What Does an AI Agent Actually Cost? The Honest Math for SMBs
One website says AI agents cost $20 a month. Another says $50,000. Both are right. They’re describing different things.
That’s the first thing to get straight before you budget. “AI agent” has become a label applied to three very different products, each with a different cost shape. If you price one and assume it applies to the others, your budget will be off by an order of magnitude in either direction.
Here’s the model I use with clients. Three buckets: build, run, maintain. Every honest AI agent budget accounts for all three. Quotes that only cover one are giving you a partial answer.
The $20 vs $50,000 gap, explained
Three products share the name “AI agent”:
| What you’re buying | Typical price | What you get |
|---|---|---|
| SaaS chatbot or off-the-shelf agent | $20 - $200 / month | Prebuilt agent for a narrow use case. Limited custom logic, shared infrastructure. |
| Custom agent built for your workflows | $5,000 - $15,000 build + ongoing run cost | Agent that uses your tools, your data, your business rules. You own the setup. |
| Managed deployment with SLA | $3,000 - $10,000 / month | Custom agent plus ongoing ownership, monitoring, and 24/7 support. |
Three different products, three different buyers, three different problems. Before you can price anything, decide which one you’re actually shopping for.
Most SMBs land in the middle row. That’s the shape the rest of this post covers.
The three-bucket cost model
Think of an AI agent like a vehicle. You pay once to get it (purchase price), you pay to use it (gas, insurance, wear), and you pay to keep it running (maintenance, repairs). Skip any of the three and the math misleads you.
Same for an AI agent:
- Build. The one-time cost to get the agent deployed: scoping, integrations, testing.
- Run. The variable cost to operate it: model tokens, tool calls, storage.
- Maintain. The ongoing cost to keep it working: monitoring, updates, optimization.
Now let’s price each bucket.
Build cost
The build bucket covers getting from “we want an agent” to “the agent works in production.”
For a typical SMB with one or two workflows, build cost breaks down roughly like this:
- Audit or scoping: $2,500. Maps your workflows and decides what to automate first. Most Meridian engagements start here. Here’s what an audit covers.
- Core implementation: $5,000 to $10,000 for one workflow. Add $2,500 to $5,000 per additional workflow.
- Integrations: built-in to the above if you’re connecting Gmail, HubSpot, Slack, Google Sheets, or similar. Custom or legacy systems push the number up.
Total build cost for a typical SMB: $7,500 to $15,000. Full breakdown: AI implementation pricing.
This is where a lot of SaaS vendors hide cost. They quote the monthly run price and don’t mention you still need someone to set it up. If you’re scoping “free” agents, ask about setup time and who is doing it.
Run cost
The run bucket is the part nobody explains clearly, so here’s the math.
Every time your agent runs, it reads some context (inbox, CRM, documents) and takes action. That action costs money. The cost is almost entirely model tokens, with a smaller component for tool calls and storage.
Rough benchmarks at today’s model prices:
- Simple agent run (small context, one tool call): $0.005 to $0.02 each
- Typical agent run (reads an email thread, checks CRM, drafts a response): $0.02 to $0.05 each
- Heavy agent run (reads multiple documents, multi-step reasoning): $0.10 to $0.30 each
Multiply by your run volume. Most SMB agents run 500 to 2,000 times a month. That’s $10 to $100 a month in raw model cost.
Add tool and data costs: API calls to your CRM, document storage, vector DB queries. For most SMBs, another $30 to $150 a month.
Typical run cost for one agent: $40 to $250 a month. Higher if you’re running high-volume customer-facing workflows.
Maintain cost
Agents break. Models change. Your business changes. The maintain bucket is what keeps the thing working after launch.
For SMBs, two maintenance models are common:
- Light retainer: $1,000 to $1,500 a month. Monthly check-in, minor fixes, performance monitoring, one small change per month.
- Active retainer: $2,500 to $5,000 a month. Continuous optimization, new workflows, monthly review calls, priority support.
You can skip maintenance for a while. Most clients wish they hadn’t. The agent works fine until your CRM admin renames a field and suddenly nothing is routing correctly.
Honest answer: budget $1,500 a month minimum if the agent touches a revenue-critical workflow.
Year-one TCO for a typical SMB
Let’s put it together. A realistic first engagement for an SMB with one workflow and moderate volume:
| Bucket | Amount |
|---|---|
| Build (audit + one workflow) | $10,000 |
| Run (12 months at $150/mo avg) | $1,800 |
| Maintain (12 months at $1,500/mo) | $18,000 |
| Year-one TCO | $29,800 |
Year two drops to about $19,800 because the build cost is amortized. Year three, same.
A SaaS chatbot advertising $99/mo would be $1,188 a year. But it’s a different product, doing a different job, without your custom workflows or tool integrations. Compare like with like.
The comparison people actually want: agent vs new hire
Most SMBs aren’t really choosing between agents. They’re choosing between an agent and hiring a person to do the same workflow.
Rough math for a part-time admin or ops assistant in Montreal or most Canadian cities:
| Role | Salary / hourly | Year-one all-in |
|---|---|---|
| Part-time VA (20h/week) | $25/hr | $26,000 |
| Full-time admin (40h/week) | $45,000 - $55,000 salary | $55,000 - $68,000 (includes payroll taxes, benefits, equipment) |
| Ops coordinator | $55,000 - $70,000 salary | $68,000 - $85,000 |
At $29,800 year-one, an agent lands between a part-time VA and a full-time admin. By year two at $19,800, it’s below the part-time VA. That’s the number SMBs usually care about.
The agent doesn’t replace a whole role. It replaces the part of the role that is repetitive and doesn’t need a human. If that’s 60% of the admin job, the math is obvious. If it’s 10%, hire someone.
What drives your number up or down
Six factors. Quick reference you can screenshot:
- Number of workflows. Linear. One workflow is cheap. Five interconnected workflows multiply build cost.
- Integration complexity. Standard SaaS (Gmail, HubSpot, Slack) is built-in. Legacy ERP or custom databases add real hours.
- Volume. Run cost scales with how often the agent runs. Thousands of runs a day change the math.
- Model tier. Cheaper models for simple triage, better models for reasoning. Mixing them saves a lot.
- Data readiness. Clean, documented processes deploy fast. Messy undocumented workflows need prep work.
- Maintenance appetite. Light retainer is fine for stable workflows. Fast-changing businesses need active.
If you want the cheapest honest number: one workflow, standard integrations, moderate volume, light retainer. Roughly $20,000 year-one.
If you’re budgeting for the full operating-scale agent: three workflows, custom integrations, high volume, active retainer. Closer to $60,000 to $90,000 year-one, still cheaper than two full-time hires.
Where to start if you’re budgeting today
Don’t budget for an AI agent yet. Budget for an audit.
A $2,500 audit tells you which workflow to automate, how much it will actually cost, and what the ROI looks like before you commit. If you move forward, the audit fee is credited toward the build, so it costs nothing extra. If you don’t, you’ve spent $2,500 to avoid spending $30,000 on the wrong thing.
That’s the order of operations I’d give any SMB looking at this for the first time.
Want a realistic year-one number for your actual workflows before you commit to anything? Book a free 30-minute call and we’ll ballpark build, run, and maintain together. Here’s what the call covers. No commitment, no pitch.
And if you’re still figuring out whether you even need an agent or whether a Zap will do, start here: Zapier vs AI agent.