AI automation cost: what drives the price and how to model it

AI automation cost has two parts: a one-time build and a monthly run cost. Build cost depends mostly on how many systems the workflow touches, how varied the inputs are and how many exceptions need rules. Run cost depends on volume, model usage, hosting, monitoring and the human review time that remains. Scenox does not publish prices, so this guide gives you the cost model and a worked example built from stated assumptions, which you can replace with your own numbers.

#How AI automation agencies charge

Agencies mostly use one of four pricing models, and the model matters as much as the number.

1. Fixed fee per workflow: one price for a defined scope. It works when the workflow has been mapped first, because the agency can see the exceptions before quoting.
2. Time and materials: you pay for hours at a day or hourly rate. It's flexible, and the risk of overruns lands on you.
3. Monthly retainer or managed run: a flat fee covers monitoring, fixes and small changes after launch.
4. Usage-based: you pay per document, call or run. It tracks volume, and it can surprise you at the end of a busy quarter.

A quote that arrives before anyone has looked at your workflow is a guess. If a vendor names a price on the first call, ask what it assumes about systems, volumes and exceptions. The answer shows how much scoping went in.

#Build cost drivers

Build cost is mostly engineering and testing hours. These factors move the hours.

The last row surprises people. Waiting on credentials and sample files burns calendar time, and often billable time too.

#Run cost drivers

Run cost repeats every month for as long as the system is live.

• Model usage is priced per token, so it scales with volume and document length.
• Hosting covers servers or cloud services, queues, storage and secrets management.
• Software licenses: a workflow tool plan, an OCR or document service, a telephony or CRM add-on.
• Monitoring and maintenance. Someone watches failures, re-tests when a supplier changes its invoice layout or a model version changes, and handles small requests.
• Human review is staff time on assisted items and exceptions. It is usually the largest run cost. It sits on your payroll, so vendor quotes tend to leave it out.

Our ROI guide shows how to put these lines next to the savings. This post focuses on the cost side.

#Public price points we checked

Some inputs are published, so you can price them yourself. We checked these on 2026-10-15. Prices change, so confirm before you budget.

Token prices are the model cost only. They exclude the engineering to build prompts, validation and review screens around the model, which is where most of the budget goes.

#A worked example (illustrative)

These numbers come from assumptions, not from any Scenox price list or client. Change the assumptions and the output changes.

The workflow. A finance team receives 2,000 supplier invoices a month in about a dozen layouts. The automation reads each invoice, matches it to a purchase order in the ERP, posts it or sends it to an approver, and flags exceptions. Systems touched: the ERP, a shared inbox and an approval channel.

#Build: hours times rate

Assumed blended engineering rate: $100 to $200 an hour. Hours per phase:

Low case: 200 hours at $100 is $20,000. High case: 380 hours at $200 is $76,000. The spread is wide on purpose. A scoped audit exists to shrink it, because it replaces "about a dozen layouts" with a counted sample and "three integrations" with checked API access.

#Run: monthly

On your side, add review time. Using the split from the ROI guide (600 assisted invoices at 1.5 minutes, 200 exceptions at 6 minutes, 10 oversight hours), that is 45 hours a month, or about $2,025 at $45 an hour.

#First-year view

Build plus 12 months of vendor-side run cost: $20,000 + $7,824 = about $27,800 at the low end, and $76,000 + $40,560 = about $116,600 at the high end. Now set that against the manual process. At 6 minutes an invoice and $45 an hour, 2,000 invoices cost $9,000 a month, or $108,000 a year, before errors. In this scenario the case holds up. Drop volume to 300 invoices a month and it falls apart. Run your own numbers before you trust either.

Key point: The build price is the visible part of AI automation cost. Over three years, monitoring, model usage and human review can add up to more than the build. Ask every vendor for both numbers.

#Where costs get out of hand

• Scope drift. "Also handle credit notes" is a new workflow with its own exceptions. Price it separately.
• Long documents. A 40-page contract uses far more tokens than a one-page invoice, so measure on your own samples.
• Retry loops. An agent that retries and calls tools several times multiplies token use per item.
• No owner after launch. A system nobody maintains degrades, then needs an expensive rescue.
• Automating a low-volume task. A job that takes ten hours a month rarely repays a custom build.

#Build custom, use a workflow tool, or buy SaaS

If the workflow is simple, structured and sits between popular apps, a hosted workflow tool is usually the cheapest start. Our comparison of n8n, Zapier and Make covers what each one charges for. If the job is a standard category such as accounts payable, packaged software may already do it, and our AP automation software guide compares the options. Custom work pays off when your process, ERP or approval rules don't fit the packaged product.

#How to get a quote you can compare

Ask each vendor to state, in writing:

• the volumes and number of systems the price assumes
• what is included after launch, and for how long
• who owns the code, prompts and evaluation set
• how model usage is billed, and whether you pay the provider directly
• how change requests are priced
• what a 90-day pilot would cost versus the full build

Our sample audit shows the kind of scoping output that makes a fixed price possible: a baseline, ranked opportunities and an estimate of the automatable share. If you want one for your workflow, book a strategy call and we will tell you whether it is worth doing.

#Frequently asked questions

#How much does AI automation cost?

It depends on scope, and anyone quoting one number without seeing your workflow is guessing. Think in two parts. The build is a one-time cost driven by integrations, input variety and exceptions. The run cost is monthly: model usage, hosting, monitoring and review time. The worked example above shows how to get a range from your own assumptions.

#How much do AI automation agencies charge?

Agencies charge fixed fees per workflow, hourly or daily rates, monthly retainers, or usage fees. Rates and totals vary widely by region and team. Compare proposals on what they include, such as testing, handover and support, and not only on the headline price.

#Is it cheaper to use Zapier, Make or n8n than to build custom?

For structured data moving between common apps, usually yes. Workflow tools are cheap to build with and have published plans. Once you need document reading, complex approval rules or strict audit trails, you may add custom code, and the comparison becomes about total cost of ownership.

#What running costs do people forget?

Monitoring and maintenance, and the human time for review and exceptions. Model usage is typically the smaller line.

#Can I get a fixed price?

Often, once the workflow has been mapped and tested on a real sample. Before that, a range is the honest answer.