The cost to build a custom AI agent is the question SMB owners ask earliest and get the most varied answers on. I've seen quotes for "the same" agent range from $8,000 to $120,000 depending on who's quoting and what's included. The wide spread isn't because the market is broken. It's because "build me an AI agent" can mean very different things, and most quotes don't define what they're including.
This post is the honest breakdown. What an AI agent costs to build, by complexity tier, with the line items broken out so you can spot what a vendor is or isn't including.
For an SMB building a first custom agent in 2026:
Low-complexity build (one input source, one output target, light integration): $10,000 to $20,000.
Medium-complexity build (two integrations, basic human-in-the-loop UI, eval harness): $20,000 to $40,000.
High-complexity build (three or more integrations, multi-step reasoning, custom UI, real eval infrastructure): $40,000 to $80,000.
Ongoing operating cost: $50 to $500 per month for model API and infrastructure, plus internal staff time (3 to 8 hours a week in the first quarter, 1 to 2 hours a week ongoing).
These ranges hold for typical SMB engagements. They go up for regulated industries (healthcare, financial services) because compliance and audit add real work. They go down somewhat for SMBs that already have strong internal engineering and just need help with the AI-specific parts.
I cover the full ROI math against these cost ranges separately. The TL;DR: at the middle of the cost range, payback is realistic in 6 to 12 months for a well-scoped triage or document agent.
When I look at an agent quote, I'm checking for six categories of work. If any of them is missing or under-priced, the quote is incomplete.
Time spent understanding the workflow, defining inputs and outputs, picking the model, sketching the agent architecture.
Typical cost: $2,000 to $8,000. This is usually folded into the project's first 2 weeks.
Underpriced quotes skip this and discover the workflow during the build, which costs more later.
The actual code that calls the model, applies the business rules, formats outputs, handles errors. The "AI part" of the project.
Typical cost: $4,000 to $15,000. Surprisingly small as a fraction of the total. The model orchestration is usually a few hundred lines of code.
This is the line item vendors love to talk about because it's where the AI happens. But it's not where the project succeeds or fails.
Wiring the agent into your existing systems. Reading from your CRM, support inbox, document store, whatever. Writing back to the queue, database, or downstream tool.
Typical cost: $5,000 to $25,000. This is usually the biggest line item.
Underpriced quotes treat integration as an afterthought. The pattern is "we'll get the agent working in isolation, then figure out the integration". That's the pattern that produces 3-month overruns.
If your quote has integration at less than 30% of the total, push back. Either the workflow has unusually simple integration, or the vendor is under-scoping.
The UI or workflow tooling that lets a human review and approve the agent's outputs before they take effect.
Typical cost: $3,000 to $15,000.
Some workflows can use existing tools (a queue, a CRM stage, a Slack approval flow) and need no custom UI. Some need a real review interface, which is real software development.
Under-budgeted quotes assume the human-in-the-loop will be solved by existing tools and discover later that it won't be.
The infrastructure to know whether the agent is working in production. Eval datasets, scoring rubrics, dashboards, alerts.
Typical cost: $3,000 to $12,000.
The single most under-priced line item in the market. Many vendor quotes don't include evaluation at all. The result is an agent that ships and silently degrades.
If your quote has nothing called "evaluation" or "monitoring", you don't have a complete project. Either ask the vendor to add it, or budget to add it yourself before launch.
The work to hand the agent off to the team that will operate it. Documentation, training, first-quarter tuning support.
Typical cost: $2,000 to $10,000.
Often missing from initial quotes because vendors think their job ends at deployment. Real builders include this because they've felt the pain of agents that drift in week 5 with no operating owner.
A complete quote includes all six line items, sized appropriately for the workflow. Roughly 10% scoping, 15% agent code, 35% integration, 15% HITL, 15% eval, 10% handoff.
The biggest single variable in the cost is the number and complexity of integrations.
Single integration (one read source, one write target, well-documented APIs): low end of the range.
Two integrations: middle of the range. Most SMB first agents land here.
Three or more integrations: high end of the range. Each additional integration adds non-linear cost because of error handling and consistency concerns.
Integrations with systems that have weird APIs, no API at all, or undocumented behavior: anything from 1.5x to 3x the simple-integration estimate.
For an SMB scoping a first agent, the cheapest move is to pick a workflow that integrates with one or at most two systems that have good APIs. The agent that reads from HubSpot and writes to Slack is cheap. The agent that reads from your custom internal database and writes to three downstream tools is not.
I've covered the build vs buy decision separately. Integration cost is a major reason build sometimes loses to buy: if the SaaS handles your integrations natively, you skip the most expensive line item.
The other big cost driver is how accurate the agent needs to be.
"Good enough that a human catches the bad cases" is the cheap version. The eval bar is permissive. The error budget is generous. Fits triage and drafting agents where human review catches mistakes.
"Production-grade with measured reliability" is the expensive version. Real eval infrastructure, careful prompt engineering, multiple model A/B tests, formal accuracy targets. Needed for document extraction agents that write to accounting systems, or anything where errors propagate downstream without human review.
The cost difference between the two versions of the same agent can be 2x. Owners who don't think about this end up either over-spending on eval rigor they don't need, or under-spending on eval and shipping a fragile agent.
The honest version of this conversation: most first SMB agents should be the cheaper "human catches the bad cases" version. Get the agent in production, build operating muscle, then invest in the heavier eval for the agents that need it.
The third driver is whether the partner stops at deployment or sticks around for the first quarter.
Stop-at-deployment quotes are 15 to 25% cheaper than full-handoff quotes. They look like a better deal upfront. They produce more failed agents downstream because the operating ownership question doesn't have a clear answer.
Full-handoff quotes include the partner's involvement in tuning, eval, and operating support for 8 to 12 weeks after deployment. The premium is worth paying for a first agent because the team is learning to operate AI for the first time. For agents three and four, you can do without it.
A good vendor offers both versions explicitly and tells you which one fits your situation. A bad vendor only offers stop-at-deployment because they don't actually want to operate agents.
The "hidden costs" people complain about with AI projects mostly aren't hidden. They're just rarely quoted explicitly. I cover them in detail in hidden costs of AI implementation. The short list.
Model API costs above the initial estimate. Quotes use the cheap model tier. Production runs use a mix that ends up 2 to 4x the original estimate.
Internal operating time. The owner spends 3 to 8 hours a week in the first quarter on operating the agent. At a loaded rate of $50 to $100 per hour, that's $4,000 to $10,000 of internal cost per quarter that doesn't show up on any invoice.
Model swap costs. Every 6 to 18 months you'll want to swap to a newer model. That's a few thousand dollars of work each time to re-evaluate and re-tune.
Infrastructure that grows with volume. Logging, eval storage, monitoring all scale with usage. Start at $50 a month, end at $400 a month for a busy SMB agent after a year.
None of these is a deal-breaker. All of them should be in your model when you're sizing the budget.
To make this concrete, a real worked example from a recent SMB engagement.
Workflow: triage agent for inbound support tickets at a B2B SaaS. Reads from Zendesk, classifies by category, drafts response for human review, queues exceptions.
Quote breakdown:
Total build: $35,700
Operating cost: $220/month model API + $80/month infrastructure + ~4 hours/week internal time ($800/month at loaded rate) = ~$1,100/month total operating.
Hours saved: 24 hours per week of support triage, at $35 loaded rate = $3,640/month savings.
Net monthly value: $3,640 - $1,100 = $2,540.
Payback: $35,700 / $2,540 = 14 months. Year-two net positive of about $30,000.
This is the honest version of the math. Vendors who quote $35,000 builds with $7,000 monthly value and 5-month payback are overselling. Real SMB AI agents pay back well, but in months not weeks.
For owners scoping their first build, the right move is to demand a quote with all six line items broken out, push back on any line item that's missing or obviously under-scoped, and budget your operating costs from the start. The vendors who can produce a quote in this shape are the ones worth working with.
RELATED READING
When SMB owners ask me about AI agents, the first thing I usually have to do is unscramble the word "agent". The term is used to mean five different things in five different vendor pitches, and most…
The build vs buy decision is older than AI. Every IT shop has had this conversation about every internal system since the 1990s. The AI version has a few new wrinkles but most of the framework is…
Vendor quotes for SMB AI projects routinely miss 20 to 50% of the actual cost. The missing pieces aren't hidden in the sense of dishonest. They're missing because the quote was scoped to what the…
Most small business owners I talk to are stuck in the same place with AI. They've watched ChatGPT do something impressive. They've signed up for two or three SaaS products with "AI" in the name.…
The first AI agent works. The team trusts it. The numbers add up. The owner is asking what's next, and the answer is usually "the second agent". Then the third. Then the fourth.
The ROI of AI for small business is one of the most lied-about numbers in the current vendor pitch landscape. Every vendor deck has a chart showing 5x or 10x return in the first year. Most of them…
FREQUENTLY ASKED