No queue reachable without consent
Consent capture with basis-derived expiry, permanent opt-out, a suppression list and the blocked-dial log all work locally. The National DND feed needs a telecom access-provider credential.
Needs a credentialOne function authorises every dial. Nothing else can.Telecaller & Voice Follow-up
Wani places no call without recorded consent, re-checked at the moment of dialling, inside the recipient's own hours.
A queued call, an attempt and a placed call each carry a required foreign key back to the consent that permitted them — there is no column path to a dialled number any other way. The calling window is enforced as a database constraint, narrowed by a workspace and never widened, and every blocked dial is kept as a row naming why, because that log is the evidence a business would need to produce, not an error list to hide.
What Wani refuses to dial is on this page, first
Consent re-checked at dispatch, not just at import9am to 9pm, in the recipient's own zoneThe AI-disclosure node cannot be deleted
How it works
Earn consent, call inside the window, disclose.
Three steps, and the disclosure at the end cannot be edited out of existence, only reworded.
- 01
Earn consent
Capture consent on one of three bases — explicit opt-in, an existing customer relationship, or an inbound enquiry — each with its own expiry rather than one shared toggle. Opt-out is one-way: nothing in this product clears it, and nothing grants consent back to a number that opted out.
- 02
Call inside the window
Every dial passes through one function that can authorise it, and nothing else in the codebase can. Wani only dials between 9am and 9pm, computed in the recipient's own local time zone rather than the caller's. That floor and ceiling are CHECK constraints on the database itself — a workspace can narrow its own calling hours; nothing, not the API, not a migration, not direct SQL, can widen them past 9 to 21. This specific window has not been confirmed against a primary regulatory source, so it is described here as what the product enforces — a deliberately conservative choice — rather than as a citation of a rule nobody here has verified.
- 03
Disclose and hand off
The call opens by saying it is an AI and names the business, and a caller who asks for a person is transferred without a branch that can refuse. Both nodes are rewritten back into a flow on every save — wording is editable, existence is not.
The boundary
Consent is re-read at the moment of dialling, not just at import.
A campaign imported on Monday with a valid basis can reach its last row on Thursday, by which point that particular contact may have opted out on a different call entirely. Queueing and dispatch both run the same gate, but only the dispatch answer authorises anything — validating at import produces a decision that was true once, and validating at dispatch produces one that is true now.
- Explicit consent to receive a promotional call expires in seven days, checked against the 2025 TRAI amendment that shortened it, and read the conservative way: consent gone stale is treated as no consent at all, not as slightly old. A customer relationship licenses a service call for longer, and someone who reached out first is expected to hear back, but not indefinitely — three bases, three different clocks, none of them permanent, and the gate re-checks the clock at the moment of the dial, not just when the number was queued.
- A queued call names the consent that permitted queueing it; an attempt row is written only when the gate passed at dispatch, naming the consent that authorised that particular dial.
- A blocked dial produces no attempt row at all — it produces a compliance-block row instead, which is why the block log is a complete account of every refusal rather than a sample.
- There is exactly one function in the codebase that can authorise a dial. Every route that queues, re-queues, retries or dispatches a call goes through it — a boundary with two entry points is one entry point too many.
In the recipient's clock
The calling window is a database constraint, not a setting.
Wani only dials between 9am and 9pm, computed in the recipient's own local time zone rather than the caller's. That floor and ceiling are CHECK constraints on the database itself — a workspace can narrow its own calling hours; nothing, not the API, not a migration, not direct SQL, can widen them past 9 to 21. This specific window has not been confirmed against a primary regulatory source, so it is described here as what the product enforces — a deliberately conservative choice — rather than as a citation of a rule nobody here has verified. The same machinery that resolves a recipient's time zone already exists elsewhere in the product, reused here rather than rewritten, because this codebase has already paid once for getting local time wrong.
- start_hour_local cannot go below 9 and end_hour_local cannot go above 21 — CHECK constraints, not application-layer validation a migration could quietly skip.
- A workspace that wants a narrower window — 11:00 to 19:00, no Sundays — can set exactly that; nothing can push the floor or ceiling outward.
- A deferred call is re-dispatched and re-gated from scratch, so the worst outcome of the deferral arithmetic being approximate is one wasted queue pass, never a call outside the window.
- Absent DND data blocks promotional dialling rather than waving a number through — an unknown answer is treated as a refusal, not a formality.
What cannot be configured away
The AI says so, and the refusals are kept, not discarded.
Two things stay true no matter what a workspace configures: every call opens with a disclosure that it is an automated assistant, and every blocked dial leaves a row. Neither is a nice-to-have — the disclosure is what a caller needs to know before the call continues, and the block log is the record a business would have to produce if it ever needed to show what it did not do.
- Deleting the AI-disclosure or human-handoff node from a flow does not take effect — normalise_flow_nodes re-inserts both on every save, wording editable, existence not.
- A caller can ask for a human at any point, and the handoff node has no branch that refuses that request.
- The blocked-dial log records why a dial was refused, not just that it was, so 'we never reached this contact' and 'we were not allowed to reach this contact' stay distinguishable.
- Recording a call needs a separate, NOT NULL consent reference tied to a logged disclosure — the schema cannot hold a recording that names no disclosure.
What it does
Thirteen things Wani does today.
Each card carries its own limit, and the label is derived from the backend's capability module rather than written here. Most end at awaiting-provider — the gate is real; the telephony credential is the operator's to add.
9am to 9pm, in their time zone
The permitted window is evaluated in the recipient's own local zone, and calls outside it are deferred with a visible next-attempt time. The floor and ceiling are CHECK constraints — a workspace can narrow them, nothing can widen them.
Works todayQueue, gate, retry — dial pending
Queue, compliance gate, deferral, per-contact frequency cap, retry backoff, dispositions and callbacks are all real. Dispatch ends at awaiting-provider: placing the call needs a telephony credential.
Needs a credentialQueued off a real appointment
Reminders are queued from a Wavy appointment as service-category calls, each gated independently through the same authorisation function. Placing them needs a telephony credential.
Needs a credentialNo call for an unconsented lead
Queues a qualifying call against a Jeet lead through the same gate as every other entry point, so a lead with no recorded consent basis is refused before it reaches a queue.
Needs a credentialNo recording without a disclosure
Disclosure logging, the consent gate and a mandatory retention clock are enforced: the recording's consent reference is NOT NULL, so a recording naming no disclosure cannot be written.
Needs a credentialNo branch that refuses a person
An explicit request for a person is classified before scope or confidence, the route has no branch that refuses one, and a staff roster with round-robin routing now decides who the call goes to. The warm-transfer telephony leg and paging the receiving agent's device both still need providers this deployment doesn't have — the response names which, rather than claiming a device was paged.
Needs a credentialTwo nodes no workspace can delete
Templates, a hand-built configuration, or a sentence describing what the agent should do — all three write through the same normaliser, so the AI-disclosure and human-handoff nodes are re-inserted on every write.
Works todayA summary and a tag, never a message
Writes the disposition and summary onto the contact's Wavy conversation as a summary and a tag, never as a message in the customer's own thread. A failed sync surfaces rather than retrying quietly.
Works todayNull, not zero, when nothing happened
Daily volume, outcome breakdown and blocked-dial counts from this workspace's own rows only. Connect rate and average talk time are null rather than zero when nothing has an outcome yet.
Works todayPaced to the window, not around it
Pacing is recomputed per read from the minutes left in the recipient's window and clamped to the carrier's documented ceiling. A paused campaign is refused at the gate, not filtered from the queue, and launching one now queues every contact on a list straight through the same gate a single call uses — reporting queued, deferred and blocked per contact rather than aborting the batch on one bad row. Dialling at the paced rate still needs a telephony credential.
Needs a credentialA real catalogue, or an honest empty one
Voice selection over a real provider catalogue; listing returns the real list or an empty one naming the missing credential, never a plausible voice id nobody has checked. Holding a conversation needs speech-to-text, not built yet.
Needs a credentialA registry of numbers, before any dial
A register of the numbers a business calls from with its own capability, and a record of which published agent answers on which number. Nothing is provisioned and no call is answered — there is deliberately no inbound webhook with no provider behind it to sign one.
Needs a credentialA summary that says what it inferred
Summary generation runs end to end off the call's captured fields, notes and disposition through the shared LLM layer, tagging every point as a confirmed field or the model's own inference rather than inventing what was said. This deployment records no call audio and has no speech-to-text credential, so the transcript itself always reports why rather than fabricating one — a summary can never be saved without an explicit language, to guard a Hindi or code-mixed call against being silently mangled by an English default.
Needs a credentialA drop message that cannot be saved incomplete
The voicemail drop message is real: sender identification and an opt-out instruction cannot be saved blank, enforced by the same kind of constraint that keeps the AI-disclosure node from being deleted. Actually detecting a machine needs the telephony provider's answering-machine-detection feature, and speaking the message needs text-to-speech plus that same provider — neither is configured, so nothing but this API writes a row here yet.
Needs a credentialA flag a human always reviews, never an auto-verdict
Quality flags are classified in the same pass that builds a call's summary; a flag naming a type outside the closed set is dropped rather than stored under a guess. Listing, review — which always requires a human's own note — and a summary that groups repeated flags into a suggested script fix all run locally. The only blocker is an LLM credential.
Needs a credentialThis page still needs a telecom-compliance reviewer's eyes on it.
Every claim above traces to a route, a CHECK constraint or a status entry in the code, and a test fails the build the moment the two disagree — including a test that this page never cites a regulation number for the 9am-to-9pm calling window, because that specific figure has not been confirmed against a primary regulatory source. That is honest labelling, not compliance sign-off. Before this page goes live, a qualified telecom-compliance reviewer familiar with TRAI's current rules should confirm the calling window, the consent-expiry figures and the DND-scrub description are accurate as stated, and correct anything that is not. Saying that here is a trust signal, not an admission.
Before you switch
Earn the consent. Wani will not dial without it.
Tell us what your calling programme looks like today, and we will tell you exactly what Wani can run now — and where it insists on a telephony credential before a single call is placed.
A missed call, returned
The one sequence Wani does end to end today: somebody rings and cannot get through, Wani rings back, and an appointment is booked only after the date is said out loud and agreed to.
Scripted demo. These are written lines, not a recording of a real call — no call has been placed to a handset from this deployment yet.
Turn 1 of 9
- Wani
Hello, this is an AI assistant calling on behalf of Sharma Dental. You rang us a few minutes ago and we could not pick up — sorry about that.
The AI disclosure opens the call. It is re-inserted into every flow by normalise_flow_nodes, so no agent can be configured without it.
- Caller
Oh, yes. I wanted to move my check-up.
The missed call was recorded as an inbound_request consent, valid for 30 days, with the caller's number and the hangup cause as its evidence.
- Wani
Of course. Which day would suit you better?
- Caller
Tuesday, if you have anything in the afternoon.
- Wani
I can do Tuesday the first of September, at four o'clock in the afternoon. Shall I put that down?
The read-back. No appointment exists at this point — a CHECK constraint means a hold cannot carry an appointment until it is confirmed, and cannot be confirmed without these words stored.
- Caller
Yes, that works.
- Wani
Booked. You will get it in writing from Sharma Dental. Anything else I can help with?
Only now is the appointment written, and its notes carry both what was read back and what the caller said in reply.
- Caller
No, that is all. Thank you.
- Wani
Thank you for your time. Goodbye.
The call is billed from the carrier's own duration, rounded up to the minute, when the CDR arrives — never from our clock.
Scripted demo. These are written lines, not a recording of a real call — no call has been placed to a handset from this deployment yet.
What a call costs
One credit is ₹1. You are charged per connected minute, rounded up, using the carrier’s own duration rather than ours — so ringing time is not billed and a call nobody answered costs nothing.
| Plan | Credits per connected minute | What it includes |
|---|---|---|
| Free | 10 | English only, on free voice models. The signup bonus is 100 credits, so the free tier is ten minutes of talking — enough to hear it work. |
| Standard | 5 | Hindi and the other Indian languages, on Sarvam. |
| Professional | 4.5 | Adds ElevenLabs. Fractions of a credit are carried between calls rather than rounded up on each one, so ten minutes costs 45 credits and not 50. |
| Advanced | 3.5 | Adds Grok and the self-hosted Hindi voice. |
| Enterprise | 3 | Adds the seam for a voice model trained on your own calls. Not on sale yet: the carrier's per-minute cost is still unknown here, so this rate is flagged as unverified in the database and Wani will not sell it until it is. |
GST at 18% is charged once, when you buy credits — 100 credits costs ₹118. Calls then draw down the credits you already own, with no tax applied again.
A phone number is separate. The carrier charges a one-off setup fee and a monthly rental, and both pass through at cost — outside credits, with nothing added on top.
Which languages anyone has actually listened to
No language has been verified end to end. A voice product that says “supports 12 Indian languages” without saying which of them anyone has actually listened to is making a claim it cannot back, so this table separates what is written from what has been heard.
- EnglishChecked by tests
The spoken date read-back is checked by tests: no digits in the date, the weekday always said, twelve-hour time with the part of day. What has not happened is a real call — the words are right, nobody has heard them down a phone line.
- HindiWritten, never heard aloud
The read-back is implemented from written grammar and has never been checked by a native speaker or heard aloud. Every appointment booked in Hindi is stored with read_back_verified set to false, so the gap is visible in the data rather than only in this paragraph.
- Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, AssameseNo spoken read-back yet
Selectable on the paid tiers for speech, but there is no spoken date read-back for them yet, and Wani refuses to book an appointment it cannot read back rather than falling back to English.
Questions
Straight answers.
Does Wani ever call a number without consent?
No — not as a policy, as a database shape. There is exactly one function that can authorise a dial, every entry point that queues or dispatches a call goes through it, and a queued call, an attempt and an outbound call each carry a NOT NULL foreign key back to the consent that permitted them. A blocked dial produces a compliance-block row instead of an attempt row — there is no column sequence that reaches a dialled number by another path.
What counts as consent, and how long does it last?
Explicit consent to receive a promotional call expires in seven days, checked against the 2025 TRAI amendment that shortened it, and read the conservative way: consent gone stale is treated as no consent at all, not as slightly old. A customer relationship licenses a service call for longer, and someone who reached out first is expected to hear back, but not indefinitely — three bases, three different clocks, none of them permanent, and the gate re-checks the clock at the moment of the dial, not just when the number was queued.
What hours does Wani place calls in?
Wani only dials between 9am and 9pm, computed in the recipient's own local time zone rather than the caller's. That floor and ceiling are CHECK constraints on the database itself — a workspace can narrow its own calling hours; nothing, not the API, not a migration, not direct SQL, can widen them past 9 to 21. This specific window has not been confirmed against a primary regulatory source, so it is described here as what the product enforces — a deliberately conservative choice — rather than as a citation of a rule nobody here has verified.
Can a business widen the calling window if it wants to call later in the evening?
No. start_hour_local and end_hour_local are bounded by CHECK constraints that cap the floor at 9 and the ceiling at 21 — not through the API, not through a workspace's own migration, not through direct SQL. A workspace may narrow its own hours; nothing in this product can widen them past that bound.
Does the AI tell people they are talking to an AI?
Yes, at the start of every call, and that opening cannot be removed. Deleting the disclosure node from a flow does not take effect — the same normaliser that writes any flow re-inserts it on every save. The wording can be customised; the node's existence cannot.
What happens if someone asks to speak to a person?
The request is classified before anything else about the call — before scope, before confidence — and the handoff route has no branch that refuses it. The warm transfer itself is a telephony-provider capability this deployment does not have configured yet, but the request to be handed off is never declined.
What is the blocked-dial log actually for?
It is the evidence a business would need to produce if it ever had to show what it did not do, not a hidden error list. Every refusal — no consent, an expired one, an opted-out number, a suppressed one, an unresolved DND answer — writes a row naming the reason, so 'we never reached this contact' and 'we were not allowed to reach this contact' stay two different, checkable facts.