AI agent development costs $350 to $20,700 for a single agent built by a development shop, depending on how much planning and coordination the agent has to do, and $10,000 to $80,000 or more for fully custom builds with integrations and ongoing hosting. A managed agent service costs far less up front: VantaSoft's plans start at $200 a month with setup scoped during discovery. The gap between those numbers is not margin. It is what you are buying: a one-time build you then run yourself, or a running service someone else operates.
The ranges above are published by the vendors themselves and linked in the table. Prices move, so confirm before you budget.
AI agent development cost compared
| Option | Published cost | What it covers | Source |
|---|---|---|---|
| Simple reflex agent, dev shop | $350 to $3,500 | Rule or model-based responses, light memory | DevCom 2026 guide |
| Goal-based agent, dev shop | $6,000 to $9,500 | Step planning and tool routing | DevCom 2026 guide |
| Hierarchical or multi-agent system, dev shop | $11,200 to $20,700 | Planners, executors, handoffs, coordination | DevCom 2026 guide |
| Rule-based custom agent | $10,000 to $30,000 and up | Custom build plus hosting and support | SoftTeco 2026 |
| Model-based custom agent | $40,000 to $80,000 and up, plus $20 to $5,000 a month cloud | Custom build with integrations and data pipeline | SoftTeco 2026 |
| Managed agent service | From $200 a month, setup scoped per customer | Agents built, hosted, monitored, and maintained for you | VantaSoft pricing |
Two things to take from that table. First, the same word, agent, covers a $500 script and an $80,000 program, so any quote without a description of the workflow is a guess. Second, custom builds put the cost up front and leave you running the thing; managed service spreads it into a monthly fee and leaves the running to the vendor. Which is cheaper depends on how long you keep it and who on your team would otherwise operate it.
What you are actually paying for: the four cost layers
An AI agent does not have one price. It has four cost layers, and vendors publish the one that is easiest to publish.
| Cost layer | What it covers | How it is billed | What drives it up |
|---|---|---|---|
| Setup | Discovery, workflow design, integration configuration, environment provisioning, QA, support through testing | One time, scoped per customer | More workflows, more integrations, systems without a documented API, messy source data, complex approval rules |
| Subscription | The managed environment the agents run in, monitoring, updates, issue investigation, support, minor configuration changes | Recurring monthly, by agent capacity | More agents in production, dedicated support, service level commitments |
| Model and third-party usage | Language model API calls and any outside subscription the agent needs | Metered by volume, or a separate vendor subscription | Higher run frequency, longer documents, larger models, more reasoning steps per run |
| Your team's time | Discovery sessions, granting access, testing, approving output, reviewing exceptions | Not invoiced, but real | Unclear workflow ownership, slow access approvals, no named reviewer |
Only the second layer looks like a subscription. The other three are where budgets break, and they are the three that get left out of the spreadsheet.
Why setup is scoped instead of listed
Almost no serious vendor publishes a setup price, and buyers reasonably read that as evasion. It usually is not. Setup is the part of the work that is different for every customer, because it is the part that touches your systems.
Configuring an agent to read one shared inbox and draft replies for review is a different job from wiring the same agent into a CRM, a project board, and an accounting file with three approval gates. Same product, very different amount of work. What you should expect is a setup number that arrives after a scoping conversation and before you sign anything. If a vendor wants a deposit before it can describe your workflows back to you, that is the problem worth reacting to.
How to make setup smaller: start with one workflow rather than a department, prefer systems that already have a documented API, decide who approves what before the build starts, and clean the data the agent has to read.
The six cost drivers that move the number
When two quotes for the same idea come back far apart, it is almost always one of these differences.
- How many workflows. One well-defined workflow is a project. Five is a program. Cost tracks workflows far more closely than it tracks agents.
- How many integrations, and what kind. A system with a documented API and clean permissions is straightforward. A system with no API is a build.
- How much judgment the work needs. Classifying and routing is cheaper than reconciling conflicting sources or weighing exceptions.
- Volume and frequency. Volume shows up in the usage layer, not setup. A workflow that runs twice a day and one that runs two thousand times a day cost about the same to configure and very different amounts to run.
- Approval and oversight design. Every action that must pause for a person is a rule someone has to define, build, and test. See what access an AI agent should have to your business systems.
- Data readiness. If the records the agent reads are inconsistent, someone pays to make them consistent, either the vendor during setup or your team afterwards.
If you have not chosen the first workflow yet, where AI fits into existing workflows and the eight practical agent roles are the faster place to start.
Model usage: the line item people forget
Language models are billed by the token, quoted per million tokens, and split between input and output. Output costs several times more than input at every major vendor.
As of August 2026, Anthropic's published list price for Claude Sonnet 5 is $2 per million input tokens and $10 per million output tokens. OpenAI's published list price for gpt-5.6-terra is $2.00 per million short-context input tokens and $12.00 per million output tokens.
A worked example. Suppose one run of your workflow reads about 20,000 tokens of context and produces about 2,000 tokens of output. At Claude Sonnet 5 list prices that is $0.04 of input plus $0.02 of output, roughly $0.06 per run. A thousand runs a month lands near $60. A real agent rarely makes one model call per run, so multiply by the number of steps. Any third-party usage cost should be identified during discovery, named in your order form, and never charged without your written approval.
Frequently asked questions
How much does it cost to hire someone to build a custom AI agent?
Published development-shop rates run from a few hundred dollars for a simple rule-based agent to about $20,000 for a multi-agent system, and custom builds with integrations and hosting are commonly quoted at $10,000 to $80,000 or more. Ask for the quote by workflow, with integrations named, and ask who runs the agent after launch.
What does an AI agent cost per month to run?
Two recurring lines: the environment it runs in, and model usage. A managed service bundles the environment from about $200 a month. Model usage scales with volume; a workflow running a thousand times a month on a mid-tier model is on the order of tens of dollars, and a high-volume, many-step workflow can be hundreds.
Is it cheaper to use a no-code AI agent builder than to hire a developer?
For a single simple workflow that lives inside one tool, a no-code builder is usually cheaper. Once the agent has to touch two or more business systems with approval rules, the no-code path tends to cost more in your team's time than it saves in fees. Build vs buy for AI agents works through the decision.
Are there hidden fees in AI agent development?
The three that surprise buyers: setup that was never scoped, model usage that scales with volume, and your own team's hours for access, testing, and review. Ask every vendor to put all four cost layers in writing before you sign.
What the same thing costs to build in house
The in-house comparison is worth doing and usually done wrong, because people compare a build estimate against a subscription and forget who operates the agent afterwards. Count the developer time to build, the time to keep integrations working as vendors change their APIs, monitoring, and the person who investigates when a run fails at 2am. If that person is your best engineer, the true monthly cost of an in-house agent is that engineer's hours, not the cloud bill.
VantaSoft builds and runs agents as a managed service from $200 a month, with setup scoped against your actual workflows during discovery. If you want a number rather than a range, bring the one workflow you would automate first and we will scope it.
