What AI workflow automation really costs

Buyers ask what AI automation costs and get either a shrug or a number with no shape to it. Both are unhelpful. The honest answer has four lines and the fourth one is where projects go wrong.

Line one: the build

Ranges I see holding across North America and Europe, for work that is genuinely production-grade rather than a demo:

• Single workflow, no retrieval. 3,000 to 12,000 dollars. Two to four weeks. Intake, classify, extract, route, notify, log, with error handling and monitoring.
• Agent system with retrieval and review steps. 15,000 to 60,000 dollars. Four to ten weeks. Document pipeline with citations, approval gates, audit trail, evals.
• Platform work. Above that, and now you are hiring, not buying a project. Be honest with yourself about which one you are doing.

Roughly 60 percent of that build time goes to things nobody demos: edge cases in real data, retry and idempotency logic, permissions, logging, and the eval suite. That is not padding. That is the difference between a working demo and a system that runs unattended.

Line two: model spend

Usually smaller than people fear, and almost always larger than it needs to be on the first pass.

A workflow processing a few thousand documents a month, built without cost discipline, might run 400 dollars in model calls. The same workflow with prompt caching, context trimming, and difficulty-based model routing typically lands closer to 80. That is not a trick, it is just engineering, and the sequence is in prompt caching, and the rest of the bill.

Line three: hosting and platform

• Orchestration. n8n Cloud on a small plan, or self-hosted on modest infrastructure with managed Postgres and Redis. Either way, tens of dollars to low hundreds per month. Sizing detail in self-hosting n8n at scale.
• Database and object storage. Usually trivial at these volumes.
• Retrieval infrastructure if you need it. A managed vector index or Postgres with a vector extension. The second option is cheaper and sufficient far more often than vendors suggest.
• Monitoring. Frequently free at low volume, and worth paying for the moment it is not.

Realistic total for a mid-sized system: 100 to 500 dollars a month, plus model spend.

Line four: maintenance, the line that gets omitted

This is where projects quietly fail. AI systems do not sit still, for four reasons that all arrive on their own schedule:

1. Model versions change. Providers deprecate and update. Behavior shifts. Without an eval suite you find out from a customer.
2. Your inputs drift. A new client sends a different document format. A form gains a field. The distribution moves and accuracy moves with it.
3. Integrations break. APIs deprecate, credentials expire, rate limits tighten.
4. Requirements grow. The system works, so people want it to do more. That is success, and it is also work.

Budget realistically at 10 to 20 percent of build cost per year, or a few hours a month of a competent person's attention. Systems with no maintenance owner are the ones I get called to resurrect, which is the subject of why AI pilots die at handoff.

It ran perfectly for five months. Then the vendor changed a field name and nobody owned it, so it stopped and nobody noticed for two weeks.

Worked example

Invoice intake for a firm processing about 2,000 invoices a month. Extract fields, validate against purchase orders, flag exceptions for a human, push the clean ones into the accounting system.

• Build: 9,000 dollars, three weeks.
• Model spend: about 60 dollars a month, cached and routed.
• Hosting and monitoring: about 120 dollars a month.
• Maintenance: about 1,400 dollars a year.
• Replaced: roughly 60 hours a month of manual entry and checking, at a loaded cost well above the run cost.

Payback in a few months, and the more durable benefit is the exception rate dropping because validation happens on every invoice instead of on the ones someone had time for.

Where money gets wasted

• Automating the wrong process. A workflow that runs twelve times a month cannot repay a serious build, no matter how annoying it is.
• Paying for discovery instead of software. Discovery is days, not months, and it should end in a scoped build.
• Per-seat AI tools nobody adopted. Check usage at 60 days before renewing anything.
• Skipping evals to save a week. The cheapest false economy in this field. You pay it back with interest the first time quality drifts silently.

How to compare two quotes properly

1. Whose accounts does it run in. Yours is the only good answer.
2. What is the expected cost per run, and what caps it.
3. What is the quality signal after launch.
4. What does the handoff include, and who owns it in month three.
5. What happens to the system if the engagement ends.

A proposal that answers all five is usually the more expensive one on paper and the cheaper one over two years. The deliverable standards behind those questions are in what an AI automation agency actually delivers.

If you want a specific number for a specific process, the AI production readiness audit produces a ranked plan with cost estimates in 48 hours, and the fee credits toward the build. Scope and approach are on the services page.

Frequently asked