What an AI Agent Actually Costs in 2026: Build Price, Monthly Run Cost, and Payback

The short answer: build $3k-$60k, run $90-$800 a month

An AI agent that drafts work inside one system costs roughly $3,000-$12,000 to build. One that takes actions across several systems, with approvals and an audit trail, costs $25,000-$60,000. Once live, most small-business agents cost $90-$800 a month to operate, and high-volume ones reach about $2,400.

Those ranges are wide because "AI agent" describes everything from a support-ticket drafter to a multi-step operations system. The rest of this guide breaks the number into the pieces you can actually control, so you can check any quote — including ours — against the math.

Build cost by scope: five agent types, five price bands

Build cost tracks two things above all: how many systems the agent integrates with, and whether it acts or only drafts. An agent that proposes an answer for a human to send is a fraction of the cost of one that writes to your CRM on its own, because the second needs permissions, validation, rollback and an exception queue.

If a proposal lands far below these bands for the same scope, check what is missing — usually evaluation, error handling or the integration work. If it lands far above, ask what is being built that you did not ask for. Either way, the AI automation service page describes how we scope these engagements.

Why two quotes for the same agent differ 10x

Two quotes differ by an order of magnitude when the vendors have made different assumptions about six variables. Put these in writing before you compare prices, and most of the gap disappears.

1. 1Integration count and quality. A clean REST API with good docs takes days; a legacy system with no API, rate limits or a CSV export takes weeks.
2. 2Act versus draft. Letting the agent write to a system of record adds permission design, validation, idempotency and rollback.
3. 3Required accuracy. "Usually right" and "right 98% of the time on this test set" are different projects. The second needs an evaluation suite.
4. 4Data condition. Messy, duplicated or unlabeled source data turns into cleanup work that someone has to pay for.
5. 5Human-in-the-loop design. A review screen that people will actually use is real product work, not a checkbox.
6. 6Security and compliance posture. Region pinning, PII redaction, retention settings and audit logs add effort, especially for EU customer data.

The line most quotes hide: what running an agent costs

Monthly run cost is the sum of four things: model tokens, orchestration and hosting, supporting services, and human review time. Most proposals quote only the first, if they quote anything at all, and it is often the smallest of the four.

Token math, worked end to end

Token cost per month equals tokens per run x runs per month x price per token, calculated separately for input and output because output tokens cost several times more. Here is the full calculation for a support-ticket triage agent.

Do this calculation before you sign anything. If a vendor cannot tell you roughly how many tokens a run consumes, they have not designed the agent yet.

Model choice swings run cost 7-13x

The same agent costs roughly seven to thirteen times more per run on a frontier model than on a flash-tier model. The practical answer is to route: use a cheap model for classification and extraction, and reserve the expensive one for the steps where quality measurably matters.

Cheaper is not automatically better. A flash model that misroutes 8% of tickets can cost more in human correction than the token saving. Choose per step, based on an evaluation set, not on the price table alone.

Cost centers nobody quotes

The hidden costs in an agent project are the ones that make it reliable: evaluation, retries, data cleanup, exception handling and prompt upkeep. Leaving them out is how a $6,000 quote becomes a $15,000 project.

• Evaluation suite. A labeled test set of 100-300 real examples and a script that scores every change. Without it, nobody can tell whether a prompt edit made things better or worse.
• Retries and fallbacks. Model APIs time out and rate-limit. Someone has to design what happens next.
• Data cleanup. Knowledge bases with contradictory articles produce contradictory answers. Fixing the source is often the highest-return work in the project.
• Exception handling. The 5-10% of cases the agent should not handle need a queue, an owner and a notification.
• Prompt and model maintenance. Vendors deprecate models. Each migration needs a re-run of the evaluation suite.

Maintenance: what 15-30% a year pays for

Plan on 15-30% of the original build cost per year to keep an agent working. The low end applies to a drafting agent on stable systems; the high end to action agents connected to several SaaS tools whose APIs keep changing.

That budget covers model migrations, integration updates when a vendor changes an API, prompt adjustments as your products and policies change, and periodic re-evaluation. An agent with no maintenance budget does not stay the same — it slowly gets worse as everything around it moves.

Three cost profiles

These are modeled profiles built from the ranges in this article, not specific client invoices. They show how build and run cost relate for three common first agents.

The pricing profile is the most expensive to build and cheapest per run — the difficulty is trustworthy data handling, not token volume. That is the pattern behind the pricing intelligence platform in our portfolio.

Fixed price, time and materials, or retainer

Insist on a fixed price for a defined first phase once scope is written down, use time and materials only for genuinely exploratory work with a not-to-exceed cap, and move to a retainer after launch for maintenance and improvements.

Agent projects tempt vendors toward open-ended time-and-materials billing because accuracy work feels unbounded. It is bounded if the acceptance criteria are: a named evaluation set and a target score. Put that in the contract and a fixed price becomes reasonable for both sides.

What $2k, $7k and $30k+ actually buy

Our public pricing maps onto agent work like this. Launch, from $1,900, suits validating an idea — a narrow prototype on one data source to see whether the approach works before funding a production build. Build, from $6,900, is a custom application that includes one production AI automation or integration. Scale is quoted to scope and covers advanced multi-system agents with ongoing engineering.

Most first production agents with one or two integrations land in Build territory; multi-system operations agents belong in Scale. The full tier breakdown is on the pricing page.

Payback math: will it pay for itself?

Payback in months equals build cost / (hours reclaimed per month x loaded hourly rate - monthly run cost). If the result is over 12 months, the scope is probably wrong — either automate a higher-volume task or build less.

Be honest about "hours reclaimed." Count only time that is redeployed to something valuable or that avoids a hire. Time that simply becomes slack does not show up in your bank account.

Seven red flags in an AI agent proposal

A proposal that cannot answer basic operating questions is a proposal for a demo, not a production system. Walk away from, or push back hard on, any of these:

1. 1No estimate of monthly run cost, or "it depends" with no formula.
2. 2No evaluation set or measurable acceptance criteria.
3. 3The agent can take irreversible actions with no approval step.
4. 4The vendor keeps the model API keys and the workflow in their own accounts.
5. 5No plan for model deprecation or maintenance.
6. 6Pricing based on "per agent" with no definition of what an agent is.
7. 7No mention of where your data is processed, retained, or whether it is used for training.

Five ways to cut agent cost 30-40% without cutting scope

The largest savings come from sending fewer, smaller requests to the right model — not from negotiating the build price. These five changes routinely cut a meaningful share of operating cost.

• Route by step. Use a flash-tier model for classification and extraction; reserve the stronger model for drafting or judgment.
• Use prompt caching. Major providers discount repeated prompt prefixes such as long system instructions. Structure prompts so the static part comes first.
• Batch what can wait. Nightly reports and backfills can use batch APIs, which several providers price well below real-time calls.
• Retrieve less, retrieve better. Five relevant passages beat twenty loosely related ones — cheaper and usually more accurate.
• Move review to sampling. Once the evaluation score is stable, review a sample instead of every output.

When an agent is the wrong answer

If the task happens fewer than a few dozen times a month, follows fixed rules, or already has an off-the-shelf tool that does it well, do not build an agent. A standard workflow automation or a seat license will be cheaper and more reliable. Build when the work uses your own data, spans your own systems, and happens often enough for the payback formula to work.