What AI Automation Actually Costs in 2026

"How much does AI automation cost" is one of the most-asked questions in this category and one of the worst-answered, because most of the answers come from firms whose pricing page says "contact us."

So here are actual numbers, including ours, plus the parts of the bill that do not show up in a proposal.

The Three Pricing Models

Almost every automation engagement is one of three shapes.

Scoped project (sprint)

One defined deliverable, fixed price, fixed end. From $9,999, typically 4 to 8 weeks.

This is the right shape when you know what you want built and it has a clear boundary. It is the wrong shape when the requirement is "make our operations better," because a fixed scope against a vague goal produces something that technically satisfies the contract and helps nobody.

Retainer

Ongoing capacity. Ours runs $6,499 per month on a 6 month engagement, or $7,999 per month month-to-month with no cancellation penalty. That covers all five service pillars rather than automation alone.

The retainer is right when automation is a program rather than a project: multiple processes, discovered in sequence, with the second one informed by the first. It is wrong if you have exactly one thing to build.

Per-workflow

Some shops price per automation, roughly $1,500 to $7,000 depending on complexity. Rough bands from what we see in the market:

Per-workflow pricing is honest and easy to compare. Its weakness is that it prices the build and ignores the thing that actually determines whether you got value, which is the next section.

The Line Nobody Budgets

Maintenance is not optional and it is not free. Expect somewhere between 15% and 25% of the build cost annually, or a support retainer in the $250 to $500 per month range per significant integration.

People hear that and assume it is padding. It is not. Here is what it pays for:

• API deprecations. Vendors retire endpoints on their schedule. You get an email you did not read, and six weeks later a workflow stops.
• Schema drift. Someone adds a required field to a form. Your automation starts writing incomplete records and reports success while doing it.
• Auth expiry. OAuth tokens expire, get revoked when an employee leaves, or break when someone changes a password.
• Volume changes. The design that worked at 200 records a day behaves differently at 5,000.
• Business change. The process you automated in March is not the process being run in September.

The failure mode is specific and expensive: the automation does not error, it degrades. It keeps running, keeps reporting success, and produces subtly wrong output for weeks. By the time someone notices, you are doing data archaeology.

An automation with no maintenance budget is a one-year asset being sold to you as a permanent one.

What Should Change the Price

Legitimate reasons for a quote to move:

• Number of systems touched. Each integration is its own auth, rate limits, error semantics, and edge cases. Three systems is more than three times the work of one.
• Whether an API exists. If the answer is "we will scrape it" or "there is a CSV export," the cost and the fragility both go up sharply.
• Data volume and cleanliness. Dirty data is the single most common reason a project runs long. Nobody knows how dirty their data is until someone tries to automate against it.
• Compliance surface. Anything touching payments, health data, or consented communications carries real requirements. Our own SMS consent handling is a good example of why that is not a checkbox.
• Whether AI is genuinely needed. Model-based steps need evaluation, guardrails, and a fallback. That is real engineering, not a prompt.

What Should Not Change the Price

• The platform. Building the same logic in n8n, Make, or code differs in effort by less than most people assume. If a quote doubles because you asked for a specific tool, ask why.
• "AI" as a descriptor. A workflow does not cost more because a model is somewhere in it. It costs more if the model's output feeds a consequential decision and therefore needs validation. Those are different claims and only the second one justifies the number.
• Urgency, past a point. A rush fee is fair. A rush fee that doubles the project is a bet that you will not shop around.

Doing the Arithmetic Yourself

Before you take any quote seriously, work out what the manual process costs today. The workflow automation ROI calculator does this properly, but the back-of-envelope version is:

(hours per week) x (fully loaded hourly cost) x 52 = annual cost of the manual process

Fully loaded means salary plus employment costs, not the hourly rate on the offer letter. Then compare that to build plus one year of maintenance.

Two things fall out of this that people find uncomfortable:

Some processes should not be automated. If a task takes 20 minutes a week, that is roughly 17 hours a year. At $40 fully loaded, $700. A $3,000 automation does not pay back inside three years, and the process will have changed by then. Leave it alone.

The best candidates are rarely the most annoying ones. The process people complain about loudest is often small and infrequent. The one worth automating is usually boring, high-volume, and so normalized that nobody thinks to mention it.

A Reasonable Way To Buy

1. Map one process end to end before requesting a quote. A vendor pricing against a vague description will price the risk, and you pay for that.
2. Ask what happens in month 13. If the answer does not include monitoring and a named owner, the price is not the real price.
3. Start with one process in production rather than a department-wide program. The second build should be informed by the first.
4. Ask which steps will be deterministic and which will use a model, and why. A shop that cannot answer that crisply is going to use AI where an if statement belongs.

For what a scoped engagement covers on our side, pricing is published rather than gated. If you want the mapping done first, that is what the strategy session is for, and it includes telling you when the honest answer is that a process is not worth automating.