How much does AI process automation cost?
Four layers of spend — platform, implementation, inference, and change — with real ranges for RPA, agencies, SaaS agents, and an open-source company OS.
Discuss this post in AI
Send a pre-filled prompt to ChatGPT, Claude, Gemini, or Perplexity — get a summary, ask follow-ups, or compare ideas from this guide.
The honest answer is three numbers, not one: what you pay to stand it up, what you pay to run it, and what you keep paying when the process changes. Most “how much does AI automation cost?” pages collapse those into a single licence fee. That is why the quote you got last quarter and the invoice you got this quarter do not match.
This post is a cost model, not a price list. Figures below are industry ranges from public RPA, systems-integrator, and LLM pricing — not a Pattern Automation quote. Use them to interrogate a vendor, a consultancy, or your own finance model. If you want the product that sits under the last column, that is Neuro OS.
The bill has four layers
| Layer | What it actually is | Typical range |
|---|---|---|
| Platform | Licences, seats, bots, or a hosted agent runtime | $0 (open source) to $50k–$250k+/year |
| Implementation | Discovery, connectors, skills, UAT, change management | $40k–$400k for a first process; $150k–$2M for a programme |
| Inference | Tokens, sandbox compute, retries | $0.02–$8 per run, depending on model and length |
| Change | Every exception, new system, new policy | 20–40% of year-one spend, every year after |
If a vendor quotes only the first row, they are selling you a product. You are buying an operating cost.
What companies actually spend
Traditional RPA
UiPath, Automation Anywhere, and Blue Prism still dominate the “we automated a process” slide. Public list pricing and integrator case studies cluster around:
- $8,000–$20,000 per attended/unattended bot per year, plus orchestrator and studio seats.
- $50,000–$500,000 to implement the first production process with a systems integrator.
- A dedicated RPA CoE once you have more than a handful of bots — two to six people, fully loaded $250k–$900k/year.
The hidden cost is brittleness. UI selectors break when a screen moves. The “cheap” bot that saved two FTEs can consume a developer-week every time Salesforce ships a layout change. That is not a footnote. It is the run-rate.
AI wrappers on top of RPA
The 2024–2026 pattern is “RPA plus a model.” You keep the bot farm and add document extraction, classification, or a chat front-end. You now pay both stacks: bot licences and tokens and a second integration layer. For a mid-size finance process that is often $150k–$400k year one before anyone has measured error rates.
Agent platforms (SaaS)
Hosted agent products typically bill seats, runs, or a platform fee plus your model keys. Public pricing in this category is still noisy. Treat $25–$150 per user/month plus usage as the band, and assume implementation is still a project: connectors, policies, evaluation, and a human in the loop. A 30-person operations team on a mid-tier plan is $15k–$50k/year in software before tokens.
Open source + your keys (Neuro OS)
Neuro OS is open source and self-hostable. You do not pay us a per-bot tax. You pay:
- Your model bill. A close pack, an access review, or a reporting run is a session on a real machine. Frontier models on a long finance task can be a few dollars; a short Slack answer is cents. Budget $500–$8,000/month for a serious internal fleet, then read the actual receipts — Neuro OS attaches spend to the work.
- Compute for sandboxes. Isolated Linux machines per session. Cloud VMs are cheap relative to people; they are not free. A busy team might look like $200–$2,000/month.
- People who own the repo. Not a 12-person CoE. One or two engineers plus the operators who write skills. That is the real line item, and it is the one that compounds: skills, memory, and policies stay in git.
Self-hosting on your VPC or on-prem shifts cost from SaaS margin to infra and security review. For a bank that already runs Kubernetes, that is often cheaper than another enterprise licence. For a 20-person firm with no platform team, hosted is cheaper until it is not.
The unit of cost in an agent system is not a bot. It is a session with a receipt.
A worked example: monthly financial reporting
Assume a 40-person finance team, one close pack, three source systems, human approval before anything posts.
| Approach | Year 1 | Year 2 (run + change) | What you own |
|---|---|---|---|
| Integrator + RPA | $180k–$350k | $80k–$160k | Bots in their orchestrator |
| AI automation agency | $120k–$280k | $60k–$140k retainer | Prompts and a Slack bot you cannot diff |
| In-house on a SaaS agent product | $90k–$200k | $50k–$120k | Workspace in their cloud |
| Neuro OS, self-hosted, your keys | $40k–$120k (mostly people + models) | $25k–$70k | The repo |
The Neuro OS column is lower on software and higher on your people writing skills. That is the trade: you stop renting a black box and start maintaining an operating system. See how to automate financial reporting with AI agents for the workflow, and AI automation agency vs in-house for who should own it.
Where the money actually leaks
- Retries and evaluation. A 4% failure rate that the agent silently re-runs will dominate the model bill. Log it.
- Unscoped connectors. One agent with payroll and email and the warehouse is an incident waiting to be priced in legal hours.
- No approval gates. The cheapest run is the one that sent the wrong vendor payment. Tool access is a cost control, not only a security control.
- Vendor lock on the workflow. If the process lives in a proprietary compiler, changing it is a change request to them. That is year-two spend you cannot staff internally. Who owns the code?
How to budget without lying to the board
- Pick one process with a known FTE cost and a known error cost. Do not start with “AI transformation.”
- Split capex-like implementation from opex. Implementation is a project. Inference and change are a run-rate.
- Put a hard cap on agent spend the way you cap a cloud account. Neuro OS was built so overruns are visible on the task, not in a surprise invoice. Budgets.
- Measure cost per accepted artifact, not cost per token. A $0.40 run that produces a reviewed variance note is cheap. A $0.02 run that a human has to redo is expensive.
- Assume the process will change. If your model only works if the chart of accounts is frozen, it is not a model — it is a hope.
The short version
AI process automation costs as much as the operating model you choose. RPA sells you bots. Agencies sell you a project. SaaS agent products sell you seats. Neuro OS sells you an open control plane: you bring models, you keep the files, you pay for what the sessions actually did.
If you want a number to put in a deck: a first production process is usually five figures to low six figures in year one, dominated by people and integration, not by the model. The question worth asking is whether year two is a licence cliff or a repo you already know how to maintain.