1. Which channels do you answer today?
List them: email, a chat widget, phone, SMS, social messages. A tool that does not cover a channel you use is out, however good it is. Smaller tools often skip phone and social, and say so; larger suites cover everything and charge for it.
2. What does a seat really cost?
Compare per-agent price on the billing interval you will actually use, and ask which features sit in higher tiers. Check minimum seat counts, whether read-only teammates need a paid seat, and whether annual billing is required to get the headline price.
3. How is AI billed?
This now changes the bill more than the seat price does. Ask what unit is counted, whether an allowance is included, what happens at the limit, and whether you can see each charged conversation. The guide on AI resolution pricing explains the units.
4. Can you bring your history, and take it back out?
Ask what an import covers (tickets, attachments, macros, rules, help articles) and what it does not. Then ask the exit question: can you download everything, attachments included, without asking support?
5. What does it say it cannot do?
A vendor that lists its gaps plainly is easier to trust than one that lists none. Look for a security page that states what is not in place yet, such as a SOC 2 report, an uptime commitment or a penetration test, and check that against what your customers require of you.
6. How good is the AI on your tickets?
Do not rely on a demo. Run it on a sample of your own past tickets. The guide on testing AI support has a fifty-ticket method that works for any vendor.
7. Who do you talk to when it breaks?
Find out whether support is a person or a queue, what the response times are, and where outages are posted. For a new vendor, ask who runs it and how long it has been selling to paying teams.
8. What will the setup cost you in time?
Count the steps: forwarding email, adding DNS records, installing chat, rebuilding rules, training the team. Tools built for small teams can take an afternoon; suites built for large ones can take weeks and a consultant.
Run a real trial
Pick two finalists, import or forward real tickets, and have the team use each for a week. Ask them afterwards which one they would be annoyed to lose. That answer beats any feature matrix.