Questions to Ask Before Buying an AI Solution

What problem, measured how?

Start with the problem, not the product. A useful AI purchase replaces or speeds up a specific, repeated task: answering the same ten customer questions, sorting incoming email, drafting quotes from a form, taking bookings on the phone. If the vendor cannot name the task in one sentence, or you cannot, the project will drift into a general "AI capability" that nobody owns and nobody can evaluate.

• Which task, exactly, and who does it today?
• How many times a week does it happen, and how long does each take? Count it for two weeks before the first meeting.
• What does a good outcome look like: time saved, faster answers, fewer missed enquiries, fewer errors?
• Who in your company owns the result after launch?
• What would you do instead if you did not buy this? That alternative is the real price comparison.

Data handling and training

This is where Swedish buyers carry real legal risk. If the solution handles customer names, emails, phone calls or case details, the vendor is almost certainly your personal data processor under GDPR, and you need a data processing agreement (personuppgiftsbiträdesavtal) before any real data goes in. IMY expects you, as controller, to know where data goes and on what legal basis.

If the tool will be used by your own staff rather than customers, the same questions apply to what employees paste into it. Our guide to a staff AI usage policy covers that side.

Accuracy, testing and fallbacks

Language models produce fluent answers whether or not they are correct. That is a property of the technology, not a flaw a vendor can promise away. The honest questions are therefore about containment: what the system is allowed to answer, what it does when it is unsure and how you find out when it went wrong.

• What sources does it answer from, and can it answer outside them? A good setup restricts answers to your approved content.
• What does it say when it does not know? The right answer is a handover to a person, not a guess.
• Can we test it on our own real questions before we sign or pay in full? Ask for a trial set of your actual enquiries, including awkward ones.
• Where do we see conversations that failed or were escalated, and how often will someone review them?
• Who fixes wrong answers, how fast, and is that included in the monthly fee?

The mechanism behind invented answers, and how to reduce them, is explained in stopping a chatbot making things up. If a vendor tells you their system "doesn't hallucinate", treat everything else they say with the same scepticism.

Running costs over time

AI products are cheap to start and can be expensive to run, because the cost follows usage: messages, minutes, documents processed, seats. A price that works for your first month can double when a campaign works or when staff start using the tool for tasks nobody planned.

Ask for the total over 24 months at your expected volume and at twice that. Compare it with the alternative you wrote down in step one. Our pricing page lists fixed setup and monthly fees for the AI services we build, so you can use it as one reference point; for budgeting a first project in general, see what a first AI project should cost.

Exit: your data and prompts

The most expensive question to skip is how you leave. AI solutions accumulate value in things that are easy to lose: the prompts and instructions tuned to your business, the knowledge base, the escalation rules, the conversation history that shows what customers actually ask. If those stay locked in a vendor's platform, switching means starting over.

• Can we export conversation logs, knowledge content and configuration in a standard format?
• Who owns the prompts, instructions and custom workflows written for us?
• What is the notice period, and is there a minimum term?
• How is our data deleted at the end, and do we get confirmation?
• If the vendor is acquired or shuts down, what happens to the service and the data?

Ownership terms deserve the same attention as in any software contract; our guide to IP clauses in software contracts explains what to look for.

When you do not need outside help

If you are buying a mainstream tool for internal use, such as an AI assistant built into software you already pay for, you can usually answer these questions yourself from the vendor's documentation and data processing terms in an afternoon. You do not need a consultant for that, and we would tell you so.

An independent review earns its cost when the purchase touches customers directly, handles sensitive data, needs integration with your own systems or involves a contract you cannot easily leave. Our AI integration audit is a fixed-price review of exactly that: which tasks are worth automating, which options fit, what they will cost to run and where the risks sit. It is also useful before you ask anyone, including us, for a quote. If you want to talk it through first, book a short call.

Frequently asked questions