What it costs to put an AI tool into production.
A demo takes a few days. A system your business actually runs on takes months. Most of the money and nearly all of the difficulty sits in that gap, and almost nobody will tell you what it costs before you have signed something. Here is one of ours, with the numbers.
The job
A B2B distributor was capturing orders by hand. Orders arrived in whatever format the customer felt like sending. A PDF from one, a photograph of a written order from another, a WhatsApp message, an email with the detail buried in the body. Someone in the office read each one, found the right products and the right customer in the ERP, and typed it in.
Thirty minutes per order. People make mistakes when they are copying part codes, so some of those orders went out wrong and were fixed later at a cost nobody was counting.
The system now reads any of those formats, matches the products and the customer against the ERP, and produces the order. Twenty seconds. No typing.
What it cost
Between R100,000 and R300,000 to build.
The range is wide and the width is the honest part. What moves a job towards the top of it:
- How many different input formats you actually receive, and how bad the worst ones are.
- How reachable your ERP is. A system with a decent API is a different job to one that was never meant to be integrated with.
- How clean your product and customer data is. If the same product exists three times under three codes, that gets sorted out before anything else works.
Where the money went
Not where people expect. The model is the cheap part.
- Choosing and standing up a local language model, meaning one that runs on infrastructure the client controls rather than sending order data to an outside API.
- Building the harness around it. This is the part that takes the messy input, hands it to the model, checks what comes back, and talks to the ERP.
- Configuration against the client's real product and customer data.
- Tuning on their actual order formats, so it handles the specific ways their customers write things down.
What it costs to run
This depends on what you have built. But there is one thing worth settling before you choose an approach.
If the system calls an outside AI service, you pay per use. Every order costs you something, and the bill grows as the business grows. If the model runs on your own infrastructure, there is no per-use charge. What you pay for is the infrastructure it sits on, which is fixed, and it does not move when your order volume doubles.
That is the difference between a running cost that scales with your success and one that does not.
How long it takes
You do not flip a switch. The demo runs for a week or two first. Then it goes out in beta on real orders. Then it is simply how orders get captured.
Work out your own number
The saving is twenty-nine minutes and forty seconds per order. What that is worth depends on your volume.
| Orders per day | Hours released per month |
|---|---|
| 10 | about 99 |
| 25 | about 247 |
| 50 | about 494 |
| 100 | about 989 |
Twenty working days a month.
What actually pays for it
Not headcount. The people capturing orders are the people who know the customers, and they are now free to sell to them. The return is admin capacity turned into sales capacity, plus the orders that no longer go out wrong.
If the plan is to remove those people instead, the numbers work differently and probably worse. This pays best in businesses with more demand than they can currently reach.
Four questions before you fund a pilot
- Can it reach your live data, or is the demo running on a spreadsheet someone exported? This is the most common reason a working demo never ships.
- Who owns it on the Monday after go-live, and does that person exist yet?
- What happens when it is wrong, and who finds out?
- What does it cost to run each month, and does that number move with your volume?
If those answers are not clear before you fund the pilot, they will not get clearer afterwards.
We ask these four before we start. It is the reason every pilot we have taken on has gone into production.