💼 Cost & business

Twelve security questions to ask a chatbot platform

Updated 2026-09-24 · 5 min read · By the Convo360 team

In short

Ask about webhook signature verification, tenant isolation, where data is stored, who can read conversations, how credentials are held, and what happens to your data if you leave. A vendor who answers these specifically has done the work; one who answers in adjectives has not.

Key takeaways

Ask for demonstrations, not assurances. A refusal should be shown, not described.
Webhook signature verification is the single most-skipped control.
Tenant isolation failures are the most damaging and the least visible.
Your data exit route matters most on the day you want to leave.

Why the usual security page is not enough

Most vendor security pages say the same things: encryption in transit and at rest, regular audits, role-based access. These are table stakes and almost nobody lies about them, which makes them useless for telling vendors apart.

The questions that actually distinguish a platform are specific and slightly awkward, because they ask about the mistakes that are easy to make and invisible from outside.

You are not trying to catch anybody out. You are trying to find out whether somebody has thought carefully about the particular ways a messaging platform can fail, which is different from whether they have a certificate.

It is also worth asking them at the right moment. Before signing, a vendor answers carefully because they want the business. After signing, the same questions get a slower and vaguer response. Ten minutes in an evaluation call is worth more than an hour of chasing later.

Questions about the messaging layer

1. Do you verify webhook signatures from Meta? This is the one most often skipped. Without it, anyone who learns a phone number ID, which is not secret, can inject fake customer messages into a workspace, create conversations and make that customer's own AI reply at their expense.

2. How do you verify it? The answer should mention the raw request body and a constant-time comparison. Re-serialising parsed JSON breaks the check subtly.

3. What happens when a signature fails? It should be refused and logged, not processed with a warning.

4. How do you handle the 24-hour window? A platform that lets you send a free-form message outside it is not tracking state properly.

A follow-up question is worth asking: what happens when no app secret has been configured yet. Refusing everything would break live customers mid-migration, so a sensible platform allows it through and records it loudly rather than silently. Either extreme, always refusing or quietly accepting, suggests the question has not been thought about.

Questions about tenant isolation

These matter most for anyone on a shared platform, which is almost everybody.

5. Can one customer's widget be pointed at another customer's AI agent? If the platform accepts an agent identifier without checking ownership, the answer is yes, and the cost lands on the wrong account.

6. Can a user from one workspace be assigned work in another? If identifiers are accepted without an ownership check, foreign names leak into reports and notifications.

7. How are anonymous chat sessions identified? A session identifier derived from a timestamp is guessable, and it is what reads the transcript back.

8. Can I see an example of a request being refused? Ask for a demonstration rather than a description.

Isolation failures are worth taking seriously because of how they surface. Nothing errors and nothing alerts. A customer simply sees a name they do not recognise in a report, or a bill that is larger than their usage, and by then the problem has been running for weeks.

Questions about data and credentials

9. Where is my data stored, and in which country? Under India's DPDP Act you are the data fiduciary, so you need to be able to answer this to your own customers.

10. How are my provider keys stored? They should be encrypted at rest and never displayed back in the interface. Ask whether a support engineer can read them.

11. Who at the vendor can read my conversations? There is usually some access for support purposes. What matters is whether it is limited, logged and rare.

12. If I leave, how do I get my data out? Ask for the format and whether export works while an account is suspended for non-payment. That second part is where policies quietly fail.

Ask about retention too. Chat transcripts contain addresses, phone numbers and sometimes photographs of documents, and keeping them indefinitely by default is both a liability and, under the DPDP Act, hard to justify. A vendor who can set a retention period has thought about it.

How Convo360 answers these

For transparency, here is how Convo360 answers each of the twelve. Meta webhooks are verified against an HMAC of the raw body, compared in constant time, and refused when a signature does not match. Widget configuration checks that the agent belongs to the requesting workspace. Assignee identifiers are checked for workspace membership before they are stored. Anonymous web-chat sessions get long random identifiers rather than anything derived from the clock.

Provider keys are encrypted at rest and never shown back in the interface. Customer text that appears in internal emails, such as an escalation, is escaped, because a customer can write anything and a support inbox is a trusted place to click a link.

Data export is available in CSV and JSON for contacts, conversations, call transcripts, consent records, templates and flows, and it deliberately keeps working when an account is suspended, because the moment you most need your data is the moment you are leaving.

What a good answer sounds like

Specific, slightly technical, and occasionally admitting a limitation. "We verify the signature against the raw body using a constant-time comparison, and we refuse unsigned requests once an app secret is configured" is a good answer. "We take security very seriously" is not an answer at all.

Be suspicious of a vendor who claims no limitations. Every platform has trade-offs, and a team that can describe theirs has thought about them.

Ask for demonstrations where you can. A platform that cannot show you a refusal probably has not implemented one, and the ten minutes it takes to ask is the cheapest security work you will ever do.

Finally, write down the answers. Six months later, when a customer of yours asks where their data is stored, you want a note rather than a memory of a sales call. It also gives you something to check against if the vendor's practice quietly changes.

Ask your vendor the awkward questions

Start free in minutes — no code, no credit card. Bring your own AI keys and pay no per-message markup.

Start 14-day free trial →See pricing

Frequently asked questions

Which question matters most?

Webhook signature verification, because it is the most commonly skipped and the consequences are immediate. Without it, anyone who learns a non-secret identifier can inject messages into a workspace.

Is a certification enough to rely on?

It tells you a process exists, not that a specific control does. Certifications are worth having and they do not answer whether signatures are verified or whether tenants are properly isolated.

What does the DPDP Act require of me here?

You are the data fiduciary for your customers' data, so you must be able to say where it is stored, who can access it, and how long it is kept. Choose a vendor who can answer those in writing.

Should I worry about the AI provider seeing my data?

Know what is sent and say so in your privacy notice. On a bring-your-own-keys platform the relationship is directly between you and the provider, so their terms apply to you rather than to an intermediary.

What is the single best test of a vendor?

Ask them to demonstrate something being refused. Descriptions are easy; a live refusal proves the control exists, and the reaction to the request tells you a good deal on its own.