How much does it cost to build an AI agent in 2026?

If you've asked around about the cost to build an AI agent, you've probably had quotes that don't seem to describe the same thing. One person says a few hundred a month. Another says six figures. Both can be honest, because "AI agent" covers everything from a subscription you switch on this afternoon to a system wired into five of your internal tools.

The bigger fear I hear isn't the build price. It's the monthly bill: the worry that you'll sign off on something and then watch a usage meter climb with no ceiling. So this post covers both. What the build costs, what running it costs, and which of the two deserves more of your attention.

Why quotes for the same "agent" vary so much

Four things drive most of the spread.

• Scope. "Answer customer emails" and "answer customer emails, issue refunds and update the CRM" are different projects. Every action the agent can take has to be built, tested and guarded.
• Integrations. Each system the agent reads from or writes to adds work. A well-documented API is cheap to connect. An old system, a shared spreadsheet or a portal with no API is not.
• Data readiness. If the information the agent needs is scattered, inconsistent or lives in someone's head, that has to be fixed first. This work is often bigger than the AI part.
• Risk tolerance. An internal tool that drafts summaries can be wrong now and then. An agent that talks to customers or touches money needs approval steps, logging and testing against messy real cases. That is where a lot of the budget goes, and it should.

What does it cost to build an AI agent? Three tiers

The figures below are ranges that vendors publish, not quotes, and they vary widely from one source to the next. Treat them as a rough map, not a price list.

Tier 1: an off-the-shelf tool or a no-code setup

If your problem is a common one (answering FAQs, booking appointments, routing enquiries), a ready-made product may be enough. One SMB pricing guide puts purpose-built conversational platforms at roughly $500 to $5,000 a month, and a contractor to wire up a no-code tool at $1,000 to $10,000.

This is the right answer more often than developers like to admit. If a subscription does most of what you need, start there and learn what the missing part really is.

Tier 2: a single-workflow custom build

This is one process, built properly. For example: reading incoming supplier documents, pulling out the key fields and putting them into your system, with a person checking anything unclear. SoftTeco publishes $10,000 to $30,000 for a prototype and $20,000 to $60,000 for an MVP, with simple agents at $20,000 to $80,000.

Tier 3: a multi-system agent

Here the agent works across several tools, makes decisions along the way and needs serious guardrails. SoftTeco's range for complex agents runs from $100,000 to $500,000 and beyond, and another guide on Indie Hackers lists retrieval-based (RAG) agents at $100,000 to $300,000.

Most small and medium businesses don't need tier 3 to start. Plenty need tier 2 for one process. Some need tier 1 and a clearer process written down.

Build cost vs running cost

The build is a one-off. Running costs are the ones that keep people up at night, and they come in four parts:

• Model fees. What you pay the AI provider, per token, every time the agent reads or writes text.
• Hosting. Servers, databases, queues and whatever else keeps it running.
• Monitoring. Logs, alerts and someone who actually looks at them.
• Maintenance. Fixing what breaks when an API changes, a model is updated or a user does something nobody tested.

The same SMB guide cites industry figures of 15% to 30% of the original build cost, every year, for maintaining custom software. For most builds, that line is bigger than the token bill.

A worked example: what the model actually costs

Let's do the arithmetic with real prices. Today, Anthropic's pricing page lists Claude Sonnet 5 at $2 per million input tokens and $10 per million output tokens, and Claude Haiku 4.5 at $1 and $5. The same page says a token is roughly three quarters of an English word.

Say your agent handles 2,000 documents a month. Each run sends about 6,000 tokens in (the instructions plus the document) and gets about 800 tokens back.

• Input: 2,000 × 6,000 = 12 million tokens, at $2 per million = $24
• Output: 2,000 × 800 = 1.6 million tokens, at $10 per million = $16
• Total: about $40 a month on Sonnet 5

Two features on the same page can bring that down:

• Prompt caching. If every request starts with the same long instructions, you can cache them. On Sonnet 5, reading from the cache costs $0.20 per million tokens instead of $2.
• Batch processing. Work that doesn't need an instant answer can go through the Batch API at 50% off both input and output.

Anthropic's own example on that page puts 10,000 support tickets on Haiku 4.5, at around 3,700 tokens per conversation, at roughly $37 in total.

One honest caveat: the page also notes that newer models use a tokeniser that produces about 30% more tokens for the same text, so estimate with some headroom. Even then, for most single-workflow agents the model bill is small next to what you paid to build it and what you'll pay to keep it working. SubTrada, the procurement platform I built for UK construction, runs its Claude features (drawing reviews, metadata extraction and quantity takeoff) with prompt caching switched on.

The costs nobody puts in the quote

Model fees are predictable once you've measured them. People's time isn't. A running-cost write-up on Indie Hackers from May 2026 put its infrastructure at $69 a month, then pointed out that the saving didn't include six hours spent chasing an out-of-memory problem. That's the usual pattern. The invoice looks small, and the hidden cost is someone's afternoon, or their weekend.

Gartner lists escalating costs among the reasons it expects over 40% of agentic AI projects to be cancelled by the end of 2027, alongside unclear business value and inadequate risk controls. On a small build, the thing that escalates is rarely the tokens. It's the maintenance nobody priced, or scope that keeps growing.

It's the same point I made in Most AI projects don't fail during the build. Budget for month three, not just launch day.

How to keep the monthly bill predictable

• Set a spending limit with your model provider, and an alert well below it.
• Log token use per run from day one, so you know the real cost per task instead of guessing.
• Use the smallest model that passes your tests. Anthropic's own advice on the pricing page is Haiku for simple tasks, Sonnet for most production work and Opus for the most complex reasoning.
• Agree the maintenance plan before the build. Who fixes it, how quickly, and what that costs.
• Start with one workflow. A narrow agent is cheaper to build, test and run, and it tells you whether the next one is worth doing.

How to narrow your own range

The quickest way from "somewhere between $500 and $500,000" to a real number is to spend a small amount finding out what you actually need. That means mapping the process, checking your data and systems, and deciding which tier you're in before anyone writes code. It's what my AI consulting audit is for, and it takes one to two weeks.

Sometimes the answer is "buy the subscription". That's a good outcome, and a cheap one.

If you want a second opinion on which tier you're in, send me what you're building.