How to evaluate an AI customer messaging platform

Similar feature lists hide real differences in ownership, permissions, integrations and cost.

1. Who owns the WhatsApp account?

Ask this first. If the WhatsApp Business Account sits in your own Meta business, you see the number and templates in WhatsApp Manager and keep them. Meta says businesses onboarded through Embedded Signup own their WhatsApp assets.

Follow up:

• Can we see the account in our own Meta Business Manager today?
• Do we pay Meta directly, or do you resell messages?
• How do we export consent records, which the platform holds?
• What is the exit process, step by step?

2. What does "supports Instagram" mean?

"Supports" can mean a full two-way inbox or a form that posts a message somewhere. For each channel you need, ask: can we receive, reply and see delivery status? And what happens when the channel's rules block a send?

WhatsApp has its own 24-hour customer service window; Messenger and Instagram have a separate window with narrower exceptions.

3. What does the AI answer from?

Ask to see an answer with its sources, then a question your material does not cover. Then test customer-specific questions: can it read the right appointment, order or invoice, and what does it do when it cannot? The grounding guide has the full test.

4. What does it do on its own?

Drafting and acting are both valid; the boundary must be explicit and yours to set.

• What sends automatically, and can we turn it off per chat and for the whole business?
• What happens when information is missing: does it ask, pass the chat to a person, or guess?
• If it offers approvals, does an approval cover the exact text that sends?

5. How is consent handled?

Ask to see one contact's consent record: the wording shown, where and when it was collected, its purpose and every change, including opt-outs. Suppression should be checked at send time, not when the audience was built.

When Meta reports that a person stopped marketing messages, future marketing to them should stop automatically.

6. Can we see quality and template health?

Meta rates quality per number and per template, and can pause a template. Ask whether you see both, and whether a paused template shows up before the next campaign or only after it fails.

7. Does reporting show cost?

Meta includes billing details in message status webhooks. Ask for a campaign report that separates sends, deliveries, billable messages, estimated cost and invoiced cost. Failures should be broken out, because a throttle, a closed window and an opt-out each need a different response.

8. How is our data kept separate?

Ask how brand isolation works across storage, search, integrations and staff access, and for evidence that one business cannot reach another's records. Also ask who can access channel credentials, how access is revoked, and where data is stored.

9. What does the activity history record?

It should show every send and its delivery status. Otobiz keeps each send in the conversation history, so your team can trace any reply a customer quotes back.

10. What do the integrations actually do?

Name the two or three systems you need before looking at any integration list. For each, ask what is read, what is written and what stays in sync. Then ask for a customer-facing API and outbound webhooks.

11. What is the pricing unit?

Start with Meta's per-message rates for your markets; the pricing guide lists them. Then pin down the vendor's unit:

Price your own volume through each model. On Otobiz, paid plans from US$19 a month include unlimited messages, Free includes 1,000 messages a month, and every paid plan includes AI replies, then US$0.02 each if you choose. See Otobiz pricing.

12. How do we get our data out?

Ask before you sign: conversation history, contacts with consent records, and templates. In what format, by what mechanism, and how long it takes.

Tests to run before you commit

Use a test number and contacts you control, never real customers.

• Ask something your content does not cover. Watch whether it admits the gap or fills it.
• Ask something two of your documents answer differently. See which wins, and whether you can tell.
• Let a window close. Reply free-form the next day. A good product says the window closed and offers a template.
• Opt out, then send. Reply "stop" from a test number and try a broadcast. The contact should be excluded everywhere.
• Hand a thread to a colleague. Can they find the owner, last promise and next action on their own device?
• Message from two channels. Two people with the same name must not share a history.
• Read a real campaign report. Does it answer "what did this cost" and "why did these fail"?
• Break a template. In a draft, omit a required variable and check the feedback before any send.

Discount the headline numbers

• Integration count. Three that work beat ninety listed.
• AI feature count. Count useful outcomes, not labels.
• Deflection rate. Every vendor measures it differently, and it rises when reaching a person gets harder. Ask whether resolved and abandoned chats are told apart.
• Customer logos. A useful case study names the workflow, the period and the result.

Keep a scorecard

Mark each scenario demonstrated, partly demonstrated or not demonstrated, with the evidence.

Then change the underlying record and ask again: the answer should follow. To size the opportunity, try the missed-opportunity calculator with your own numbers.