Nobody in this industry publishes numbers, which is how you end up on a discovery call being told "it depends on scope" for thirty minutes.
Part of our guide to AI enablement for small business.
Here are the actual costs, broken into the four things you will pay for, with honest ranges for the three ways you can go about it.
The Four Cost Buckets
Every AI enablement budget breaks into the same four pieces. Vendors tend to quote you one of them and stay quiet about the rest.
1. Platform costs
The automation layer that runs your agents.
- Zapier — free tier at 100 tasks a month (two-step Zaps only), paid from about $20/mo for 750 tasks. Metered per task, meaning every action in every run.
- Make — free tier at 1,000 operations, Core from about $12/mo for 10,000 operations. Metered per operation, similar counting to Zapier but cheaper per unit.
- n8n Cloud — Starter around $20–24/mo for 2,500 executions. Metered per full workflow execution, so step count does not change the bill.
- n8n self-hosted — license free, you pay roughly $5–20/mo for a small server and own the upkeep.
Most small businesses land somewhere between $20 and $60 a month here. The variable that bites is per-task pricing: a six-step workflow firing a thousand times a month burns six thousand tasks on Zapier and one thousand executions on n8n. Same work, very different invoice.
2. Model usage
What you pay the AI itself.
For a small business running a handful of agents, this is usually $20 to $150 a month. A per-seat subscription like Claude or ChatGPT covers hands-on tool use, but agents typically run on API access, which is metered by volume.
Content generation is the expensive category because it processes and produces a lot of text. Routing and categorization agents are cheap because they read a little and output a little.
This is the bucket people most consistently overestimate at the start and then find is smaller than they feared once it stabilizes.
3. Build cost
Getting the agent actually working. This is the big variable, and it is a one-time cost per agent.
DIY. Free in cash, expensive in hours. A first simple agent takes most people 4 to 12 hours including the false starts. Once you understand your platform, subsequent agents take 2 to 4.
Contractor. $75 to $200 an hour. A well-scoped single agent is typically 5 to 15 hours, so $500 to $3,000 per agent.
Agency. Usually packaged rather than hourly. Expect $3,000 to $15,000 for a multi-agent buildout, depending on how many agents and how messy the integrations are.
4. Maintenance
The bucket nobody quotes.
Agents need upkeep. Tools change their APIs. Your process evolves. Edge cases surface that the original build never anticipated. Budget 10 to 25% of build cost annually, or a few hours a month of someone's attention.
Across our own forty-agent fleet, maintenance is about 23% of our total AI spend. Build is 41%, platform fees 22%, model usage 14%. If you had told me at the start that maintenance would outweigh model costs, I would not have believed you. The full breakdown is in what running 40+ agents actually costs and saves.
Three Paths, Three Budgets
Path 1: DIY
Cash cost: $40 to $200 a month, all-in Time cost: 10 to 30 hours to get your first two agents running Right for: Businesses with someone technically curious and more time than budget
You will build something that works. You will also spend a weekend learning why your first trigger did not fire. The knowledge you gain is genuinely valuable, because you end up able to maintain and extend what you built.
The honest risk: the project stalls because the person building it has a day job. Most DIY AI efforts die from competing priorities, not technical difficulty.
Path 2: Contractor
Cash cost: $500 to $3,000 per agent, plus $40 to $200/mo running costs Time cost: A few hours of your time defining what you want Right for: You know exactly which workflow you want automated and want it done properly once
Good fit when the scope is clear and self-contained. Weaker fit when you do not yet know what to build, because you will pay a contractor to help you figure that out at implementation rates.
The honest risk: you get an agent nobody owns. Contractors build and leave; if nobody internally understands it, the first time it breaks it stays broken.
Path 3: Agency or fractional
Cash cost: $3,000 to $15,000 for a buildout, or $2,000 to $10,000 a month ongoing Time cost: A discovery process, then review checkpoints Right for: Businesses that want a system rather than an agent, and want someone accountable for it working
You are buying the diagnosis as much as the build: which workflows to automate, in what order, with what gates. For a lot of businesses, that sequencing decision is worth more than the code.
The honest risk: paying for strategy you did not need on a workflow simple enough to have handled yourself.
What We Charge, For Reference
Since I am asking you to trust numbers, here are ours.
AI Workflow Audit — $997, one-time. We map your operations, identify the three to five highest-impact automation opportunities, and hand you a prioritized 90-day roadmap. Five business days. No obligation to build with us; several clients have taken the roadmap and built it themselves.
90-Day Operations Buildout — from $5,000/month, three months. Audit included. We design, build, test, and deploy three to five custom agents, wire them into your existing tools, train your team, and support for 30 days after handoff.
Ongoing AI Operations — custom, month to month. We run and evolve the system as the business grows.
The audit exists because the most common expensive mistake is building the wrong agent well. A thousand dollars to find out what to build is cheap insurance against a fifteen-thousand-dollar build of something nobody uses.
What Actually Drives Cost Up
Four things, in order of impact:
Integration mess. If your data lives in systems with poor or no APIs, build cost multiplies. An agent that needs to read a Google Sheet is cheap. One that needs to read a legacy system with no API is a project.
Undefined scope. "Help our marketing team move faster" costs three times what "QA every draft against its brief before human review" costs, because someone has to do the definition work either way, and it is more expensive at build rates.
Agent count. Each agent is largely independent work. Five agents is not much cheaper than five times one agent, though shared architecture does bring the marginal cost down.
High-consequence output. Anything client-facing or financial needs more testing, better gates, and more careful error handling. That is real additional work and it is worth paying for.
What Drives It Down
Pick workflows with existing triggers. If something already happens on a schedule or a status change, you are most of the way there.
Start with two data sources, not ten. Every additional integration adds build time and failure surface.
Accept a human gate. Agents that need to be right unsupervised cost dramatically more than agents that need to be right enough for a human to approve in ten seconds.
Build one, then reuse. Our fortieth agent took a fraction of the time our first one did, because the architecture and the patterns were already settled.
The Return Side
Cost only means something against return. From our own fleet: roughly 63 hours of capacity returned per week across a six-person team, against a spend where build is the largest line item and shrinking as a share over time.
Our first agent found $4,200 in unlogged billable time in its first month, which covered the entire program.
Your numbers will differ. But the shape usually holds: the first agent, if you pick the right one, tends to pay for the next several.
And to be honest about the other direction: four of our agents had negative ROI. We built them, they did not earn their keep, and we killed them. That is a normal hit rate and worth budgeting for emotionally as well as financially.
The Cheapest Possible Start
If budget is the binding constraint, here is the floor:
A free Make or Zapier tier, a $20 AI subscription, one workflow you already understand, and a weekend. Total cash cost: $20. If it works, you have proof. If it does not, you have learned something for the price of a lunch. If you want a shortlist first, we compared the options in the best AI agents for small business.
The expensive mistake is not spending too little. It is spending on capability before deciding what job you are handing over.
Want to know what your specific first agent should be? The audit answers exactly that, and you can take the roadmap wherever you like. Book a free AI Workflow Audit →
FAQ
How much does AI automation cost for a small business? Running costs typically total $40 to $200 a month for platform fees and model usage. Build cost is the larger variable: free if DIY, $500 to $3,000 per agent with a contractor, or $3,000 to $15,000 for an agency buildout.
Is it cheaper to build AI agents in-house or hire an agency? In-house is cheaper in cash and more expensive in time and risk of stalling. An agency costs more upfront but includes the sequencing decision of what to build first, which is where most of the expensive mistakes happen.
What are the ongoing costs of AI agents? Platform subscriptions ($20 to $60/mo), model usage ($20 to $150/mo), and maintenance at roughly 10 to 25% of build cost annually. Maintenance is the bucket most quotes leave out.
How long until AI automation pays for itself? A well-chosen first agent often pays back within one to three months, usually by returning hours or catching revenue that was leaking. Poorly chosen agents never pay back, which is why the selection decision matters more than the build.
Can a small business afford AI enablement? Yes, at the low end. A meaningful first agent can run for under $50 a month in tooling if someone internally builds it. Cost scales with agent count, integration complexity, and how much you outsource.
More in this series: AI agents for small business is the hub, and what running 40+ agents actually costs and saves covers the numbers from our side of the table.
