The agent operating system
Every business has a gap between the person who talks to customers and the person who knows what's actually going on.
Wetel closes it. One agent — with a face, a voice, and real knowledge of what's happening inside your company — that serves the outside: the customer on your site, the candidate on Telegram, the lead who called at 11pm. And when it's about to do something that costs you something, it stops and asks.
Most AI agents are a voice on top of an FAQ.
The front desk smiles but has no idea whether the workshop is backed up. The chatbot answers questions but has never seen the calendar. The screener collects résumés but can't tell a candidate whether the role is still open. So the real work falls to a human — someone copy-pastes between the tool that talks to customers and the tools that know the truth. Someone is the router.
An agent that only knows what a PDF said last quarter is a search box with a personality. An agent that knows what's true today can be trusted to speak for you.
Wetel gives a business a teammate that shows up — and that asks before it speaks on your behalf.
Seven roles you haven't hired yet.
All of them are the same agent, pointed at a different job, sharing one picture of the business.
Your AI recruiter
Runs first-round screening end to end — on whatever channel the candidate arrived through. Takes the résumé, files it to your ATS, books the ones worth meeting. It knows which roles are open now.
You meet four good people this week instead of reading sixty CVs.
Your AI front desk
The face on your homepage. Answers in your voice, from your real policies, and knows when the answer is “we can't do that.” Books, reschedules, routes.
Nobody waits until Monday for an answer you already have.
Your AI support rep
It answers the ticket — and it also knows the shipment is delayed and the workshop is full. So it stops promising Thursday.
Fewer apologies, because fewer wrong promises.
Your AI ops brain
The one facing inward. Watches bookings, capacity, backlog — and speaks up when asked. “You've said yes to three things next Tuesday.”
You stop finding out about the crunch during the crunch.
Your AI sales concierge
Handles the 11pm enquiry. Qualifies, quotes from real pricing, checks whether you can actually deliver on that date — and when it needs a decision only you can make, it asks.
Deals don't cool off overnight waiting for you.
Your AI coach
Role-plays the hard conversation with a new hire, face to face, as many times as they need. Then tells them — and you — how they did.
Training that scales past whoever was free that afternoon.
Your AI account manager
The one who remembers. Follows up when a follow-up is due. Notices the renewal before the renewal notices you.
The quiet revenue you keep losing to being busy.
Why this holds up when the conversation leaves the script.
Four of these are the reason the agent is trustworthy. Two are the reason it's still yours in three years.
A workflow you can look inside
Behaviour is a visual graph, not a prompt you hope holds. Follow the path: here it checks stock, here it branches if the candidate has no résumé, here it hands to a human. When it does something surprising, you can see exactly what it did — every run traced, node by node.
See how workflows workInterruption that actually works
It stops mid-sentence when you start talking. Genuine barge-in, not a queued turn that finishes its paragraph while you're already speaking over it. The difference between a walkie-talkie and a conversation.
Voice and turn-takingActions from a vetted menu
The agent can't invent a new way to touch your CRM. It acts through pre-reviewed operations with a known contract, known failure mode, and known blast radius. Credentials are resolved on our servers and never placed into the model's context.
Certified OperationsIt knows today, not last quarter
Per-tenant knowledge, scoped and isolated, plus live reads from the systems that actually hold the truth — the calendar, the stock list, the pipeline. Not a PDF upload with a personality on top.
Knowledge and retrievalYour data in its own database — if that's what it takes
Shared-mode tenancy with row-level isolation enforced at every layer, or silo mode: a dedicated physical database per tenant, provisioned on demand, with its own credentials. The sentence that survives a security review.
How tenancy worksOne agent, every surface
The same agent definition, the same graph, the same knowledge, the same governed actions — presented as a 3D avatar, a voice line, a chat widget, a Telegram bot, or invisibly inside your own product. Embodiment is a rendering decision, not a rebuild.
Agents and surfacesThe canvas is the product, not a wrapper around a config file.
Nearly every “workflow” in this market is one of two things: a linear pipeline with no real branching, or a fully autonomous reasoning loop where the model decides everything and you pray. The first can't model a real business process. The second can't be reproduced, can't be audited, and terrifies anyone with a compliance function.
Wetel puts deterministic structure around non-deterministic reasoning. The model decides which branch. The graph decides what is even possible.
Draw the behaviour
Intent routing, tool calls, knowledge lookups, human handoff — laid out as a graph with labelled success and failure edges, not a wall of prompts.
Test it as a draft
Publish is a deliberate act. A draft runs against real conversations before anyone else sees it, and every run leaves a per-node trace you can read afterwards.
Ship it to any surface
One script tag, or a handful of GraphQL calls, and the same agent is live — on your site, in a voice line, or inside your own product under your own brand.
Small stuff, it just does. Big stuff, it asks.
That's the honest answer to “aren't AI agents dangerous?” They're dangerous when they're all-or-nothing — either lobotomised into uselessness, or free to invent a 40% discount at 2am. This is neither.
It can only do what someone already approved.
Actions come from a reviewed catalog, not from handing a model an API spec and hoping. Each one is a contract: this call, these arguments, this failure mode, this credential scope.
Everything it did is on the record.
Every run is traced node by node — the rendered input, the output, the branch taken. Conversation quality is scored. Spend is capped per tenant. Nothing is a black box after the fact.
Consequential things come to a human.
Look something up, answer a question, log a note — it just does it. Money, a promise, a person's record, anything irreversible — that stops and waits for a person.
Approval routing across a team and swipe-to-approve on mobile are in active development. Today the same boundary is enforced through the operation catalog, credential scoping, spend caps, and admin override.
One agent. Wherever the conversation is.
Add a surface without rebuilding the agent — same graph, same knowledge, same permissions.
3D avatar
Real-time presence that looks at the customer while it talks.
Voice
Natural speech with true mid-sentence interruption.
Embedded web
One tag on your own site.
Telegram
Live messaging, already in production.
Partner SDK
The invisible brain inside someone else's product.
More surfaces
Each new surface is an adapter over a stable agent contract — and lands for every existing agent the day it ships.
For platform teams
If you're not buying an agent — you're building on one.
Wetel is infrastructure: the runtime, the governance layer, and the control plane other companies ship agent products on top of. The right mental model is the payments processor, not the checkout page. That isn't a roadmap statement — partner products in customer service, education, and recruitment already run on it today, with no Wetel branding anywhere in them.
ISVs and platforms
You need agent capability inside your own product, under your own brand, for your own customers. You get a stable, versioned SDK contract, programmatic tenant provisioning, credential isolation per your customer, and drift detection in both directions. We will not compete with you for your customers.
Agencies and implementation partners
You deliver agent solutions for clients and need to reuse patterns across engagements. Reusable workflow patterns per client, multi-tenant management from one console, white-label branding, and a certified-operation catalog so you're not re-litigating the safety of every integration on every project.
Enterprise platform teams
You're standardising agent capability across business units. Dedicated-database isolation per tenant, SSO, audit trails, per-unit spend controls, and a clean answer to “what did the agent do, when, on whose authority.”
Nothing gets thrown away. The agent built alone on a Sunday is the same agent serving fifty thousand conversations two years later — same runtime, same graph. Nobody graduates off a starter tool.
Running in production today.
Real external integrations generating real sessions — not pilots sitting unused after a demo. Names withheld at partner request; the technical details are unaltered.
A university library assistant
RAG-grounded Q&A, room booking and loan renewal, across two cloud regions behind a white-labelled gateway.
A customer support platform's embedding layer
A third-party helpdesk product consuming Wetel's embedding pipeline directly.
An HR screening assistant
Conversational candidate screening with a swappable digital-human avatar backend.
An insurance claims coordinator on WhatsApp
A real-time multi-party conversation state machine, not a single-user chatbot script.
Or open one yourself — every demo on the docs site is a live application built against the exact same API.
Everything is documented. Including what isn't built yet.
A full integration guide, a complete GraphQL API reference generated from the live schema, per-node workflow reference with worked examples, and an honest roadmap page. Plus a drop-in SKILL.md if your team codes with an AI assistant.
Talk to us
Tell us what you're building. Whether you're evaluating an agent for your own business, embedding Wetel inside your product, or standardising agents across an enterprise — we'll reply with a real answer, not a drip sequence.