“AI agent” is a broad label. A useful cost estimate starts with one narrow task, one success metric and one owner. If those are missing, the budget is a guess.
This is a model, not a quote. Use the AI agent cost calculator to plug in your own numbers.
The five cost components
- Build: discovery, prompt/tool design, integration, evals and deployment.
- Inference: input and output tokens per run, multiplied by volume.
- Tools and data: search, CRM, vector database, document storage, APIs.
- Maintenance: monitoring, prompt updates, model changes, eval reruns.
- Human review: the time a person spends approving or correcting outputs.
A worked example
Assume a lead-qualification agent that runs 2,000 times a month. Each run uses 8,000 input tokens and 1,000 output tokens. Model prices are illustrative, not a quote.
| Line | Assumption | Monthly cost |
|---|---|---|
| Input tokens | 16M tokens at $2.50 / 1M | $40 |
| Output tokens | 2M tokens at $10 / 1M | $20 |
| Tools/search | Search + CRM API | $50 |
| Monitoring | Logs, alerts, eval runs | $30 |
| Maintenance | 4 hours at $100/hr | $400 |
| Human review | 10 hours at $40/hr | $400 |
| Total run cost | ~$940/mo |
Build cost is separate: a 60-hour pilot at $120/hr is ~$7,200. If the agent saves 40 hours of manual qualification a month at $40/hr, the monthly saving is $1,600; payback is roughly 5 months before maintenance.
What makes it more expensive
- A broad use case (“do everything for sales”) instead of one narrow task.
- Dirty data and missing integrations.
- High-stakes actions without a human approval step.
- No eval set, so every change is a guess.
- Multiple models, tools and vendors with no observability.
- Compliance review, data residency or on-premise requirements.
Build vs buy
| Option | When it fits | Trade-off |
|---|---|---|
| Buy a vertical SaaS agent | A standard task and a small budget | Less control, per-seat pricing, vendor lock-in |
| Build on n8n + LLM | Custom integrations and internal data | You own evals, monitoring and maintenance |
| Build a custom app | Complex logic, multiple tools, strict SLAs | Highest cost, highest control |
How to budget
- Pilot: 3–4 weeks, one task, one metric, a fixed budget.
- Production: add monitoring, evals, access control and a runbook.
- Maintenance: 10–20% of the build cost per year, depending on volatility.
- Human review: budget it explicitly; it does not disappear.
FAQ
Written by
KEA CORP
SEO and automation team. We publish what we learn from live B2B projects, with sources where the numbers allow.
About KEA CORP →