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

An AI agent has two prices, and most quotes show one. The build is engineering time, paid once. The running cost is a monthly model bill set by which model you choose and how often the agent runs.

Both are worked out below from published figures.

The build is mostly engineering time. Salesforce staffs each agent deployment with one strategist and two engineers for about three months, covering one or two use cases.
At published subscription rates for senior AI engineering capacity, $3,500–$12,000 a month in August 2026, three months of one engineer is $10,500–$36,000. That is our arithmetic on published rates.
The same three months of a US hire is about $81,000, at a fully loaded first year near $325,000, after a median 71 days to fill a senior role.
Running cost is a separate monthly line. At Anthropic's list prices, the same agent task costs 40 times more on the largest current model than on the smallest.
A quote without an eval set, approval gates and logging is lower because it prices a demo.

The direct answer: building a production AI agent in 2026 is mostly the cost of engineering time. Published figures put three months of one senior engineer at $10,500–$36,000 on subscription capacity, or about $81,000 as a fully loaded US hire.

Three months is the anchor because that is how long Salesforce's own deployment pods spend per agent: one deployment strategist and two forward-deployed engineers, full time on one client, for one or two use cases.

Running the agent is a second, monthly cost: tokens per task multiplied by how often it runs. That line depends far more on model choice than on anything in the build, and the worked examples below show by how much.

Most of the spread between quotes is what they include. An eval set, human approval on irreversible actions, and decision logging separate an agent you can operate from a demo, and they are the first things a low quote drops.

Our AI agent development is priced as a fixed fee against a written scope, with all three included.

Three ways to pay for the build

	In-house hire	Subscription capacity	Fixed-scope build
Published anchor	About $325,000 fully loaded in year one on a $220,000 base	$3,500–$12,000 a month, most $3,500–$8,000 (August 2026)	Priced per written scope
Before work starts	A median 71 days to fill a senior role, then ramp	Onboarding, then you direct the work	A signed scope and a fixed fee
Who carries overrun	You	You, month by month	The firm, inside the written scope
Who directs the work	You	You	The firm, against the scope
What you hold at the end	An employee and what they know	The code; the context leaves with the engineer	Code, prompts, eval set and runbook, assigned to you

The cheapest line depends on how long the need lasts. Permanent AI work favours a hire; one agent with a deadline favours a fixed scope. The hire-versus-engage model runs both with your numbers.

What the build pays for

Five pieces of work sit inside any production agent build. Quotes differ mostly in how many of them they include.

Scope and the failure map. Which task, which data, which systems, and what the agent must never do, written before any code.
Integrations. Read and write access to the CRM, ledger, inbox or database the agent acts on, with permissions that match your own. Count the systems: each one brings its own authentication, rate limits and edge cases.
The eval set. Examples from your own data that define good output and failure, run on every change. Without it, nobody can say whether the agent improved or regressed.
Approval gates. A person signs before anything irreversible: money moved, a customer contacted, a record deleted.
Logging and rollback. Every decision stored with its inputs, and a way to switch the agent off without breaking the workflow around it.

What it costs to run, worked from list prices

Running cost is tokens in and tokens out, multiplied by volume. Anthropic's published API prices in October 2026, per million tokens: Claude Haiku 5. 5 at $0. 10 input and $0. 50 output, Claude Sonnet 5. 5 at $2 and $10, Claude Opus 5.

5 at $4 and $20. Web search adds $10 per 1,000 searches.

Take an agent task that makes five model calls, each reading about 6,000 tokens and writing about 600: 30,000 tokens in and 3,000 out per task. At 500 tasks a working day, about 11,000 a month, our arithmetic at list price is:

Haiku 5.5: about $0.0045 a task, or about $50 a month.
Sonnet 5.5: about $0.09 a task, or about $990 a month.
Opus 5.5: about $0.18 a task, or about $1,980 a month.
Two web searches per task add about $220 a month at that volume, before the tokens the results bring back.

Where the running bill is won or lost

Model choice is the largest lever: at list price the same task costs 40 times more on Opus 5. 5 than on Haiku 5. 5.

The engineering answer is to send each step to the smallest model that passes its evals, which is only possible when the evals exist.

Two published discounts compound with that. Anthropic's Batch API takes 50% off input and output for work that can wait, and cached prompt reads on Sonnet 5.5 and Opus 5.5 are billed at 0.05 times the base input price.

Retries are the hidden multiplier. An agent that retries a failed call three times costs three times as much on its bad days; a spend ceiling and a timeout on each call path keep a bad day bounded.

What a credible quote includes

Ask every firm for the same breakdown, in writing. A quote that cannot itemise these is pricing a demo.

The task, the data and the systems, named.
The eval set: who writes it, how many examples, and the pass bar.
Which actions need a human approval, and how the agent asks for it.
Expected running cost per month at your volume, by model.
What you own at the end: code, prompts, eval set and runbook.
How the price moves if the scope does.

Common questions

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

Mostly engineering time. Salesforce's deployment pods spend about three months per agent with one strategist and two engineers.

At published subscription rates for senior AI engineering capacity, $3,500–$12,000 a month in August 2026, three months of one engineer is $10,500–$36,000; as a fully loaded US hire it is about $81,000.

The number of systems, the approval gates and the eval work decide where a given agent lands.

How much does an AI agent cost per month to run?

Tokens multiplied by volume, plus any web search. At Anthropic's October 2026 list prices, a task of 30,000 input and 3,000 output tokens costs about $0. 0045 on Haiku 5. 5, $0. 09 on Sonnet 5. 5 and $0. 18 on Opus 5. 5.

At 11,000 tasks a month that is roughly $50, $990 or $1,980. Hosting, monitoring and maintenance sit on top.

Why do AI agent quotes vary so much?

Because they price different products. A low quote is usually a prompt wired to an API, with no eval set, no approval gates and no logging. A higher one includes the work that lets you run the agent in production.

Compare quotes line by line against the list above.

Is it cheaper to hire an in-house AI engineer?

Over a long enough horizon, often yes. A $220,000-base US hire costs about $325,000 in the first year once employer tax, benefits and recruiting are counted, and a senior role takes a median 71 days to fill.

For one agent with a deadline, a fixed scope costs less; for permanent AI work, a hire wins. The hire-versus-engage model runs it with your numbers.

How long does it take to build an AI agent?

Salesforce's own pods plan for about three months per agent, covering one or two use cases.

A narrow agent over one system with a clear eval set can be faster; one that writes to several systems of record, or needs approval gates on money, takes longer.

What should an AI agent development quote include?

The task, data and systems named; the eval set and its pass bar; which actions need human approval; expected running cost per month at your volume, by model; what you own at the end; and how the price moves if the scope does.

Who owns the code and the IP?

You do, outright. IP assignment is signed before repository access, by everyone who touches the code, and it covers AI-assisted output. Handover includes the runbook and the rollback path. Nothing depends on a licence we control.

What are the exit terms?

Retainers end on 30 days written notice from either side, with no minimum term. Milestone billing means you never pay ahead of delivered work.

Want the number for your agent?

Bring the task, the systems it touches and a rough monthly volume. That is enough to scope the build and estimate the running cost.

Book a call →