One Agent, Four Channels, One Memory of the Customer: What ThinnestAI Actually Is

What this is

One AI agent that answers your customers on your website, WhatsApp, Telegram and the phone. It remembers who it has spoken to across those four places, records a lead in the middle of a conversation, calls people back, and hands the thread to a person when the conversation stops being one it should be having.

That is the whole product. What follows is each part, with the limit stated next to it, because the limits are where you find out whether a platform fits.

There are exactly four channels

• Website widget. One script tag. Cross-origin isolated, so nothing on your page can read the conversation and nothing in the widget can touch your page. A returning visitor gets their thread back — the last 30 messages, for 30 days.
• WhatsApp. Direct on Meta's Cloud API, not resold through a third party. Templates, broadcasts, one-time codes and transactional messages. Connecting an account is self-serve through Meta's Embedded Signup — we are an approved Tech Provider, so you sign in with Facebook, choose the number, and it is live in minutes.
• Telegram. Paste a bot token. The only channel that is fully self-serve end to end right now.
• Voice. Inbound calls, outbound calls and in-browser calls you can make before a phone number is involved at all.

There is no email channel. We send platform mail — invitations, notifications — and that is all. No Instagram, no Messenger, no SMS, no Slack. Four is the number.

One memory of each customer, with a deliberate hole in it

This is the part most worth understanding, because it is the thing the product is actually built around and it is also the thing we refuse to do the easy way.

A customer who messaged you on WhatsApp last week and rings your number today is not a stranger to the agent. Their history from the other channels is injected as a short digest before the agent speaks — on the website, on WhatsApp, on Telegram, and on a live phone call. It is capped: six messages, 160 characters each, 700 characters in total. A customer who has only ever used one channel costs nothing extra, because there is nothing to inject.

Phone callers are matched to their WhatsApp and web history by phone number. Which brings us to the limit.

Caller ID is spoofable. So records are linked, never merged, and what the agent may reveal is governed by a disclosure tier you set. The default tier is the middle one: the agent will say "we have spoken on WhatsApp before" but will not repeat what was said until the person verifies who they are. You can set it to reveal content, or to reveal nothing at all until verification.

A competitor advertising total cross-channel recall is advertising a privacy hole. Anyone can ring a number claiming to be your customer. We would rather ship the cautious default and let you loosen it than ship the demo-friendly one.

Two things this is not: there is no retrieval tool that fetches the full earlier conversation on demand, and the agent cannot start a conversation with someone on a different channel from the one they are on.

Lead capture is the thing to turn on first

The agent can record a lead in the middle of a conversation — name, email, phone, what they wanted — without stopping to run an interrogation. Partial details are accepted, because a person who gives you a name and a phone number and then leaves is still a lead, and a form that demands all four gets you none of them. It is a per-agent switch, and every captured lead deep-links back to the conversation it came out of, so the person following up reads what was actually said.

Reaching out, rather than waiting

Calling campaigns

• Number pooling. Dial from a set of numbers rather than burning one, so no single number gets flagged as a spam dialler. The same customer keeps getting the same number. 200 calls per number per day.
• Pacing. Cap calls per rolling hour, restrict which weekdays you dial in the customer's timezone, and skip holidays in the business's timezone. Those are two different clocks and treating them as one is how a campaign rings people on a public holiday.
• Outcomes, not activity. Connect rate, average talk time, and why the rest failed — busy, declined, unreachable, voicemail. "Calls placed" is a receipt, not a measurement.
• Retries that make sense. Busy retries in minutes, no answer in an hour or more. Someone who actively declined is never rung again, and neither is a dead number.
• Voicemail detection leaves a campaign-specific message instead of talking to a machine. Available on ThinnestAI, Plivo and Telnyx numbers.
• Callbacks. "Call me back after five" resolves against your workspace clock, respects your calling hours, and can be cancelled from the console.
• A do-not-contact list that outlives any single campaign, separate from per-contact consent.
• Per-person variables fill both the opening line and the agent's instructions from a spreadsheet column, and A/B testing runs two opening lines against complementary halves of one audience.
• Every call is logged with its transcript, recording playback and a CSV export.

WhatsApp broadcasts and drips

Broadcasts run against consent-gated audiences — four tiers by risk — with tag segmentation and behavioural retargeting: people who did not reply, people who did, people never reached, people who refused. Sends can be scheduled, resume safely if something crashes mid-run, and replies are attributed back to the send that caused them on a 72-hour last-touch window.

Drip sequences put each person on their own clock and exit them on a reply, an opt-out or a tag. Consent is re-checked at every send, not just at enrolment — because a person who opted out on Tuesday should not get Wednesday's message from a sequence they joined last month.

Your own backend can post a fact to the Events API — an order was delivered, say — and that enrols someone in a sequence. Order confirmations and one-time codes go out through a transactional messaging API with idempotency keys, rate limits and API keys.

The shared inbox, and what takeover means

Conversations from all four channels land in one live inbox. The agent escalates on triggers you pick per agent — it cannot answer, or the customer is upset and asking for a person. Turning a trigger off removes the tool rather than asking the model nicely, so the behaviour becomes impossible rather than discouraged.

When a teammate replies, the agent stops answering that conversation. Two voices in one thread is a failure everyone has seen once. You also get media replies both ways, conversation assignment, saved replies, and internal notes the customer never sees — including the agent's own reasoning about why it escalated, so whoever picks the thread up reads why before they type.

What the agent knows, and what it can do

Knowledge comes from a website crawl one to three levels deep, a sitemap import, text you paste, or files you upload — txt, md, csv and the text layer of a PDF. There is no OCR, so a scanned document is not a knowledge source here. CSV chunking is row-aware, so a price never gets separated from its product.

For anything that is not in a document, the agent can call a tool: 21 pre-registered MCP server connections plus a directory of 1,434 services, your own MCP address, your own HTTP endpoints, Shopify in two separate lanes, and built-in web search. Every tool arrives switched off, and an agent stops at eight. These tools are reachable from a public chat by a stranger; nothing should be live because somebody clicked through a setup screen.

The agent replies in 37 languages, including all 22 scheduled Indian languages in their own scripts, and handles Hinglish code-switching. One knowledge base serves every one of them — you do not maintain a Tamil copy of your policies.

What it costs

• Voice from ₹2 a minute — ₹2 for standard voices, ₹2.5 for premium ones.
• That rate is all-inclusive. Telephony, speech recognition, the language model and speech synthesis are all inside it. No per-language surcharge, no provider passthrough, no platform fee plus costs.
• Billed in 30-second pulses, not rounded up to the whole minute. Four short calls cost about what four short calls should.
• Chat replies are usage-based.
• Plans: a free trial with 25 voice minutes and 200 chat replies and no card, then Pay As You Go with no monthly fee, then Scale at ₹4,999 a month or ₹54,989 a year. Enterprise is negotiated.
• Prices are pre-GST. Paid in rupees through Razorpay and UPI.

An Indian phone number is ₹349 a month from us (₹249/mo on Scale), or nothing if you bring your own number from Plivo, Vobiz, Twilio or Telnyx and keep paying that carrier directly.

Where the data sits

• GCP Mumbai (asia-south1).
• DPDPA 2023: consent capture, configurable retention with automatic purge, right to erasure, 72-hour breach notification.
• TRAI: DND and NCPR scrubbing, calling time-boxed to 9am–9pm IST, an AI disclosure toggle, and 140-series support. Telemarketer registration itself is yours to hold, not ours.
• Payments localised through Razorpay for RBI.
• Prompt-injection fencing — everything a tenant adds is labelled as data before it reaches the model, and eight live attacks run against the public endpoint as part of the test suite.
• Call recordings are deleted after 30 days on Pay As You Go, 75 on Scale.

What we deliberately do not have

• No visual flow builder. No canvas, no nodes, no edges. Configuring an agent is a tabbed settings screen: instructions, knowledge, tools, voice, channels.
• No email as a channel the agent answers on.
• No speech-to-speech or realtime voice model. The voice pipeline is cascaded — speech to text, then a language model, then text to speech.
• No bring-your-own model keys. Models are curated and platform-managed, which is the only reason a flat per-minute rate can include the language model at all. Bringing your own number is the one bring-your-own thing here.
• No automatic learning loop. The agent does not quietly rewrite its own behaviour between conversations. You change it, and the change is visible.
• No agent-to-agent delegation, and no AI judge scoring your conversations out of ten.
• No published latency or call-quality figure. Carrier-call performance is not something we have measured to a standard we would defend, and a number we cannot defend is worse than no number.

Get started

Create an agent, give it your website to crawl, and talk to it in the browser. Then connect Telegram — it takes a pasted token — and watch the same agent answer somewhere else. Add a phone number when you actually want one to ring.

Create your first agent free →

No credit card required • Trial includes 25 voice minutes and 200 chat replies