Last month a client sent me a one-line brief:
"Can we have an AI agent handle invoices, CRM updates and the weekly report? Budget is small."
I've been writing software for 25+ years, so I know that sentence. It's the start of a project that is either a weekend script or a six-month integration, and the budget alone doesn't tell you which.
The tricky part is that both answers are partly true. AI really has made some parts of the job dramatically faster. It has barely touched others. And most estimates I see treat the whole project as one lump with one discount.
So I did what developers do when they get tired of repeating themselves: I wrote the explanation as a tool. It's called the AI Business Process Automation Budget Calculator, it's free, it runs entirely in your browser, and it's deliberately a bit cheeky.
This post is the "how the model thinks" version. If you estimate work for clients, or you're the one receiving estimates, I think the structure is more useful than any specific number.
The core idea: split the work into two piles
If you look at an automation project honestly, the hours fall into two piles.
Pile 1: things you type. Glue code, notifications, dashboards, boilerplate. AI assistants are excellent here. In the calculator's model these get 25-35% off.
Pile 2: things you negotiate and verify. Understanding how the process is really done (not how the wiki says it's done), connecting old systems, privacy, security. Here the model takes off only 8-15%, because this work is conversations, access rights and testing with real cases. There's nothing to autocomplete.
Then a flat +15% goes on top for testing, QA and project management. Yes, flat. No, you can't negotiate it away. (Well, you can. You'll just pay it later, with interest.)
If your plan applies one "AI makes it 40% faster" multiplier to everything, it's pile 1 maths applied to pile 2 work. That's usually where the surprise invoice comes from.
The slider that matters most
The calculator asks how many systems the process touches. In my experience that's the single most predictive input, more than the type of automation.
Every extra system adds:
Access to arrange, and a human who has to approve it
Data mapping between whatever formats each side uses
Limits to respect: rate limits, quotas, maintenance windows
Failure handling: what happens at 3 a.m. when the other side times out
And then there's the special category: systems without a proper API. Lots of older accounting and ERP software lives here. If your process touches one, expect the number to react. Anyone who has automated a screen-scraped ERP knows why.
The 97% problem
This one is less about cost and more about design.
A model that's right 97% of the time is wrong 3 times in 100. In a demo that's invisible. In a business process, one of those three can be a payment or a contract.
That's why the calculator's "what has to be in place" list includes a pilot and approval steps, not just the automation itself. The pattern I trust:
Run the automation alongside the manual process.
Let a person approve the risky cases.
Expand what it does alone only as the track record builds.
It's the same logic as shipping behind a feature flag with a human reviewing the diffs. Boring, and it works.
The "could I do it with AI alone?" quiz
There's also a six-question reality check. It scores your experience and applies a skill level to a "DIY" estimate: the less you know, the longer the same work takes (up to 4x a specialist's AI-assisted time), and the more likely something breaks in production.
I want to be clear about what this is: a joke with maths in it, not a prediction. I'm all for people building their own automations with AI, and plenty will do it well. The point of the quiz is that the typing got cheap, while knowing what to test and what can go wrong did not.
Try it with two numbers
Open the calculator, pick a process, then compare:
€500: at the lowest rate (€25/h) that's about 20 hours of work.
€15,000: at €60/h that's about 250 hours.
Watch what the right side says about what each budget actually buys. Then change the hourly rate and the number of systems and see which input moves things most. (Spoiler: it's usually the systems.)
It's a rough model, not a quote. Real cost depends on your systems and your exceptions. If you want a real number, tell us about the process.
There are also sibling calculators if you're scoping something else: website, SaaS, web automation.
Over to you
I'd love a sanity check from people who've shipped this stuff:
Is a 25-35% saving on glue code about right for you, or are you seeing more?
What's the nastiest legacy system you've had to connect an automation to?
Where do you put the human in the loop?
Drop your answers in the comments. I read all of them, and the best ones tend to change the model.
Top comments (0)