Ask what automation costs and you will usually get "it depends", which is true and useless. The reason nobody gives a number is that the same sentence, "we want to automate our invoicing", can mean two days of work or two months. What follows is what actually moves the price, so you can estimate your own job before you speak to anyone.
The four things that set the price
Almost all of the variation comes down to the same four questions. None of them are about the AI.
1. How many systems have to talk to each other
One system is cheap. Two is fine. Four, where one of them is a twenty-year-old package with no API and a CSV export, is where the cost lives. Every connection is a place things can break, and every one needs handling when it does.
2. How consistent your inputs are
If every invoice arrives as a PDF from the same supplier in the same layout, that is straightforward. If they arrive as photos, scans, emails and the occasional fax, each variation is more work. This is usually the single biggest surprise in a quote.
3. What happens when it gets something wrong
Automating an internal report that a human reads is low risk. Automating something that sends money, or emails a customer, needs approval steps, audit trails and a way to unwind a mistake. That safety work is often larger than the automation itself, and it is not optional.
4. Whether anyone has written the process down
A process that lives in one person's head has to be extracted before it can be built. That is discovery time, and it is real. If you can already describe the job step by step, you have saved money before anyone starts.
Rough shapes
We quote every project individually because of the above, but jobs tend to fall into recognisable shapes:
- A single, well-defined task with one or two systems and predictable inputs. Days rather than weeks.
- Several steps joined into one flow, with a couple of integrations and some human approval. Two to four weeks is typical.
- A company-wide system touching many tools, with reporting and a staged rollout. Longer, and usually better split into phases that each pay for themselves.
A good supplier will tell you when the third option should be the first option done three times. Phasing costs slightly more in total and dramatically reduces the risk of paying for something that turns out to be wrong.
Working out whether it pays
The arithmetic is simpler than people expect. Take the task, estimate honestly how many hours a week it consumes across everyone who touches it, and multiply by what that time costs you. Compare that against a one-off build plus any ongoing running cost.
Two things people forget. First, count the rework: the invoices keyed wrong, the quote sent to the wrong address, the chase that never happened. That is usually worth more than the typing time. Second, count the jobs you turn down because nobody has capacity, if that applies.
And be willing to reach the opposite conclusion. Something that takes ten minutes a fortnight is not worth automating, however annoying it is. The jobs worth doing are the small, dull, daily ones.
Costs that show up after the build
A build price is not the whole picture, and anyone who does not mention the following is either inexperienced or hoping you will not ask:
- Hosting and monitoring, so somebody notices when it stops.
- Model and API usage, if the work involves a paid AI provider. Usually small, but it is not zero.
- Maintenance, because the tools you connect to change their APIs whether you like it or not.
- Your own time during the build, mostly answering questions and testing against real cases.
We price the build as a fixed figure agreed before work starts, and hosting and support separately as an optional monthly plan, so you always know which part you own outright.
What to ask before you commit
- Is this a fixed price, or an estimate that can move? Get the answer in writing.
- What happens if it breaks in six months, and what does that cost?
- Who owns the finished thing, and can we run it without you?
- What does it do when it is not confident? Guessing is worse than stopping.
- Can I see it working against our real data before we go live?
The last one matters most. Anything can be made to work on a tidy example.