The short answer
Cost scales with how many decisions the agent is allowed to make on its own, not with how impressive the demo looks. A scripted assistant that answers questions from a knowledge base is a different project — and a different price — from an agent that reads your CRM, calls three internal APIs, and takes actions without a human in the loop.
These are build-cost ranges for a production system, not a weekend prototype — the gap between the two is most of this article.
What actually drives the price
Six variables move the estimate more than anything else:
- Channels. Web chat only is cheap. Add voice, WhatsApp, Slack, or email and you're maintaining several integrations with different latency and formatting constraints, not one.
- Knowledge and retrieval. A static FAQ is trivial. A retrieval system over your actual docs, tickets, and product data — kept in sync as those change — is most of the engineering effort on a lot of projects.
- Tool integrations. Every system the agent can call (your CRM, billing, internal APIs) needs an integration, error handling for when that system is slow or down, and a decision about what the agent is allowed to do versus just read.
- Memory. "Remembers this conversation" is cheap. "Remembers this customer across sessions, weeks apart, and updates its understanding of them" is a real data architecture problem.
- Autonomy level. The biggest cost multiplier. An agent that drafts an action for a human to approve is far cheaper to build safely than one that takes the action itself — the second one needs guardrails, rollback paths, and monitoring the first doesn't.
- Evals and guardrails. Testing an agent isn't like testing a form. You need a test suite of realistic conversations and edge cases, and a way to catch regressions when you change the prompt or swap models — this is ongoing work, not a one-time cost.
The gap between a working demo and a production agent is almost always the guardrails, not the "AI part." Getting an agent to sound smart in a demo takes a weekend. Getting it to fail safely when a user says something unexpected, a tool call times out, or the underlying model has a bad day — reliably, at scale — is the actual engineering work.
The cost nobody budgets for: running it
Build cost is a one-time number. Running cost is monthly, and it's easy to underestimate because a demo with ten test conversations looks free. At production volume, the model calls themselves are usually the smallest running cost — infrastructure, monitoring, and the human time to review flagged conversations tend to add up faster. A few things that keep running cost sane at scale:
- Model routing — sending simple queries to a cheaper, faster model and reserving the expensive one for genuinely hard requests.
- Semantic caching — not re-running the full pipeline for questions you've effectively already answered.
- Context management — trimming what actually gets sent to the model on each call, since cost scales with tokens in and out, not just calls.
None of this is exotic, but if nobody designs for it, running cost quietly creeps until someone asks why the AI feature costs more than the sales rep it replaced.
Build vs. buy vs. no-code
| Approach | Good for | Where it breaks down |
|---|---|---|
| No-code agent builder | Simple, single-purpose bots; validating an idea fast | Custom logic, non-standard integrations, and anything with real compliance requirements |
| Open-source framework + your team | Teams with in-house ML/engineering capacity and time to spend on infrastructure | Time-to-production — most of the work is undifferentiated plumbing, not your product |
| Custom-built for your workflow | Anything customer-facing, revenue-affecting, or handling sensitive data | Higher upfront cost than the other two options |
A rough way to estimate your own project
Before getting a quote from anyone, it's worth being honest with yourself about three things: how many systems does the agent actually need to talk to, does it take actions or just answer questions, and what happens when it's wrong. Those three answers narrow the estimate more than any feature list. If you want a number without waiting on a call, our AI Agent Builder walks through exactly this — channels, tools, memory, autonomy — and gives you a live cost and timeline estimate as you configure it.
Get a real estimate, not a guess
Configure your agent's channels, integrations, and autonomy level in our free builder and see the cost update live.
Frequently asked questions
Is a chatbot the same thing as an AI agent?
Not quite. A chatbot typically answers questions from a fixed script or knowledge base. An agent can take actions — look something up in a real system, update a record, trigger a workflow — which is what introduces most of the cost and risk discussed above.
Can I start small and add autonomy later?
Yes, and it's usually the right approach. Starting with a human-in-the-loop version (the agent drafts, a person approves) lets you validate the workflow and build trust in the outputs before removing the human step — which is also far cheaper to build safely the first time around.
Which model should the agent use?
Whichever is cheapest for a given request type while still meeting your accuracy bar — which is usually not one model for everything. This is what model routing solves, and it's a design decision, not a one-time choice you make on day one.
How long does a production agent take to build?
A single-channel assistant with one knowledge base is typically 3–6 weeks. A multi-tool agent with real integrations and evals runs 8–16 weeks. Enterprise systems with compliance and audit requirements can run longer — scope, not raw complexity, is usually the deciding factor.