Ask five vendors what an automation project costs and the quotes will differ by a factor of ten. This is usually taken as proof that someone is lying. It is closer to the truth to say that the five vendors are quoting five different objects — and that the buyer's real job is to work out which object each number describes. This piece breaks the price apart so that comparison becomes possible. It ends with the five variables that set your number, because a market this opaque does not need another article that ends in "it depends."
The three costs, and why just one appears in the quote
Every automation carries three costs. Most proposals show one.
The build cost
is the visible one: mapping the process, engineering the workflow, integrating with the systems of record, testing, go-live. It is a one-time figure and it is what the quote usually contains.
The run cost
is the recurring one: software licences, model usage, hosting, monitoring, and the maintenance that keeps an automation working when an API changes or an invoice format shifts. A system with no maintenance plan is not cheaper; its run cost has simply been scheduled as a surprise.
The internal cost
is the one nobody quotes: the hours your own people spend documenting the process, answering questions, testing edge cases, and adapting to the new workflow. It is real, it is unavoidable, and a vendor who claims your team will need to invest no time is describing a project that will not be integrated into anything.
A quote that looks dramatically cheaper than the others has usually excluded the second cost, assumed away the third, or scoped a smaller object than you think you are buying.
Why the cheapest quote is usually the most expensive number
The figure that decides whether automation was worth it is not the price of the build. It is the total cost of a working system — build plus run plus internal time — divided by what the system measurably returns. A build that never reaches production costs infinitely more per unit of value than one at five times the price that removes thirty hours of manual work a month, permanently.
The research record supports being suspicious of cheap-and-fast. MIT's 2025 study of enterprise AI deployments found that roughly 95% of pilots produced no measurable P&L impact — and that the failures clustered around tools bought on demo appeal, without integration into the systems where work actually happens. The money wasted on those pilots was not lost to overpriced vendors. It was lost to underpriced promises.
The inverse also deserves stating: paying more does not buy certainty either. The same study found more than half of AI budgets flowing to visible, presentable use cases while the measurable returns sat in unglamorous back-office work. Price discipline means matching spend to measured value — in both directions.
Why nobody can quote you a number from a website
Published, independent benchmarks in this market are scarce and mostly produced by agencies marketing themselves, so treat any precise industry figure with caution — the flattering ones included. Self-reported return figures circulate widely: an IDC study commissioned by Microsoft in 2024 put average reported returns at $3.70 per dollar invested in generative AI. That is a survey of what adopters said about themselves. It should be read as sentiment, not as a promise your project inherits.
We publish no price list either, and the reason is the argument this article has been making. The price of an automation is a property of the process, not of the vendor: it follows from how many systems the process touches, how clean its inputs are and how often it changes, and none of that is knowable from a website. A number quoted before anyone has mapped the process is a guess with a currency symbol attached — and quoting one would make us the kind of supplier the first half of this piece warns about.
What we commit to instead is the sequence. Every engagement begins with a paid process audit whose price is fixed and agreed before it starts, and the audit is what produces a fixed number for the build — so that the larger figure stands on a mapped process and a written success criterion rather than on optimism. You see both numbers before you commit to anything. We would rather lose a project at the audit stage than win one that joins the 95%.
The five questions that determine your specific price
If you want to estimate a project's cost before talking to anyone, these five variables carry most of the weight:
How many systems does the process touch?
Each integration point adds engineering. One inbox to one ERP is a different project from inbox, ERP, CRM, and a bank feed.
How structured is the input?
Clean data in consistent formats is cheap to process. Unstructured documents in fifteen supplier formats are what modern tools handle well — but handling them still costs more to build and test.
What happens when the system is unsure?
A workflow that routes uncertain cases to a human for approval is cheaper to build safely than one that must act autonomously — and for most processes it is also the right design.
How often does the process change?
A stable process amortises its build cost over years. A process that changes quarterly needs a maintenance budget written into the contract, not discovered after it.
What does the current process cost?
This is the variable buyers most often cannot answer, and it is the single variable that turns a price into a decision. Eighteen minutes per invoice at current volumes is a number; against it, any quote becomes evaluable. Without it, every quote is just a feeling.
Frequently asked questions
How much does AI automation cost for a small business? There is no honest single figure, and a supplier who gives you one before seeing the process is quoting a guess. The build cost follows from how many systems the process touches and how structured the inputs are; the recurring cost follows from volume and complexity. That is why credible providers price after a process audit, not before.
What is the ongoing cost of automation after the build? Expect licences, model usage, hosting, and maintenance. A reasonable rule: if a proposal contains no recurring line at all, the maintenance has been left for you to discover later.
Is it cheaper to build automation in-house? Sometimes on paper, rarely in total. MIT's 2025 research found purely internal builds succeeding at roughly a third of the rate of externally partnered ones — and a failed cheap build costs more per unit of value than a successful expensive one. See our build-versus-partner analysis for the full argument.
How do I know if a quote is too good to be true? Ask what the written success criterion is and what happens if it is missed. A quote without a measurable outcome attached is a price for effort, not for a result.
Sources: MIT NANDA, "The GenAI Divide: State of AI in Business 2025" (2025); IDC, "The Business Opportunity of AI," commissioned by Microsoft (2024), self-reported figures.
