How much does AI automation cost?

AI automation is quoted badly almost everywhere, because it has two costs and most proposals only mention one. There is the cost to build the thing, which is a normal software number. Then there is the cost to run it every month, which scales with your usage and which nobody discovers until the second invoice. Here is what both actually look like.

Build cost: three bands

The band is set by how messy the input is and how expensive a mistake would be — not by how advanced the model sounds.

Note that the cheapest band is genuinely cheap. If someone quotes $25,000 to classify inbound emails, they are either solving a much harder version of the problem than you described or padding. Ask which.

The running cost nobody quotes

This is where AI automation differs from ordinary software. A traditional web application costs roughly the same to run whether it serves 100 requests a day or 10,000. An LLM-based automation does not — you pay per token, every time, forever.

• Model API usage — the dominant cost, and the one that scales directly with volume. A well-designed automation handling a few thousand items a month typically runs $20–$200; a poorly designed one doing the same work can run ten times that.
• Hosting and infrastructure — usually $20–$100 a month for the service itself, queues, and a database.
• Monitoring and error tracking — often free at low volume, $20–$50 a month once you want retention and alerting.
• Maintenance — the honest line item. Models get deprecated, third-party APIs change, and your process evolves. Budget a few hours a month or the automation quietly rots.

The gap between a well-designed and a badly designed automation is almost entirely in this column. Sending a 40,000-token document to a frontier model when a 2,000-token extract and a smaller model would do is the single most common way running costs get out of hand.

Ask for a cost-per-item estimate before anyone writes code. If a supplier cannot tell you roughly what one processed document will cost to run, they have not designed the system — they have described it.

What actually drives the number

1. How consistent your input is

Three hundred invoices in one template is a small project. Three hundred invoices in eighty layouts, some scanned at an angle, is a different one. Input variety is the biggest multiplier in AI automation, and it is the thing clients most often describe optimistically.

2. The cost of being wrong

Drafting an internal summary and issuing a customer refund are not the same risk. The moment a mistake has money or legal consequences attached, you need validation, human approval, and audit logging — and those are where the engineering hours go. Price the consequence, not the task.

3. How many systems it has to touch

The model is rarely the hard part. Getting reliable data out of your CRM, writing results back without creating duplicates, and handling the case where the third system is down for twenty minutes — that is the work. Each integration is a real line item.

4. Whether anyone can tell if it is working

An automation without an evaluation set will degrade and nobody will notice until a customer complains. Building a labelled set of real examples and a harness that scores changes against it adds cost up front and is the difference between a system you can improve and one you can only hope about.

Which processes are actually worth automating

Multiply the time a task takes by how often it happens by a loaded hourly cost. If the annual figure is not several times the build cost, automate something else first. Beyond that arithmetic, the processes that work best share four traits:

1. 01It happens often enough to matter — daily or weekly, not twice a quarter.
2. 02The rules are stable. Automating a process you are about to redesign wastes the build.
3. 03Someone can tell quickly whether an output is right or wrong. If correctness is contested, no automation will settle it.
4. 04A wrong answer is recoverable. Start where a mistake costs an apology, not a lawsuit.

And one honest caveat: a good share of what gets pitched as AI automation does not need a model at all. If the rules fit in a lookup table, a lookup table is cheaper, faster, and correct every time. Any supplier worth hiring will tell you that before selling you the alternative.

How to keep the cost down

• Automate one process properly before automating five badly. The first one teaches you what your data is really like.
• Route by difficulty — send the easy 80% to a small, cheap model and reserve the frontier model for the rest.
• Cache aggressively. Identical inputs should never be paid for twice.
• Keep the human in the loop at first. Reviewing output is cheap, and it builds the labelled evaluation set you will need anyway.
• Extract before you send. Trimming a document to the relevant section often cuts token cost by an order of magnitude with no quality loss.

What we'd tell you on a call

Start with one process, budget $2,000–$6,000 for the build, and expect to spend under $100 a month running it at typical small-business volume. Insist on a cost-per-item number before the build starts and on monitoring that tells you when quality drifts. If the arithmetic on your first process does not clearly work, the answer is not a bigger model — it is a different process.