Ask what an AI automation costs and you will usually get a build price.
That number matters. It is also only part of the answer.
An automation is not a purchase you make once. It is a small system you keep running. Its total cost has two halves: what it takes to build, and what it takes to keep it useful.
This guide does not give price ranges. The honest range for “an AI automation” is too wide to help anyone. It explains what drives the cost, so you can compare quotes properly and ask the right questions.
What goes into the build cost?
Discovery
Understanding the current process, its exceptions and its data. Skipping discovery makes the quote smaller and the project riskier. A good quote either includes it or says plainly that it assumes you have done it.
Integration
Connecting the systems involved. This is frequently the largest and least predictable part. Well-documented applications with good integrations are quick. Older systems, custom software and anything without an API can take far longer than expected.
Building the logic
The steps themselves: rules, AI prompts, routing, formatting, notifications.
Testing
Running the workflow against real past cases, including awkward ones, and fixing what fails. Testing is where a demo becomes something you can rely on. It should be visible in the quote.
Review controls
The points where a person checks or approves output before it matters. These cost a little to design and save a great deal later.
Handover and training
Documentation, credentials, and showing the people who use it how it works and what to do when it does not.
What goes into the running cost?
Platform and software subscriptions
Automation platforms, hosting and any applications added for the workflow. These often scale with usage or with the number of steps run.
AI usage
Most AI models are paid for by usage. Cost depends on how often the workflow runs and how much text, imagery or data each run processes. Providers also change their pricing and models over time.
Monitoring
Someone, or something, needs to notice when runs fail or output quality drifts. Silent failure is the most expensive kind.
Support and fixes
Applications update. Connections break. Models change behaviour. Budget for this, whether it is your own time or a support arrangement.
Change requests
Your business will change the process. New fields, new rules, a new tool. Small changes are normal; who makes them and how they are priced should be agreed up front.
How should you handle uncertainty in a quote?
Some uncertainty is genuine. Nobody knows how co-operative an old system will be until they try. A trustworthy quote names its uncertainty rather than hiding it.
- Ask which parts of the quote are fixed and which are estimates
- Ask what assumptions the quote depends on
- Ask what happens if an assumption turns out to be wrong
- Prefer a small paid discovery or test over a large fixed price built on guesses
A quote that is suspiciously precise about something nobody has investigated is not more reliable. It is less honest.
When is it worth paying more?
Paying more is usually worth it for:
- Proper discovery when the process is poorly understood
- Testing on real cases rather than a demonstration
- Review gates on anything customer-facing, financial or sensitive
- Clear documentation and handover
- A defined support arrangement
Paying more is usually not worth it for features nobody asked for, complexity that impresses in a demo, or automating a task that happens too rarely to matter.
Questions to ask before you sign
- What is included in the build price, and what is not?
- Which recurring costs will we pay, to whom, and how do they scale?
- Whose accounts do the subscriptions and AI usage sit in?
- How is the workflow monitored, and who is told when it fails?
- How are changes after launch priced?
- Are prices quoted including or excluding tax?
A fair price is one where you can see what you are paying for now and what you will keep paying for later.
MonoGrain quotes engagement pricing after a discovery conversation, so the scope is known before the number is. The published starting options are on our What we do page.