The questions people ask before they ask us

Less than people fear for the model, and more than they expect for everything around it. At small-business volumes the model calls cost little each month. What you pay for is an agent that answers only from what you actually know, hands over to a person when it should, and gets better from its own conversations. A narrow agent that answers customers from your own content, with an inbox where a person can step in, is a matter of weeks rather than months.

Yes, if you treat the AI tool like any other system that handles personal data: a lawful basis for the purpose, a data processing agreement with the provider, a clear answer to where the data is processed and whether it trains anything, and a line in your privacy notice. The usual mistake is customer data pasted into a free chat tool that nobody in the company has approved.

It depends on four things more than on the app itself: whether it has to be in the App Store or can live in the browser, whether people log in and pay, how many other systems it has to talk to, and who looks after it once it is live. A narrow first version that does one job well is typically a couple of months of work. The price follows what the app has to do, and the number of screens says little about that.

RAG, retrieval-augmented generation, means the model looks up the relevant parts of your own documents before it answers, and answers from those. You need it when the answers have to come from content that is too large to hand the model every time, or that changes often: manuals, contracts, product sheets, years of email. You do not need it for a one-page FAQ, which fits in the instructions as it is. The hard part is keeping the documents current, and making sure nobody finds what they are not allowed to see.

There is no single number, but there are four questions that decide it: who pays, what happens after the payment, how many people share the same calendar, and whether something physical has to happen, such as a door opening or a heater switching on. A system that only accepts a time slot is a few weeks of work. One that takes payment, tracks memberships and lets a guest into a warm room is a different product, and that difference is what the price is about.

For most Norwegian online services the answer is both, and the real question is a different one: which system owns the truth about what the customer bought. Vipps MobilePay is how Norwegians actually want to pay and belongs at the checkout. Stripe is the toolkit for subscriptions, invoicing and payment methods beyond the Nordics, Klarna included. Decide who holds the truth before you pick a provider, and the rest becomes an integration instead of a rewrite.

Three layers, not one: the equipment that does the work, the connection that goes away sometimes, and the cloud that decides what should happen. The integration itself (getting a value out and a command in) is usually the smallest part. The work is deciding which logic has to run without a network, what the system does when the equipment does not answer, and how somebody sees it has stopped before the customer calls.

What works today is narrow jobs with an answer somebody can check: a draft a human approves, information pulled out of a mess and into a form, an enquiry routed to the right place, a question answered from your own documents. What does not work yet is the agent that “runs the business” unsupervised. And the most common failure is the system around the model throwing away an answer that was right.