The AI chatbot market is full of vendors describing themselves with the same words: smart, simple, powerful. To actually compare them you need to go deeper than the sales copy. This checklist gives ten questions, what a good answer looks like and what should make you pause.
The ten questions
| Question | A good answer | A warning sign |
|---|---|---|
| 1. Where is data stored? | A concrete answer and a written data processing agreement | General wording about "secure cloud storage" |
| 2. How good is the Swedish? | The vendor lets you test with your own questions | Only ready-made demo questions |
| 3. What is in the price? | A clear list of channels, connections and usage | Add-ons that appear after signing |
| 4. What does it cost at our volume? | Scenarios for a quiet, normal and busy month | Only the vendor's default example |
| 5. How does handoff work? | Clear rules and context that travels with the case | No route out to a person |
| 6. Which connections exist for real? | An answer on what is available for your platform and what is set up for you | Long logo lists without detail |
| 7. How do we check what the AI says? | You can read conversations and fix sources | A black box with no visibility |
| 8. How do we test before launch? | A pilot or test environment with real questions | Immediate launch to every customer |
| 9. What if we want to switch? | A clear notice period and a way to export knowledge | Lock-in or unclear terms |
| 10. Who helps us? | A named contact route and onboarding | Only a support email address |
Why each question matters
Data and contract. Require a written data processing agreement and a clear answer on data storage in the EU. Also read which subprocessors are used. Practical rules are in GDPR and data storage and common misunderstandings in AI and GDPR. Assess the legal requirements with qualified advice.
Swedish in practice. A demo in English says little about how Swedish performs. Test with your own questions, with dialect, abbreviations, typos and the way customers phrase things carelessly. That is where the differences show.
Price at your volume. Calculate several scenarios. A model that is cheap at low volume can be expensive at high volume, and the other way round. Use the pricing guide as a basis and compare against staffing.
Connections. Many vendors list dozens of integrations. Ask which are available for your platform today, which the vendor's team sets up during onboarding and which are planned. There is a big difference between a logo on a page and a working connection.
Handoff and control. A good vendor lets you see what the AI has said and fix the source when something goes wrong. Read when the AI should escalate to a human and what a handoff should include.
How to test before you sign
- Collect fifty real questions from email, phone and chat. Include the unclear and the difficult ones.
- Write down the answer you expect, or the handoff you expect.
- Run the questions through the vendor's demo or pilot.
- Mark wrong facts, invented promises and missed handoffs.
- Ask the vendor to explain each error and how it gets fixed.
How to test an AI agent before you launch it has a full test plan you can reuse.
Compare two or three vendors the same way
Use the table above as a scorecard. Give each question a score from one to three for each vendor and write down the reason. Put the most weight on data protection, handoff and price at your volume, because those are the most expensive to get wrong. See also comparisons between Aveexia and other alternatives if you are weighing several systems.
If you want to see how Aveexia answers these questions directly, there is a longer walkthrough in the guide to choosing a vendor, and you can book a demo.
