An AI chatbot that gets an answer wrong sends an embarrassing message. An AI agent that gets an action wrong issues a refund, overwrites a customer record, or emails the wrong client. That difference, not model quality, is what you are paying for when you build an agent, and it is why the price is not a bigger version of a chatbot quote.
TL;DR
- Narrow agent, 1-2 integrations: $15K-$30K. Production agent, 3-5 integrations plus approvals: $30K-$60K. Orchestrated multi-step: $60K-$150K+.
- Running costs are $200-$2,000/month, higher than a chatbot because one task burns many model calls.
- 60-70% of an agent build is containment work: permissions, approvals, rollback, evaluation, logging. Not conversation.
- Cost drivers ranked: integration count, cost of a wrong action, step count, evaluation requirements. Model choice barely matters.
- If your use case is mostly answering questions, you want a chatbot and the chatbot cost breakdown is the right page.
- Timelines run 4-10 weeks. Anyone promising a production agent on your live systems in under two weeks is skipping the safety work.
Make sure you want an agent
Most businesses asking for an "AI agent" describe a chatbot: answer customer questions from our documentation. That is a genuinely different and cheaper build, covered in AI chatbot development cost. An agent is worth its premium only when the system needs to do things across your tools, not explain things. If you are not sure which you need, AI agent vs chatbot sorts it in a few minutes and can save you a five-figure mistake.
What actually makes an agent expensive
The conversational layer of an agent is the cheap part. It is the same technology as a chatbot and it is largely a solved problem. The cost sits in everything that stops a capable system from doing damage at speed.
Permission scoping. An agent that can update records can, without constraints, update the wrong ones. Every tool the agent can call needs explicit boundaries: which records, which fields, which conditions, what it may never touch. This is design work before it is engineering work, and it is where a good build spends its first week.
Approval steps. For anything consequential, a human confirms before the action commits. Building that well means holding state, presenting the proposed action clearly, handling timeouts, and dealing with the human who approves 40 pending actions at once without reading them.
Retry and rollback. Agents fail mid-chain. Step three succeeded, step four timed out, and now your system is in a state nobody designed for. Handling partial failure is ordinary distributed-systems engineering and it is a large share of the budget.
Evaluation harness. You cannot ship an agent that touches money or customer data without a test suite that runs realistic tasks and scores the outcomes. Unlike normal software tests, these are probabilistic, so the harness needs enough cases to detect a regression that only appears one time in fifty.
Action logging. Every decision, every tool call, every input and output, retained. You need this to debug, and you need it the first time a customer asks why the system did something.
None of that is AI work. All of it is why an agent costs two to four times a comparable chatbot.
The cost bands
| Agent type | Typical build | Monthly run | What you get |
|---|---|---|---|
| Narrow, single-purpose | $15,000-$30,000 | $200-$500 | One or two integrations, one task type, human approval on everything consequential |
| Production, multi-tool | $30,000-$60,000 | $400-$1,200 | Three to five integrations, conditional approval, evaluation harness, full logging |
| Orchestrated / multi-step | $60,000-$150,000+ | $800-$2,000+ | Long task chains, multiple specialised agents, rollback handling, monitoring dashboards |
Narrow, single-purpose is where most businesses should start. One job: qualify an inbound lead and write it to the CRM, or read an invoice and file it into accounting. One or two systems. A human confirms anything that matters. This is the agent equivalent of a first automation, and it is the version most likely to survive contact with reality.
Production, multi-tool is the common serious build. The agent reads from several systems, decides, and writes to a couple of them, escalating what it is unsure about. The jump in price over the narrow band is almost entirely integration count and the conditional approval logic that comes with it.
Orchestrated is genuinely expensive and genuinely rarer than the market implies. Multiple agents, long chains, agents checking other agents' work. Most businesses quoted at this level do not need it, and should ask hard whether the requirement is real or architectural enthusiasm.
The four cost drivers, ranked
1. Integration count. This is the dominant variable. Each system the agent writes into brings authentication, rate limits, error semantics, data model mismatches, and a set of edge cases that only appear in production. Rough rule: each additional write integration adds $5,000 to $12,000 depending on how well documented the target system is. Read-only integrations are much cheaper than write integrations, often by half.
2. Cost of a wrong action. An agent that drafts internal summaries needs light guardrails. One that issues refunds, changes bookings, or sends client-facing email needs approval workflows, rollback, and audit trails. This single question can double a quote, and it should. Ask yourself what the worst plausible wrong action costs you, and price accordingly.
3. Step count. A two-step task has few failure modes. A seven-step task has many, and they interact. Each step added to a chain increases testing surface more than linearly, which is why keeping agents narrow is a cost strategy and not just an engineering preference.
4. Evaluation requirements. Anything touching money, health, legal, or customer records needs a real evaluation suite, typically $5,000 to $15,000 of the build. Internal-only tools with a human reviewing every output can get away with much less.
Notably absent: model choice. Whether you use one frontier model or another moves running cost somewhat and build cost almost not at all. Vendors who lead with model selection as a cost lever are talking about the least important variable.
Running costs, and why they surprise people
Agents cost more to run than chatbots, and the reason is structural. A chatbot answering a question makes roughly one model call. An agent completing a task plans, calls a tool, reads the result, decides what to do next, possibly calls another tool, and checks its own work. That single task might consume five to fifteen model calls.
So running cost scales with tasks x steps per task, not just with volume. A modest-volume agent doing complex tasks can cost more monthly than a high-volume chatbot.
Budget components:
- Model usage - the largest and most variable piece, $100 to $1,500/month at typical volumes
- Hosting and infrastructure - $50 to $300/month
- Monitoring and logging - $50 to $200/month, and worth every dollar the first time something goes wrong
- Maintenance - integrations break when vendors change APIs; budget a few hours a month or a small retainer
Ask any vendor for a monthly figure at twice your expected volume. If they cannot produce one, they have not run an agent in production.
Where buying beats building
Buy when your use case is mostly answering. Platforms like Intercom Fin resolve documented questions well, immediately, with no build risk, and you should not pay to rebuild that. The Zendesk, Fin, and custom agent comparison covers where that line sits.
Build when your use case is mostly doing, when the logic is specific to your business, or when per-unit platform pricing has compounded past a fixed build cost over a 12 to 24 month horizon.
The trap in the middle: buying a platform for an action-heavy use case. You pay the platform fee and then pay for custom integration work anyway, because the platform was never designed to write into your systems. That combination is the most expensive outcome available and it is disappointingly common.
What a fair quote looks like
A quote you can trust breaks out:
- Integration count, listed by system, and read versus write for each
- Which actions require human approval and which commit automatically
- Whether an evaluation harness is included, and how many test cases
- Monthly running cost at your expected volume and at twice that
- Who owns the code, the prompts, and the evaluation suite
- What happens when an integration breaks after launch
A quote that is one number with a timeline is not a quote, it is a hope. And a six-figure quote for a first agent is almost always the wrong shape of project. Start narrow, prove the task, then expand volume before scope. The 30-day first automation plan applies to agents as much as to simpler automations.
The bottom line
A custom AI agent costs $15,000 to $60,000 for most real business use cases, plus a few hundred to a couple of thousand dollars a month to run. The number is set by how many systems it writes into and how expensive it is when it gets one wrong, not by which model sits underneath.
Before you price anything, confirm you need an agent rather than a chatbot. The most expensive AI mistake available in 2026 is not choosing the wrong vendor, it is buying an agent-shaped budget for a chatbot-shaped problem.
Next step: If your use case is answering rather than acting, the chatbot cost breakdown is the accurate page. To scope an agent properly, see AI agent development.
Frequently asked questions
How much does a custom AI agent cost in 2026?+
A narrow agent with one or two integrations typically runs $15,000 to $30,000. A production agent with three to five integrations and a human approval step runs $30,000 to $60,000. Multi-step agents with orchestration, evaluation harnesses, and rollback handling run $60,000 to $150,000 or more. Running costs add $200 to $2,000 a month depending on volume and how many model calls each task consumes. Integration count and the cost of being wrong drive the number far more than the AI itself does.
Why does an AI agent cost more than a chatbot?+
Because an agent takes actions and a chatbot produces text. A wrong chatbot answer is an embarrassing message. A wrong agent action is a refund issued, a record overwritten, or an email sent to the wrong client. That difference forces engineering a chatbot never needs: permission scoping, approval steps, retry and rollback logic, evaluation harnesses, and full action logging. Typically 60 to 70 percent of an agent build is this containment work, not the conversational layer.
What actually drives the cost of an AI agent build?+
Four things, in order. Integration count, because each system it writes into brings authentication, error handling, and edge cases. The cost of a wrong action, which determines how much approval and rollback machinery you need. Step count, since each additional step in a chain multiplies the ways it can fail. And whether an evaluation harness is required, which is essentially non-negotiable for anything touching money or customer records. Model choice barely moves the number.
What does it cost to run an AI agent per month?+
Usually $200 to $2,000 a month for a production agent at moderate volume. That splits into model usage, which scales with task count and how many reasoning steps each task takes, plus hosting, monitoring, and logging infrastructure. Agents cost noticeably more to run than chatbots because a single task can consume many model calls as the agent plans, calls tools, and checks its own work. Ask any vendor for a monthly figure at twice your expected volume.
Is it cheaper to buy an AI agent platform than build one?+
Almost always at low volume, and often not at high volume or where actions are involved. Platforms are excellent at answering and weak at writing into your specific systems, so action-heavy use cases end up requiring custom engineering anyway, on top of the platform fee. The honest rule is to buy if your use case is mostly answering questions, and to price a build if most tasks end in a system change. Volume and action complexity decide it, not preference.
How long does it take to build a custom AI agent?+
Four to ten weeks for most production agents. A narrow single-integration agent can ship in three to four weeks. Anything with multiple integrations, an approval workflow, and an evaluation harness realistically takes eight to twelve. The integration and testing work dominates the timeline, not the AI development. Any vendor promising a production agent touching your live systems in under two weeks is either scoping something trivial or skipping the containment work.
Free PDF · No fluff
The 2026 AI Development Rate Sheet
Real build, agent, RAG, and consulting rates by tier — the numbers vendors quote behind NDAs, in one PDF.
Written by
Pankaj Kumar
Founder · Metageeks Technologies
Metageeks builds production-ready AI products for $1M–$15M companies — shipped in fixed-price sprints, not open-ended retainers. We write about what actually works in the field.
Connect on LinkedInThe AI Build Brief
Ship AI that actually works.
Practical playbooks on building, pricing, and shipping production AI — one email, every other week. No fluff.
