On this page
- 1What counts as an AI agent, and why it changes the price
- 2Where the build side of AI agent development cost goes
- 3How much does it cost to build an AI agent? Four tiers
- 4Running costs: the monthly bill
- 5AI voice agent cost per month
- 6AI agent pricing models vendors use
- 7Build or buy: when a custom agent pays off
- 8How to keep AI agent development cost under control
- 9Get a fixed estimate for your agent
Most quotes for an AI agent answer only half the question. They price the build and say little about what the agent costs to run once real customers start talking to it. AI agent development cost has two parts: the one-time work to design, connect, test, and launch the agent, and the monthly bill for model usage, voice minutes, hosting, and upkeep. Plan for both and you avoid the surprise that stalls a lot of agent projects around month three.
This guide covers both parts. It explains what drives the build, gives hour-based estimates for four agent tiers, walks through what a voice agent costs per month, and lays out the pricing models agent vendors use so you can compare a custom build against a subscription.
What counts as an AI agent, and why it changes the price
A chatbot answers questions. An agent takes actions. It looks up an order, books a slot in the calendar, updates a CRM record, issues a refund under a set limit, or hands the case to a person with a summary attached. Each action is a tool the model can call, and each tool is an integration someone has to build, secure, and test. That is why two projects both described as 'an AI agent' can differ in price by a factor of five.
The second factor is freedom. An agent that follows a fixed path, with one model call per step, is cheaper to build and test than one that plans its own steps, calls tools in a loop, and decides on its own when the job is done. More autonomy means more ways to fail, and more testing to prove it doesn't.
Where the build side of AI agent development cost goes
When we scope an agent, the hours fall into five areas. The shares below are a typical planning split for a multi-step agent. They shift depending on how many outside systems the agent touches.
| Work area | What it covers | Typical share of build hours |
|---|---|---|
| Scoping and conversation design | Task boundaries, what the agent may and may not do, escalation rules, sample conversations | 10-15% |
| Integrations and tools | API connections to CRM, scheduling, ticketing, ERP, or payment systems, plus authentication and error handling | 25-35% |
| Guardrails and permissions | Input filtering, action limits, approval steps, handling of personal data, audit logs | 10-15% |
| Evaluation and testing | A test set of realistic requests, automated scoring, regression runs before every change | 15-20% |
| Interface and admin tools | Chat widget, phone flow, or internal console, plus screens for reviewing conversations | 15-25% |
Two of these get underestimated almost every time. Integrations, because a vendor API that looks clean in its docs often has rate limits, odd pagination, or fields that mean different things in different accounts. Evaluation, because it is invisible in a demo. A demo shows the agent handling five friendly requests. A test set shows it handling three hundred messy ones, including the customer who asks for a refund halfway through a complaint about something else.
If the agent can move money or change records, give guardrails their own line in the budget. Our enterprise AI guardrails work covers patterns like approval thresholds, spending caps, and action logs that an auditor can read.
Key takeaway
A demo agent and a production agent can look identical on screen. The cost difference sits in the integrations, the guardrails, and the test set that proves the agent behaves.
How much does it cost to build an AI agent? Four tiers
Each estimate is typical build hours multiplied by $100-$175, the hourly range most US agencies charge. They are planning estimates, not quotes. Your number depends on the systems involved, the state of your data, and how much the agent is allowed to do without a person signing off.
| Agent tier | Example | Estimated hours | Estimated build cost | Typical timeline |
|---|---|---|---|---|
| Single-task agent | Answers warranty and policy questions from your documents and opens a support ticket when it can't help | 250-500 | $25,000-$87,500 | A few weeks |
| Multi-step agent with tool use and integrations | Looks up orders, processes returns under a set dollar limit, updates the CRM, and escalates edge cases | 600-1,200 | $60,000-$210,000 | MVP in a few weeks, then phased |
| Voice agent | Answers inbound calls, books appointments, takes messages, and transfers urgent calls to staff | 700-1,400 | $70,000-$245,000 | MVP in a few weeks, then phased |
| Multi-agent system | A coordinator routes work to specialist agents for intake, research, drafting, and review | 1,500-2,800 | $150,000-$490,000 | Phased, set by scope |
So the answer to how much does it cost to build an AI agent is a range tied to tier. A single-task agent makes a good first project because the scope is narrow and success is easy to measure. For most businesses the right first build sits in the first two rows. Multi-agent systems make sense when one agent's instructions grow so long that it starts mixing up tasks, or when separate steps need separate permissions. They are rarely the place to start.
For a step-by-step view of how a build runs from scoping to launch, read our guide on how to build an AI agent for business. Customer-facing chat agents fall under our AI chatbot development service.
Running costs: the monthly bill
The build is paid once. Running costs scale with usage and never drop to zero. Five lines make up the monthly bill.
- 1
Model usage
Providers such as OpenAI and Anthropic charge per token, with separate rates for input and output. An agent often makes several model calls per task because it plans, calls a tool, reads the result, and then writes a reply. Long system prompts and conversation history get resent on every call, so they add up fast.
- 2
Voice and telephony
Phone agents pay a carrier such as Twilio for numbers and call minutes, plus speech-to-text and text-to-speech services, usually billed per minute or per character.
- 3
Hosting
Servers or serverless functions for the agent, a database for conversation state, and often a vector database for document search.
- 4
Monitoring
Logging every conversation and tool call, tracing failures, and alerting when error rates climb.
- 5
Maintenance
Model versions get retired, vendor APIs change, and your own policies change. Budget a few hours of staff or agency time each month to update prompts, rerun the test set, and fix integrations.
Token prices change often and vary widely between models, so check each provider's current pricing page instead of trusting a figure in a blog post. Our breakdown of OpenAI and Claude API integration cost shows how to estimate token usage before you commit to a model.
A worked example makes the math concrete. Say a support agent handles 3,000 conversations a month, and each conversation uses 6 model calls averaging 4,000 tokens per call. That is 72 million tokens a month. At an illustrative blended rate of $3 per million tokens, the model bill is about $216. Move every step to a larger model at an illustrative $15 per million and it becomes $1,080. Same agent, five times the bill. That is why we pick the model step by step instead of using the biggest one everywhere.
AI voice agent cost per month
Voice agents add a per-minute layer on top of everything above. Each minute of a call can involve the carrier's charge, speech recognition, the language model, and speech synthesis. Some platforms bundle these into one per-minute price. Others bill each piece separately, which makes comparison harder.
Here is an illustrative example. A multi-location dental practice takes 1,500 calls a month with an average length of 4 minutes. That is 6,000 minutes. At an illustrative all-in usage rate of $0.10 a minute, usage comes to about $600. Add hosting, monitoring, and several hours of maintenance and the AI voice agent cost per month for that practice lands roughly between $1,300 and $3,000. Your real figure depends on current provider rates and how long your calls run.
Call length is the biggest lever. An agent that confirms an appointment in 90 seconds costs a fraction of one that holds a five-minute conversation about insurance coverage. We cover phone-specific numbers in our AI receptionist cost guide, and our AI receptionist development page shows what a custom phone agent includes.
Key takeaway
For voice agents, call volume times average call length drives the monthly bill more than any other setting. Design each conversation to reach its goal quickly.
AI agent pricing models vendors use
If you buy an agent product instead of building one, you will run into one of the pricing models below. We don't quote vendor prices here because they change too often, but the structure tells you who carries the risk when volume grows.
| Pricing model | How it works | Fits best when | Watch for |
|---|---|---|---|
| Per seat | Monthly fee for each staff member who works with the agent | Internal assistants used by a known team | Paying for seats that rarely log in |
| Per conversation or resolution | Fee for each chat, or for each case the agent closes | Support volume is steady and predictable | How the vendor defines a 'resolution' |
| Per minute | Fee for each minute of voice time | Phone agents with short, predictable calls | Hold time and transfers counting as billable minutes |
| Outcome-based | Fee tied to a result such as a booked meeting or collected payment | The outcome is easy to measure and attribute | Disputes over which results the agent caused |
| Platform plus usage | Base subscription plus pass-through model and telephony charges | You want some control over model choice | Two bills that both grow with volume |
A custom agent works on different terms. You pay for the build, then pay the model and telephony providers directly at their published rates. There is no per-conversation markup, so the gap between custom and subscription widens as volume grows.
Build or buy: when a custom agent pays off
Buy a platform when your use case is common, like answering website FAQs or qualifying inbound leads, and the platform connects to your systems out of the box. You can be live in weeks, and a subscription is easy to cancel if it doesn't work out.
Build when the agent has to act inside systems no platform supports, such as a homegrown ERP or a scheduling tool made for your trade. Build when per-conversation fees at your volume would overtake the build cost within a couple of years. And build when you need control over where customer data is stored and which model sees it.
A quick test: multiply the vendor's per-unit fee by your expected monthly volume, then by 24. If the result comes close to the build estimate plus two years of running costs, a custom agent deserves a serious look.
How to keep AI agent development cost under control
- Start with one task that has a clear finish line, such as 'book the appointment' or 'answer the warranty question correctly'.
- Collect 100-200 real requests from emails, chat logs, or call notes before development starts. They become the test set and expose scope gaps early.
- Use a smaller, cheaper model for routing and classification, and save the larger model for steps that need reasoning.
- Cap the number of tool calls per task so a confused agent can't loop and burn tokens.
- Give the agent read-only access first. Add write actions one at a time once the logs show it handles them well.
- Cache answers to repeated questions and trim conversation history the agent no longer needs.
Get a fixed estimate for your agent
The ranges in this guide are wide because agents vary so much. The fastest way to narrow them is to look at your actual task, your systems, and your volume.
Tell us what you want the agent to do, which tools it needs to touch, and how many conversations or calls you expect each month. After a free consultation we send a fixed-scope price for the build, a projection of monthly running costs, and a list of what we would leave out of version one to keep the first release lean. Book your free consultation to get that estimate.