How Winaity AI works
Your AI agent takes no code: start from a template, teach it your business, try it, then put it live.
The journey of a message, from start to finish: let's follow one, step by step. 1, customer message: a customer writes “Hi, where is my order no. 4821?” on WhatsApp; web widget, WhatsApp, Messenger… the journey is always the same. 2, guardrails: what comes in is filtered, and attempts to divert the agent from its role are blocked. 3, your agent: this channel's agent, here customer service, takes the message, with its role, its instructions and its documents. 4, knowledge base: it reads your documents, “Delivery within 3 to 5 working days after shipping”, and answers from your own passages, not from a made-up policy. 5, tools: mid-conversation, it reads the order tracking through your API: shipped, delivery on Thursday. 6, check: by default, an anti-hallucination net compares the answer with your documents and removes what they don't confirm. 7, answer: “Your order no. 4821 has shipped: delivery expected on Thursday”, sent on the same channel; guardrails also read what goes out: toxic language, profanity and, if you switch it on, personal data. Another message: “I'm disputing an invoice, I want to talk to someone.” The agent hands over by your rules: an email to your team, with a link to the conversation, or a ticket in your tool; it tells the customer their request has been passed on. In short, every message follows this journey, on every channel, day and night. Channels and integrations: one single agent, set up once, answers on the web widget, the REST API, WhatsApp, Messenger, Instagram and Telegram, connected to your APIs and your ticketing tool; the same documents and the same escalation rules on every channel you switch on. Conditional journeys: for more complex cases, “if the customer asks for a refund, check with an API call that the order has been delivered, then send the return procedure, otherwise hand over to the team”, no code, in a visual editor. Templates: three ready-to-use templates, e-commerce order tracking, customer service and B2B lead qualification; clone one, fill in its variables, make it yours. Winaity AI: create your AI agent, at winaity.com. Voice-over: “What happens to a message sent to your AI agent? Let's follow one. A customer writes, on your website or on WhatsApp: the journey is always the same. First, guardrails filter what comes in. Then this channel's agent takes it, with its role and its instructions. It looks for the answer in your documents. And it calls your API, here to track the order. Before it's sent, the answer is checked against your documents. It goes out on the same channel, read again by the exit guardrails. If the customer asks for a human, the agent hands over by your rules: your team gets a link to the conversation. Every message follows this journey, day and night. You set up the agent once, for all your channels. For complex cases, you draw journeys, with no code. And to get started, three templates are ready to use. Winaity AI: create your AI agent.”
A message's journey
From your customer's question to the answer, step by step
The film shows it in a minute; here is the same journey in writing, with what you can set at each step and what the product does not promise. It is the same day and night and on every channel, apart from the few differences noted along the way.
Your customer writes
On your website's widget, on WhatsApp, Messenger, Instagram or Telegram, or from your server through the REST API. They write the way they talk: no menu to work through, no keyword to guess. Whatever the channel, the message reaches the same agent, which you set up only once.
Guardrails examine what comes in
Three checks read the message before the agent does: detection of attempts to divert it from its role, moderation of violent or sexual content, and the detection of personal data such as an IBAN or a card number, logged by type while the message itself is left unchanged. These checks are always on. A refused message gets a fixed reply.
This channel's agent takes the message
It reads it with what you wrote for it: its mission, its tone and its reply to off-topic requests. It keeps the thread of the exchange: your customer does not have to repeat an order number in every message. If the message matches a journey you have drawn — on a keyword, at the start of the conversation or on a detected intent — the agent follows its steps.
It looks in your documents
The agent searches its knowledge base for the passages that answer the question. That base is its own: neither your other agents' documents nor another organisation's are in it. It relies on those passages to answer. In the widget, the answer can show up to three sources: the link of the web page or the name of the file.
It queries your tools
When the answer lies not in a document but in a system — the status of an order, the progress of a case — the agent calls the API you declared in a connector or in a journey. A read runs directly. Declared in a connector, an action that changes something, such as opening a ticket, first asks the visitor to confirm and is not offered on messaging channels. In a journey, the call runs where you placed it: it is up to you to put a confirmation step before it.
Its answer is compared with your documents
By default, when the knowledge base was consulted before the answer, an anti-hallucination safety net compares the answer with the passages found and can remove the sentences they do not confirm. If nothing is left, the agent says it does not have enough information to answer. An administrator can adjust this net agent by agent, down to switching it off. It lowers the risk of invented answers; it does not rule it out.
The answer is read once more
The exit guardrails read the answer: toxic language and profanity by default, personal data if you switch it on. It goes back on the channel the question came from. An answer that breaks one of them is replaced with a fixed message: before sending on messaging channels; just after it appears in the widget, where the text appears as it is written.
When needed, it hands over
A handover is triggered when the customer asks to talk to someone, when they express psychological distress or a medical emergency, when they show frustration, or when the agent fails several times in a row. You set how sensitive those last two triggers are and choose who is told: members of your team, by email with a link to the conversation, a Zendesk ticket, a Salesforce case or a webhook to your own tool. The agent tells the customer their request has been passed on, with the reply time you gave. With no handover configured, it says so and keeps helping.
No code, no developer
Your AI agent live, in four steps.
No technical skill required. The one line of code in the journey is handed to you, ready to copy.
Start from a template
Customer service, reception and appointments, software support, lead qualification: four complete templates, ready to tailor. Or a blank page, if your line of work fits none of them.
Give it your business
Your documents become its knowledge: PDF, Word, Excel or CSV files, a URL, a sitemap with its exclusions (500 pages per import, within your plan's quota). They belong to that one agent — your other agents never see them.
Try it before your customers do
A built-in test space: you talk to the agent, adjust its instructions, reload a document, start again. As many times as it takes, with nothing to install.
Put it live
One line to copy onto your site, and it answers your visitors. The same agent, with the same knowledge base and the same escalation rules, then serves your other channels.
The four steps, in detail
What you do on screen, what the agent gets out of it, and what is worth knowing before you start.
Step 1
A template, or a blank page
A template is an agent already written: its mission, its welcome message, the questions it suggests, its reply to off-topic requests. You fill in its variables — your company's name, your contact email, your opening hours — and it becomes yours: everything in it can still be changed. If no template fits your line of work, start from a blank page: the agent's name, your company, its language and its mission in a few sentences, the way you would explain it to a new colleague.
- Four official templates: customer service, reception and appointments, software support and lead qualification.
- A template is connected to none of your tools: calls to your APIs and the handover to your team are set afterwards, in the agent.
- Allow about fifteen minutes, once the account is created, for an agent started from a template; the real timeline depends mostly on the state of your documentation.
Step 2
Your documents, as they are
The agent answers from what your team has already written: terms of sale, FAQ, welcome guide, product documentation. You import these documents without rewriting them for it: each one is split into passages and indexed in the background, and the agent uses it as soon as indexing is done. The more precise and up to date they are, the better it answers: of all your settings, this is the one that matters most.
- PDF, Word (.docx), Excel and CSV files, plain text, HTML or Markdown, up to 20 MB each. Only the text of a PDF is read: a scanned page adds nothing. PowerPoint presentations are not accepted.
- A URL imports one page; a sitemap, up to 500 pages per import, with its exclusions and within your plan's quota. Imported pages can be read again at regular intervals, from once an hour to once a week.
- Each document belongs to the agent you import it into, in a collection specific to your organisation. Archiving it removes it from the answers without deleting it.
Step 3
Tests, until it answers well
The test space is a conversation with your agent, in the dashboard, with nothing to install. Ask it what your customers really ask: the most frequent questions, then the trickiest — a complaint, an off-topic request, someone who wants to talk to a human. For each answer, you see what the agent did: its searches in your knowledge base, its tool calls, the guardrails triggered. When an answer falls short, the cause is nearly always one of two: an instruction to sharpen, or a missing document.
- Tests are not counted against your monthly conversations: they have their own allowance of 200 messages a day.
- Your typical questions can be saved as test cases, with what the answer must contain or avoid: you run them again after a change.
- The agent's settings apply as soon as they are saved: the agent has no draft version, only journeys do. Once the agent is live, ask your questions again after any significant change.
Step 4
The widget first, your other channels next
For your website, the dashboard gives you a tag to copy and paste into your pages: it is the one line of code in the whole setup, and the standard job of a web developer. You set the widget's look with no code — title, welcome message, quick replies, colours, logo, position — and the list of sites allowed to display it. Messaging channels then connect from the Integrations page, with nothing to set up again: the agent keeps its knowledge base, its journeys and its escalation rules.
- The web widget is included in every plan, Free included. With several agents, from the Pro plan, you choose which one appears on which page.
- WhatsApp, Messenger, Instagram and Telegram connect from the Pro plan: a Meta Business account and the credentials of your WhatsApp Business number for WhatsApp; a Facebook Page for Messenger and Instagram (and, for Instagram, a professional account linked to that Page); a bot created with @BotFather for Telegram.
- The REST API opens a channel of your own making from your server, with a key specific to each agent.
Once it is live: measure, then enrich
An agent is not set up once and for all. Real conversations show what your documents cover poorly; the dashboard tells you where to look.
Follow the activity
The Analytics screen shows, over 7, 30 or 90 days and with a filter by channel, the number of conversations, their average duration, the topics that come up most, satisfaction and the sentiment of the conversations analysed.
Find what is missing
The Quality page groups the recurring questions your knowledge base could not answer: in practical terms, the list of what to write next. The agent does not correct itself: you decide what to add.
Adjust with nothing to reinstall
A sharpened instruction or a changed escalation rule applies as soon as it is saved; an added document, as soon as it is indexed. On every channel of the agent, without touching the tag on your website.
What stays on your side
Winaity AI spares you the development, not the groundwork. Here, plainly, is what to plan for.
- Up-to-date documentation
- It is what the agent reads in order to answer, and nobody can write it for you. The real setup time depends mostly on its state.
- Someone to connect your channels
- Pasting a tag into your website's pages is a routine job for whoever manages them, in your team or at your web agency. Connecting WhatsApp also takes someone with access to your Meta Business account; Messenger and Instagram, someone who manages your Facebook Page.
- Someone who knows your APIs
- Declaring a call takes no code, but you need to know which address to call, with which key, and what it returns. There is no ready-made connector for Shopify, Jira or HubSpot: you declare the call to their API yourself.
- A team to follow up after the handover
- The agent notifies your team, it does not replace it: your team then takes up the request, from the link to the conversation or the ticket created, with its own tools.
Ready to deploy your AI agents?
Create your account for free. 30 conversations a month, forever.
1 free AI agent
· Designed in France · Data partitioned per organisation ·