Method · 18 of 18
What this will not do
The list is the product. Every one of these is technically achievable and most have been asked for by somebody. They are not switched off pending a policy decision. They are outside what this system is for, and the boundary is published on the console so that it can be checked rather than promised.
Refused by design
Nine things this layer does not do
No individual profiling. Scoring stops at the requirement, the account and the centre. There is no per-person score, no behavioural profile against a named member, and no field in the schema that would hold one.
No scraped personal data. Trigger signals are company-level facts from public filings and announcements. Contact details come from your CRM or a licensed provider under warranty. Nothing is harvested from a social platform or a directory.
No calling outside permitted windows. DND scrubbing and calling-hour enforcement sit in the platform, not in a campaign brief. A campaign scheduled outside the window will not dial, and the refusal is logged.
No improvised commercials. The agent quotes from your approved rate card and your live inventory. It cannot invent a discount, extend a term or waive a deposit, and it says so and escalates when asked to.
No cross-client use of your book. Your member data, your rates and your conversations never inform, train or appear in any other operator's system. We work with more than one flex operator and this separation is structural: separate deployments, separate storage, separate keys.
No writing to your systems until you say so. Read-only by default. The write-back is a switch you throw after you have watched it run, field by field, not a default we turn on at go-live.
No scoring of any individual. The person on a call is checked against public filings for one thing only: whether they hold a role that lets them sign. There is no personal credit check, no personal profile, no professional network lookup, and no schema field that would hold one.
No automated adverse decision. The verified lead register recommends a commercial posture. A person accepts or overrides it and the override is logged with a reason. Nothing is declined by a model with no human in the loop.
No pretending to be a person. The agent states that it is an AI when asked, without deflecting, and hands to your team wherever judgement is required rather than improvising past the edge of what it knows.
Why publish this
A listed company buying a conversational layer is buying a disclosure risk alongside it. The first question in the room, and the first question from an infosec reviewer, is what else the thing can do with member data. Answering it on screen, inside the product, before it is asked, is cheaper than answering it in week six of a security review. It also holds us to it.
Where the line is structural
Contractual protection binds the parties to a contract. It does not survive a change of vendor, a change of owner or a subpoena. So the separations above are structural where they can be: separate deployments, no shared identifiers, no schema field for the thing we say we will not do. Where a protection can only be a policy, it is labelled as a policy rather than dressed up as architecture.
Compliance posture
- India entity, in-country data residency
- DPDP Act consent capture and purpose limitation
- TRAI and DLT registration, NDNC scrubbing
- Full transcript audit trail, retention per your policy
- Right to erasure honoured at account and conversation level