

The short answer, by tier

Published 2026 vendor pricing clusters into 4 bands. A single-task workflow agent that reads a document, classifies it and routes it runs $5,000 to $20,000. A customer-facing agent that answers from your own knowledge base with retrieval runs $20,000 to $60,000. A multi-step agent that has to call your CRM, your billing system and a human when it gets stuck runs $60,000 to $150,000. Multi-agent systems in regulated industries start near $150,000 and pass $400,000.

Healthcare and fintech carry a 25 to 40% premium on top of whichever band you land in. That is audit logging, consent handling and compliance review, and it does not go away because the model got cheaper.

Why 3 quotes come back 3 different sizes

The model is the cheapest part of an agent. In most of these projects, integration engineering and safety testing together account for 40 to 60% of the build. That is the number founders find hardest to believe, so here is where it goes.

Integrations. An agent that only talks is a demo. An agent that updates a record in your CRM, creates an invoice and books a slot in someone's calendar has to handle 3 APIs, 3 auth flows and 3 kinds of failure. Each integration is a week of engineering on a good day.

Data readiness. Retrieval is only as good as the documents behind it. If your help center has 4 versions of the refund policy, the agent will quote all 4. Cleaning that up is real work and it is almost never in the first quote.

Evaluation and safety. A production agent needs a test set of real questions with known good answers, run on every change. Without it, you find out about regressions from customers. Building that set and the harness around it is 1 to 2 weeks that cheap quotes skip.

Human handoff. The moment the agent is unsure, what happens? Who gets the ticket, with what context, and how does the customer know? This is UX work, and it decides whether the agent gets used or gets switched off in month 2.

Compliance. If the agent touches health data, payments or anything under GDPR, add audit logs, retention rules, consent and a review with someone who has done it before. This is where the 25 to 40% premium comes from.

The bill nobody puts in the proposal

The build cost is a one-time number. The running cost arrives every month for as long as the agent is live, and it depends almost entirely on which model you pick.

As of September 2026, published API prices per 1 million tokens look like this. GPT-5.6 Luna at $0.20 input and $1.20 output. Claude Sonnet 5 at $2 and $10. Gemini 3.1 Pro at $2 and $12. Claude Opus 5 at $5 and $25. Claude Fable 5.1 at $10 and $50. Prices fell roughly 80% between early 2025 and early 2026, and the $2 input tier is now where 3 of the major labs compete.

Take a realistic month: 10,000 conversations, each using about 6,000 input tokens once you include retrieved context and system prompts, and about 1,500 output tokens. That is 60 million tokens in and 15 million out. On GPT-5.6 Luna the bill is about $30. On Sonnet 5 or Gemini 3.1 Pro it is $270 to $300. On Opus 5 it is $675. On Fable 5.1 it is $1,350. Same product, same traffic, a 45x spread.

Then add hosting, a vector database, logging, monitoring and a person who reviews the edge cases. Across published 2026 cost breakdowns, infrastructure, usage and maintenance add 40 to 80% to the first-year total. A $40,000 build is a $56,000 to $72,000 first year.

None of this means pick the cheapest model. It means model choice is a product decision with a monthly invoice attached, and it should be made on purpose, with your real traffic numbers, before the build starts.

Why 4 in 10 of these projects get cancelled

In June 2025, Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. That came from a poll of 3,412 people. Two months later, MIT's NANDA initiative published a study of 300 public AI deployments and found that about 5% of pilots produced rapid revenue impact. The rest stalled.

Read both reports and the pattern is the same. Projects do not fail because the model was weak. They fail because nobody defined what done looks like, the integrations were harder than the demo suggested, and the humans who were supposed to use the output never trusted it.

That is the reason we stopped treating AI as a feature to bolt on after the product is designed. On our real estate GIS platform, the AI site scoring was designed alongside the map interface and the onboarding flow from week 1, and onboarding time dropped 55%. The scoring model was not the hard part. Making the score legible and trustworthy to a broker with 12 minutes to spare was the hard part. We wrote more about that shift in what agentic AI actually means in production.

What we charge, and where we say no

An AI integration sprint into a product you already run starts at $5,000. That covers one agent or workflow, wired into your systems, with evaluation and monitoring, in about 3 weeks. A full AI-native product build with the agent designed into the architecture runs $15,000 to $35,000, fixed price, 6 to 10 weeks, and you own everything at the end. The pricing page has the detail.

We also say no. If your data is not ready, we will tell you to fix that first because the agent will fail on it and you will blame the agent. If a well-designed form would solve the problem, we will say so, because a form costs $2,000 and never hallucinates. And if the request is for an agent that makes irreversible decisions with no human in the loop, we decline the work.