Where the money goes in an AI system

Imagine two development quotes on an owner's desk for the same knowledge-base system, one twice the price of the other. The owner asks: what's the difference?

Most people compare the one-off development fee. But with an AI system, the development fee is only part of the first year's bill, and often not the largest part. Until you know where the money in an AI system goes, you can't really compare quotes.

The one-off money

Requirements and design, development, testing and launch, training. Of the four, the first is the easiest to squeeze, because nobody can see it: "Don't you already know what we need?"

Money saved there is usually paid back double later. Requirements that weren't understood mean rework, and rework after launch is the most expensive kind. If a quote treats requirements research thinly, that is probably where the saving comes from.

The money that runs every month

Once traditional software is built, most of the cost has been paid. An AI system is different: it spends money every time it's used.

• Model usage. Mostly charged by volume — the more you use, the more you pay. Different tasks suit different tiers of model, and sending simple sorting to the biggest model is like sending the managing director to do the photocopying. How well the models were chosen shows up directly on the monthly bill.
• Servers and databases. The system has to run somewhere, and the data has to live somewhere.
• Third-party services. Document recognition, speech-to-text, email sending — each may come with its own monthly invoice.

When you ask for a quote, ask whose name these accounts will be in, who receives the bills, and roughly what they come to each month. Whoever holds the accounts holds the cost of ever switching suppliers.

The money for maintenance

After launch, the world outside keeps moving: model providers release new versions, interfaces change, business rules are adjusted. Someone has to keep the system up to date.

The line quotes most often leave blurred is the one between a fix and a new requirement. If the system gets something wrong, that's a fix; if you want it to do something more, that's a new requirement. Leave that line undrawn before launch, and every conversation afterwards turns into a negotiation.

The easiest cost to miss: your people's time

This one never appears on a quote, and it is often the biggest.

A system needs people to tidy data, confirm rules and review what the AI produces. In the first few months after launch, how many hours a week will the responsible colleague spend? Who keeps the data current? Who signs off the reviews? If that time isn't planned for, it gets squeezed out of other work — and the system slowly ends up with nobody looking after it.

Five questions to ask when comparing quotes

1. How much of the effort goes into requirements research, and whom will you interview?
2. Roughly what will the running costs be each month, and whose name are the accounts in?
3. Which tier of model will you use, and why?
4. For how long after launch do errors count as fixes?
5. After launch, how much time will our own people need to put in each week?

The quote with clear answers isn't necessarily the cheaper one. But it is very likely the one closer to the real total cost.

A cheap system doesn't always save money. A system you can rely on for three years does.