What we
shipped.

Connect a WhatsApp Business Account and Talkif answers voice calls placed to your WhatsApp numbers with the same AI agents that handle your phone lines.

Connecting

Two paths:

• OAuth Embedded Signup — connect through Meta's guided flow; Talkif discovers your WABA and numbers automatically.
• Manual credentials — paste your WABA ID, phone number ID, and access token.

A WABA registers as a provider, and you can register multiple numbers under it in one step — each number reports back individually as registered, already registered, or failed.

How numbers behave

WhatsApp numbers are receive-only: agents answer inbound WhatsApp calls; outbound dialing stays on your phone lines. A number goes Active once Meta reports calling enabled — until then it shows as Pending so the dashboard never claims an un-callable number is live. The same phone number can exist as both a regular phone line and a WhatsApp number, each with its own flow assignment.

Pricing

WhatsApp calls carry no telephony charge and no monthly number cost — Meta bills you directly for the calling service. Talkif charges only the AI legs: speech-to-text, LLM, text-to-speech, and infrastructure.

See the WhatsApp docs.

The Assistant lives in the dashboard and works against your real account data.

What it does

• Answers questions — calls, campaigns, billing, flows: it looks up your actual data, read-only.
• Builds flows — describe the agent you want (optionally point it at your website) and it researches, architects, writes the prompts, and assembles a complete draft flow. Progress streams live; the result lands as a normal draft you review before publishing.
• Edits flows — proposes changes as diffs you apply or reject. The assistant never changes anything without your confirmation.
• Escalates to support — hand a conversation off to the support team without leaving the chat.

Usage

Every account gets a free monthly allowance. Real-money top-ups add bonus assistant credit on top, and once both are exhausted, usage bills to your balance as a separate Assistant Usage line — visible in transaction history.

See the Assistant docs.

The provider lineup for flows grew substantially:

You can also add voices from the ElevenLabs shared library to your account, and the LLM catalog was refreshed to the July 2026 lineup with current official pricing.

All providers are configurable per-flow in the builder — see provider nodes.

Failure reasons

When a call fails, the call detail now shows a specific failure code and a human-readable reason taken from the carrier's status callback — busy, invalid number, carrier rejection — instead of a generic failed state.

Performance metrics

Every completed call reports: time to first response, turn count, interrupted turns, function call count and total time, user and bot speech seconds, token usage, and TTS characters — including wasted TTS characters: speech your bot generated that the caller never heard because they interrupted. A consistently high waste number is a signal your agent's turns are too long.

Calling-hours compliance

Outbound calls are checked against jurisdiction-aware legal calling hours before dialing. Calls outside the window are rejected up front with outside_callable_hours, and queued calls that drift out of the window are deferred until it reopens.

See call details and making calls.

Two flow-level settings now control how and when a call ends on its own — find them under Flow Settings → Call lifetime.

Inactivity handling

On by default. When the caller goes quiet after the agent speaks, the agent checks in — "Are you still there?" — and ends the call politely if there's still no response. The early checks are short, LLM-generated nudges that adapt to the conversation's language; the final check speaks a fixed goodbye and hangs up. Tune the silence timeout (default 10s) and how many checks to allow (default 2). The counter is monotonic by default, so background noise can't keep an abandoned call alive forever.

Maximum call length

Off by default. Set a hard ceiling — 30 seconds to 4 hours — measured from the moment the call is answered. When reached, the agent says a brief goodbye and hangs up regardless of conversation state. It's a graceful ceiling: a goodbye queued behind an in-progress sentence waits for that sentence to finish.

Use both together for full protection against stuck calls. See the Call Lifetime docs.

Connect a Facebook Page and Talkif places an AI phone call to every Lead Ads submission within seconds — running the flow you choose, from the caller ID you choose.

How it works

1. Connect your Page via OAuth — Talkif subscribes its lead webhook automatically.
2. Map each lead form to a published flow, a caller-ID number, and calling rules (delay, retries on no-answer).
3. A lead submits → Talkif creates a contact and places the call.

Closing the loop with Meta

With a Conversions API dataset configured, Talkif reports funnel-stage events back to Meta as each lead progresses:

This lets Meta's Conversion Leads optimization target leads that actually convert on the phone — not just form submitters. Event names and outcome mappings are fully customizable to match your Events Manager setup.

See the Meta Lead Ads docs.

Flow functions give your agents a way to reach live data during a call: look up an order, check availability, create a ticket, verify a customer — anything where the conversation needs to talk to your systems.

The shape

Each function defines method, URL, and three independent parameter locations:

Parameter names are unique across all three locations, URL placeholders must match pathParams keys, and the dispatcher percent-encodes path values to prevent injection.

Where values come from

Each parameter is bound to one of three sources:

• llm — the model extracts the value from the conversation
• call_context — server injects from the active call (caller identity, variables). Never shown to the LLM. onNull: reject hides the function for that call; fallback_to_llm exposes it.
• static — fixed value baked in. Never shown to the LLM.

This lets you keep sensitive identifiers server-side while still giving the LLM everything it needs to decide when and how to call the function.

Execution

Functions run server-side. Headers are encrypted at rest with AES-256-GCM. Timeouts are configurable (100–30000ms, default 5000). If the function has transitionTo, the agent transitions even when the HTTP call fails — the LLM sees the error in the new node's context.

See the flow functions docs for the full request shape, binding modes, and validation rules.

Some reasoning models reject a caller-specified temperature and use their own defaults. Talkif now strips the field before the flow definition reaches the bot, so you can see at a glance which models accept which parameters.

What's affected

gpt-5, gpt-5-mini, gpt-5-chat-latest, gpt-5.1-chat-latest, and gpt-5.2-chat-latest do not accept custom temperature. If you set one on a flow using these models, the value is removed when the flow is saved or returned to the editor.

How to tell

The model list response now includes an llmCapabilities.temperature boolean per model. The in-app model picker surfaces this directly.

Why

Before, the bot would receive a temperature value, silently drop it to call the model, and you'd never know whether your setting had any effect. Now the flow definition reflects what will actually run.

Phone-to-flow version behavior is now predictable:

• Default (unpinned): The phone follows the latest published version. Publishing a new version immediately starts serving it on all connected phones.
• Pinned: You can pin a phone to a specific version. When that version is republished, pinned phones are auto-bumped to the new version in the same transaction — preserving the intent that this phone stays in sync with this particular flow.

Previously, a phone could silently become pinned to whatever the current version happened to be at connection time. That surprise is gone — pinning is now opt-in and explicit.

The difference matters:

Publishing a flow that declares tts_say or forward_call as a function is rejected with unknown_builtin. Move them to preActions or postActions with the appropriate config ({ text } for tts_say, { phoneNumber } for forward_call).

The action nodes and flow functions docs have been updated to reflect the correct shape.

Every payment method on your account now carries a status that updates automatically based on charge outcomes.

Statuses

Status resets to Active on any successful charge. When a charge fails, the specific Stripe decline code is stored so you know exactly what went wrong.

Auto-recharge will not attempt charges against Expired or Blocked methods. If your default card enters one of these states, add a replacement before your balance runs low.

You can also now assign custom labels to payment methods (e.g., "Corporate Visa", "Marketing Budget") to stay organized across multiple cards.

The billing section now includes a full charge history alongside the existing transaction log.

While transactions track credit movements (purchases, refunds, expirations), charges track the individual billable events that consume those credits.

Charge types

• Call Usage — telephony, STT, TTS, and LLM costs per call
• Phone Purchase — one-time number acquisition
• Phone Rental — monthly number rental
• Call Analysis — post-call AI analysis
• Recording Storage — storage costs, billed when recordings expire or are deleted
• Adjustment — manual credits or debits
• Reservation — pre-authorized holds during active calls

Each charge includes a status lifecycle (Pending → Completed/Failed/Refunded), cost line items, and context about the associated entity (phone number, flow, or contact). Click any charge to see the full breakdown.

You can now export a complete copy of your account data directly from Settings > Data Export.

The export includes 12 data categories: account info, users, contacts, calls, flows, phone numbers, VoIP providers, call schedules, campaigns, charges, balance transactions, and notifications.

How it works

1. Request an export from Settings
2. Confirm via email (security verification)
3. A background job collects and packages your data
4. Download link arrives by email — valid for 24 hours

The export is a structured JSON file, machine-readable and portable. Sensitive fields like passwords and MFA secrets are excluded automatically.

Limits

• One active export at a time
• 24-hour cooldown between exports
• Download available for 24 hours after generation

This fulfills GDPR Article 20 (right to data portability). Full details on our Privacy Policy.

Every prompt in Talkif is a Handlebars template. When a call starts, the platform resolves that template against real data: who the contact is, why they're being called, what business is calling, and what time it is in their timezone.

Four namespaces are available at resolution time:

• contact.* — name, email, company, timezone, language, notes, interests, tags, custom fields
• call.* — direction (inbound/outbound), caller and callee numbers
• account.* — your business name and identifier
• system.* — current date, time, and timezone

A template looks like this:

handlebarsYou are calling on behalf of {{default account.name "our office"}}
to confirm an upcoming appointment.
 
Address the contact as {{default contact.firstName "there"}}.
 
{{#if contact.email}}
  Offer to send a confirmation email to {{contact.email}}.
{{/if}}
 
{{#if (includes contact.tags "vip")}}
  This is a high-priority contact. Be especially attentive.
{{/if}}

The engine supports conditionals, comparisons (eq, ne, gt, lt), regex matching, array operations, fallback defaults, and existence checks. Template resolution runs in microseconds — the agent picks up, greets the caller by name, and references their specific context in under one second.

Writing Handlebars templates from scratch isn't always intuitive — especially when you have dozens of available variables and helpers.

The Prompt Library now includes an AI assistant that knows your schema. Describe what you want in plain language. It generates a complete, valid template using only variables that actually exist in your account.

The assistant enforces proper Handlebars syntax, avoids literal dialogue scripts, and supports multi-turn refinement. You stay in control of the final template — the AI just gets you there faster.

textYou: "I need a prompt for appointment reminders that 
      checks if they're a VIP and mentions their timezone"
 
Assistant generates:
  {{#if (includes contact.tags "vip")}}
    This is a priority client. Be especially accommodating.
  {{/if}}
  Their local time is based on {{default contact.timezone "UTC"}}.

No more guessing which variables exist or getting the syntax wrong.

When you run an outbound campaign, contact data is now captured at enrollment time and frozen for the duration of the campaign.

If a contact's name, company, or notes change between scheduling and dialing, the call still uses the original data. Retries and rescheduled attempts stay consistent — the agent references the same context across every attempt.

This prevents a class of bugs where mid-campaign CRM updates cause the agent to reference stale or conflicting information. The prompt your agent receives on retry #3 is identical to what it received on attempt #1.