Short answer
A custom AI build is priced by the work, not by the month. The price is driven by four things: how clean your data is, how many systems it must connect to, how wrong it is allowed to be, and who maintains it afterwards. Everything else is detail.
Why per project, not per month
Subscription pricing suits products. Kateb is a product, so Kateb has a subscription. Custom work is not a product. It is a defined piece of work with an end. Charging monthly for something that finishes creates a bad incentive on both sides: we benefit from the work never quite being done, and you pay indefinitely for something you already own.
So a custom build is scoped, quoted, delivered, and handed over. If you want ongoing changes afterwards, that is a separate, explicit agreement, not a default that quietly renews.
The four things that actually move the price
1. The state of your data
This is the single biggest variable, and it is almost never the one clients expect. If your records live in one system with consistent fields, the model work is straightforward. If they live across a spreadsheet, a WhatsApp thread, and a notebook, and the same customer is written three different ways, then most of the project is not AI at all. It is cleaning. We cover this in why your data is the hard part.
2. How many systems it has to touch
A tool that stands alone is cheap. A tool that must read from your stock system, write to your invoicing, and notify someone on WhatsApp is three integrations, each with its own failure modes. Integrations are where estimates go wrong, because they depend on systems we do not control.
3. Your tolerance for being wrong
This is the cost driver nobody discusses up front. A chatbot that answers opening hours can be wrong occasionally and nothing breaks. A tool that drafts a legal document cannot. Lowering the acceptable error rate does not cost a little more. It changes the architecture. You need retrieval, verification, an audit trail, and a way for the system to refuse rather than guess. See stopping a model from inventing a citation.
4. Who maintains it
Models change, APIs change, your business changes. Someone has to own the thing after launch. If that is us, it is priced. If that is you, we build it so a competent developer can pick it up, and we say so in writing.
How we scope
Before quoting anything, we want three answers:
- What does the person do today? Not the ideal process, but the real one, including the parts done by hand.
- What does a wrong answer cost? Embarrassment, a lost sale, or a legal problem. This sets the engineering bar.
- How would you know it worked? If nobody can name a measurable outcome, the project has no finish line.
If those three cannot be answered, we would rather delay the quote than invent one.
When the honest answer is no
Sometimes the scoping conversation makes it clear that AI is the wrong tool. A well-designed form, a shared spreadsheet, or a phone number often solves the actual problem for a fraction of the cost. We would rather say that than sell a build that will not survive contact with your workflow. That case is made in full in when you do not need AI.
Frequently asked questions
Do you charge a monthly fee for custom AI work?
No. Custom builds are scoped and quoted as a defined piece of work. Ongoing maintenance, if you want it, is agreed separately rather than bundled into a subscription that renews by default.
What makes an AI project more expensive?
Four things: messy or scattered data, the number of existing systems it must integrate with, how low the acceptable error rate is, and whether we maintain it after launch. Data quality is usually the largest single factor.
Can you quote before seeing our data?
Only roughly. Data condition is the biggest cost driver, so a quote given without looking at it is a guess. We would rather look first and give you a number we can stand behind.