How much does AI automation cost for a small business in 2026?

Ask five vendors what an AI automation project costs and you will get five versions of "it depends", a request for a discovery call, and no number. The number exists; they just prefer to say it after they know your budget. This article does it the other way around.

The short answer, from my own published price list: AI automation projects start at $8,000 as a fixed quote. Advisory work, when you want an engineer's judgment rather than a build, is $150 an hour. Data and integration work, where the job is to connect two systems rather than to build an assistant, starts at $1,500 as a fixed scope. Independent testing of something someone else built starts at $3,000. In my quotes, most first automation projects for a small business land between the $8,000 floor and about $20,000 (an assistant over your own data can reach $25,000), and take two to six weeks. Everything is quoted in writing before work starts, and the number does not move unless the scope does.

The rest of this article is what sits behind those figures: what you are actually paying for, what pushes a quote up or down, where the costs hide, and how to get a fixed price out of anyone, including me.

Why almost nobody publishes a price

Three reasons, and only one of them is good. The good one: AI projects vary more than websites do, so a single number can mislead. The other two: an unpublished price can be set after the vendor has sized your budget, and a discovery call is a sales call with a nicer name.

The honest fix is not to hide the number but to publish the floor and explain what moves it. That is what the rest of this does. If a vendor cannot tell you what moves their price, they cannot tell you what you are buying.

What you are paying for

An automation project is mostly engineering time, and the time goes to four things. Each one is a question you can answer about your own project before you talk to anyone.

How many systems it touches

Every system the automation reads from or writes to is an integration: an API to learn, credentials to store safely, failure cases to handle when that system is down or changes. One system is a small project. Four systems, one of which has no API and only exports PDFs, is a different project with the same headline.

How clean the inputs are

Structured data in a database is cheap to work with. Spreadsheets are fine. PDFs, scanned documents, emails and free-text notes are where the hours go, because the automation has to extract meaning before it can act on it. A project that starts from a clean export costs a fraction of one that starts from a shared inbox.

How expensive a mistake is

This is the driver most buyers underweight. An automation that drafts internal summaries can be wrong occasionally and nobody is harmed. One that files a tax return, changes a price, or sends a customer an answer has to be right, which means evaluation cases with known-correct answers, tests that run on every change, and a human approval step in front of anything irreversible. On my IFTA filing agent that discipline is 562 automated tests and a backtest that matches a real state filing to the penny. It is the single biggest reason two projects that sound alike can be priced far apart.

What has to run unattended

A tool a person runs at their desk needs less than a job that runs at 2 a.m. with nobody watching: logging, retries, alerts, a status page, a backup. If the automation becomes part of how the business runs, that layer is not optional, and it is part of the price.

Three project shapes at the published prices

These are shapes, not quotes. The exact number comes from the four drivers above and arrives in writing.

Connect two systems

A nightly export from one platform lands in another, or a report that someone builds by hand every Monday builds itself. Mostly deterministic code, little or no language model. This is the data-and-integrations service: from $1,500 as a fixed scope. Typical range in my quotes: $1,500 to $4,000.

Automate a document-driven process

Intake forms, invoices, bills of lading or applications arrive as files; the automation extracts the fields, checks them against rules, routes the result and flags what a careful person would flag. This is where a model earns its place and where evaluation matters. This is the AI automation service: from $8,000 as a fixed quote. Typical range: $8,000 to $15,000.

An assistant over your own data

An agent that answers questions from your documents, your records or your systems, with citations, for your team or your customers. Retrieval, permissions, evaluation against real questions, and a human-approval path for anything it is allowed to change. Typical range: $12,000 to $25,000, because the evaluation layer is larger and the permissions matter.

If your project is bigger than the last shape, it is probably several projects, and the right first step is a paid scoping document rather than a build.

What the price does not include

Running costs are separate from the build, and you should ask every vendor to state them, because they are where a cheap build becomes an expensive habit.

• Model usage. The automation calls a model provider and pays per use. For most small-business workloads this is tens of dollars a month, not hundreds. I design so the model is swappable, so a provider's price change is a setting, not a rebuild.
• Hosting. A small server or a serverless host. Usually under $50 a month for the shapes above.
• Third-party APIs and licences, if the systems you connect to charge for access. Some do; find out before the scope is written, because a vendor's API fee can be the largest line.
• Changes after launch. Mine are quoted per change or at $90 an hour, with no retainer. A fixed price for the build plus a retainer you cannot leave is a structure to look at twice.

What is deliberately not on this list: a monthly platform fee for your own system, or a dependency on the builder to keep it running. You own the repository and the accounts at the end, and the code is written and documented so another developer can take it over.

The hidden costs, which are the real ones

The expensive AI project is rarely the one with the high quote. It is one of these.

The cheap build with no evaluation

A thin wrapper around a model API, no test cases, no checks. It demos well and is quietly wrong a few percent of the time, and you find out from a customer or an auditor. The repair costs more than doing it properly would have.

The prototype that became production

Built in a few weeks with an AI assistant, running from one person's laptop and one login, with nothing watching it. It works until the login expires at 6 a.m. Making it reliable is a bounded job, and I have written separately about the order to do it in, but it is a cost that was not in the original "free".

Lock-in

A system you cannot take elsewhere, hosted on the vendor's account, with your data in their store. The quote looks fine until you want to leave.

Scope that nobody wrote down

The project that was going to "automate the back office". Without a written scope that says what is excluded, every quote is a guess and every invoice is a surprise.

How to get a fixed price from anyone

Fixed pricing is possible when the scope is written. Here is what makes it possible on my side, and it works on any vendor.

• Describe the process as it happens today, step by step, including the exceptions. The exceptions are the project.
• Name the systems involved and whether each one has an API, an export, or only a screen.
• Show three real examples of the input, including an ugly one.
• Say what a mistake costs. This sets how much checking the system needs.
• Say what must run without a person, and when.
• Ask for the scope in writing, with exclusions, a delivery date, and the price, before any work starts.

A vendor who can turn those answers into a number has done this before. One who cannot is going to learn on your budget.

What a project here looks like

A free intro call of about 30 to 45 minutes, where you describe the problem and get an honest read on whether it is worth automating and roughly what it costs. If it is a fit, a written scope and a fixed price follow before any work begins. If it is not a fit, or a platform you already pay for would do the job, you hear that instead. The published prices are the floor; the written quote is the number.