The cost of building an AI agent depends much more on what it has to connect to, and how much could go wrong, than on the AI model. You pay once for design, integrations and testing, then every month to keep it running: model usage, voice and phone minutes if it takes calls, hosting, monitoring and maintenance.
We don’t publish fixed prices. The gap between a simple agent and a complex one is wide, and a number without a scope would mislead you. This guide explains how each part of the cost behaves so you can read any proposal with a critical eye, ours included.
Build cost vs running cost
Every AI agent has a one-time build cost and an ongoing running cost. A proposal that shows only one of them is incomplete, and the second one is usually where budgets go off track.
Build cost vs running cost
| Criterion | Build (one-time) | Running (monthly) |
|---|---|---|
| What it covers | Discovery, task and conversation design, integrations, knowledge prep, guardrails, testing, launch | Model usage, voice and telephony, hosting, monitoring, maintenance, human review time |
| What drives it | Number of systems and tasks, risk, channels, languages | Volume, conversation length, model choice, how often your information changes |
| How it’s usually priced | Fixed project fee or time and materials | Usage passed through plus maintenance, or a monthly plan |
| Where it catches people off guard | Integrations with legacy or poorly documented systems | Long conversations and bloated prompts at high volume |
Build (one-time)
- What it covers
- Discovery, task and conversation design, integrations, knowledge prep, guardrails, testing, launch
- What drives it
- Number of systems and tasks, risk, channels, languages
- How it’s usually priced
- Fixed project fee or time and materials
- Where it catches people off guard
- Integrations with legacy or poorly documented systems
Running (monthly)
- What it covers
- Model usage, voice and telephony, hosting, monitoring, maintenance, human review time
- What drives it
- Volume, conversation length, model choice, how often your information changes
- How it’s usually priced
- Usage passed through plus maintenance, or a monthly plan
- Where it catches people off guard
- Long conversations and bloated prompts at high volume
On our AI agent development projects, the proposal always shows both numbers: the build, and the expected running cost at your volume.
Where the build budget goes
Here’s where the work goes when we build an agent, roughly from smaller to larger. Your project may weigh things differently, but the model itself is rarely the expensive part. It’s a configuration choice.
| Component | What it involves | What makes it bigger |
|---|---|---|
| Discovery and design | Reviewing real conversations, listing tasks, defining "done" and the handoff rules | Undefined processes, lots of stakeholders |
| Knowledge prep | Gathering and cleaning up the content the agent answers from | Scattered, outdated or conflicting documents |
| Instructions and guardrails | Writing what it may do, must never do, and when it stops | Regulated topics, lots of exceptions |
| Integrations (tools) | Connecting to your CRM, calendar, help desk or order system, and handling their errors | No API, legacy systems, write access to sensitive records |
| Channel | Web chat, WhatsApp, SMS, email or phone | Phone (latency tuning, transfers), several channels at once |
| Testing and evaluation | Scripted scenarios, edge cases, attempts to push it off-script | High-risk actions, two languages, many tasks |
| Deployment and monitoring | Hosting, logging, alerts, gradual rollout | Strict data-location requirements, self-hosting |
On most agent projects, integrations and testing together make up the largest share of the build. If the agent has to reach several internal systems, the work looks more like an AI integration project than prompt writing.
What makes an AI agent more expensive to build
Run through this before you ask for quotes. Each item you check pushes the cost up, and the ones you leave unchecked are where you can keep the first version small.
Cost factors that raise the build price
How complex an agent you need
These aren’t price tiers. They’re a way to place your project on the scale before you talk to anyone. Most first projects for small businesses land at level 2.
Four levels of agent complexity
-
Step 1
Answering agent
Answers questions from approved content on a website or WhatsApp, collects contact details and hands off. Its only action is creating a lead. The smallest build.
-
Step 2
Task agent with one or two integrations
Books, reschedules or looks up records in one main system such as a calendar or CRM. Many AI receptionists sit here.
-
Step 3
Multi-system agent
Reads and writes across several systems (CRM, scheduling, billing, help desk), with different rules per task and approvals for sensitive actions.
-
Step 4
Multi-channel or multi-agent system
Phone and chat sharing one brain, or several specialized agents passing work along the same customer record. Worth it only when the volume justifies it.
How to estimate the running cost of an AI agent
You can estimate the running cost with one formula and the current rates on each provider’s pricing page. Rates change often, so use the ones posted the day you run the numbers.
Three things drive up model cost without anyone noticing:
- Long conversations. Each new turn resends everything before it, so every turn costs more than the last.
- Bloated prompts. Pasting your whole handbook into every request gets expensive. Pulling in only the relevant sections costs less and is usually more accurate. That pattern is explained in what is RAG.
- Tool loops. An agent that keeps calling the same tool because an error message is unclear burns through usage. Clear error responses fix it.
Most major model providers also offer prompt caching, which cuts the cost of resending the same instructions on every turn. Ask if your setup uses it.
Pricing models you’ll see from agencies
- Fixed project fee. One price for an agreed scope. Predictable, but changes cost extra.
- Setup plus monthly. A build fee, then a monthly fee for usage and maintenance. Common for agents that need ongoing review.
- Retainer. A monthly amount for a block of work, first building and then improving. A good fit for businesses planning several agents.
- Time and materials. Billed by the hour. Flexible for exploratory work, as long as there’s a cap.
Whatever the model, ask who pays the model and telephony providers. Some agencies pass usage through at cost, some add a margin, some bundle it in. Any of those is fine as long as it’s disclosed.
Hidden costs to plan for
How to keep the cost of an AI agent down
- Keep the first scope narrow. One task, one channel, one or two systems. Expand once you’ve read real conversations.
- Use fixed automation where the path is known. Don’t pay an agent to do what a simple workflow handles reliably. AI automation cost covers that side of the budget.
- Pick the model for each task. Small, fast models can classify and extract. Larger ones take on harder reasoning.
- Put approvals where mistakes are expensive, instead of trying to make the agent perfect.
- Favor systems with good APIs. A clean integration costs a fraction of a workaround.
Questions to ask before you sign
- What exactly is in scope: tasks, channels, systems, languages?
- What’s the build cost, and what’s the estimated monthly running cost at our volume?
- How are model, voice and telephony usage billed: at cost, with a margin, or bundled?
- How will the agent be tested, and do we review transcripts before launch?
- What does maintenance include, and what counts as a change request?
- Who owns the code, prompts, integrations and data if we part ways?
We start with one agent, one workflow and written success criteria, then send a fixed written proposal with the build cost and the expected running costs at your volume. To understand the work behind each line item, read our step-by-step guide on how to build an AI agent. If the concept itself is still fuzzy, start with what an AI agent is. Or describe the task you want an agent to handle, and we’ll tell you which level it falls under and what drives its cost.