Conversations
Chat, voice and the avatar in one window. Every answer names its source, and an answer without coverage in your knowledge never reaches the customer.
Product
Built around one question: how does the agent know that?
Most chatbots generate an answer and then try to make it credible. AIvatary works the other way round: it first finds coverage in your sources, and an answer without coverage is never generated.
Five modules, one Ada
Ada works from the first question on your website to a booked meeting. Each module has its own page with a sample conversation and a clear split between what works today and what is still being built.
Meetings
A slot picked inside the chat window, a confirmation with a calendar file and reminders. Ada runs the demo herself, prepared on your own website.
Inbound and reply in one thread, the same knowledge as on the website, and sending through a queue instead of inline.
Offers
Material surfaced when it answers the question just asked. You define the catalogue, and the agent never invents a link of its own.
Signals
The gap register, a record of every turn and quality scoring add up to a picture of what customers ask and what your materials miss.
Where Ada's behaviour comes from, meaning context, knowledge, guardrails and improvement: the platform and its working cycle.
A knowledge base with an editorial cycle
The agent's knowledge is not "an uploaded file" but content with a process, much like a newsroom's.
Draft → review → publish
An article moves through states: draft, reviewed, published, archived. The agent uses published content only, so nobody accidentally "uploads" an unfinished procedure into the agent.
Versioning
Every change creates a new version. You can see exactly what the agent knew on the day a given conversation took place.
Import from URL
Pull content from an existing help site. The importer has built-in protection against reaching into your internal network (SSRF guard).
Knowledge-gap register
Questions the agent couldn't answer land in the gap register. It's a ready-made "what to write next" list, ordered by real customer conversations rather than guesswork.
Gates before the model, not after it
Before the language model is even called, the conversation passes a cascade of checks. If any check fails, the model call simply never happens.
Citations checked before the first character
An answer must point to sources in your knowledge base. The check runs before anything is streamed to the customer: an uncovered answer is rejected in full and the conversation can be handed to a human. There is no "send now, correct later" mode.
Operations via confirmation cards
When a conversation leads to a data change (a refund, an address update, a cancellation), the agent doesn't perform it. It issues a confirmation card that a human approves with a separate click. Every card leaves an audit trail.
Quality review off the critical path
Answer quality assessment runs in parallel and never delays a conversation, because it has a hard time budget. Results go to the panel; they never sit between the customer and the answer.
Tenant separation in the queries themselves
Organisation scope is a condition of the knowledge-base query itself, not a filter on results. Another platform customer's data cannot "leak" into your conversation, because the query for it never exists.
Channels
Each of them leads to the same knowledge base, the same citation gate and the same handoff to a human.
| Channel | Status | Details |
|---|---|---|
| In-app chat | Working | Full agent: streamed answers, tools, confirmation cards. |
| Widget on your website | Working | Anonymous chat embedded with one line of script, with its own, stricter set of safeguards. |
| Outbound email | Working | Delivery through the task queue, SMTP transport. |
| Outbound SMS | Working | SMS notifications (Twilio); disabled as a conversation channel by default. |
| Push notifications | Working | Web Push to the customer's and the team's browsers; off by default, needs VAPID keys. |
| Inbound email | Working | IMAP polling: customer replies land in the same conversation; needs a mailbox to watch. |
| Inbound SMS | Working | Replies to text messages within a 24 hour window, opt out with the word STOP, delivery statuses; needs a provider to be selected. |
| Browser voice | Working | A voice conversation in the browser window over the realtime server; needs the transport configured and the realtime worker image. |
| Inbound telephony | Working | A call to the customer's own number over a SIP trunk; number and trunk stay on the customer side, handoff without an announcement, no recordings. |
You can verify channel state yourself at any time: the application
publishes a capability matrix at /capabilities. It is
the same mechanism the table above uses, described in the
documentation.
In progress
Channels on the way
We order them by what you ask for. Signing up for a notification is also a vote.
WhatsApp, Messenger, Instagram
The channels customers reach for first, wired into the shared conversation inbox.
See the roadmapAda in Slack and Teams
An AI coworker inside the company, for the team rather than for end customers.
See the Ada previewOperator panel
Everything your team needs to supervise the agent, in one place.
Conversation inbox
A view of every agent conversation with customers, including full history and answer sources.
Human handoffs
Conversations the agent couldn't or shouldn't finish go to a takeover queue with their full context.
Audit log
Who approved what, when, and in which conversation. Confirmation cards, knowledge changes, operator decisions.
Widget embedding
A panel screen recognises your site, hands you the snippet to paste or pastes it for you over SFTP or FTPS, and then checks that the widget came up. Hosting credentials are encrypted the moment they arrive and deleted once the job ends, within an hour at the latest. Over SFTP you first compare your server's key fingerprint, and assisted installation is switched on by the organisation owner, who accepts an annex to the data processing agreement.
The best test is your own knowledge
Upload a few real help articles and ask the questions your customers actually ask.