Can the assistant hand over to a person?

Yes, and we treat it as a core feature. You set the triggers, such as low confidence, a named topic or a request for a person. The person gets the transcript and customer details, so the customer does not repeat themselves.

How do you stop it making things up?

It answers only from content you approve and is told to say it does not know rather than guess. Each answer links to its source, and every release runs against a test set of real questions.

Can it connect to our help desk and CRM?

Usually, through the APIs those systems provide. It can look up an order or ticket, create a case or update a contact within permissions we agree. Read-only access comes first.

Can it work in more than one language?

Yes. Quality varies between models and languages, so we test candidate models on your own questions in each language before choosing one.

What is the difference between a chatbot and an agent?

A chatbot holds a conversation: it answers, collects details and passes work to people. An agent carries out multi-step tasks on its own. Many projects start as a chatbot and add a few agent-style actions once answers are reliable.

Assistants that answer from your content, and know when to hand over

Customer and staff chat that cites approved sources and passes the conversation to a person before it goes wrong.

Site chat is one channel for inbound leads. See SaaS inbound lead handling.

What we build for chat

Most projects start with one channel and one audience. We agree the scope, the hand-over rules and the test set before building.

Customer service assistants
Order, account and policy questions answered from your help content and systems, each with a link to its source.
Internal knowledge assistants
Staff ask in plain language and get an answer from your policies and wikis, limited to what they may already open.
Hand-over to people
The agent who takes over sees the transcript, the customer's details and what the assistant already tried.
Several channels
A website widget, a messaging app, Teams, Slack or a chat screen in your own app, all sharing the same knowledge and rules.
Languages and voice
Replies in the language the customer writes in, and speech in and out where phone calls matter more than chat.
Analytics and guardrails
A view of what people asked and where the assistant handed over. Unanswered questions become your content backlog.

How an assistant handles a conversation

Behind a good assistant is a router: it screens the message, works out what is asked, sends it down the right path and checks the reply before it goes out.

Conversation

For every message

Missing details: ask a follow-up.

Hand-over

When the assistant should not continue

Improve

On a regular rhythm

Step 1 of 13System work

Message in

From your website, app or messaging channel, in the customer's language.

In practice: Multi-channel intake
You get: One conversation history per customer

A conversation, and what runs behind it

A scripted exchange. Under each reply you can see the checks, the intent, the route taken and the moment it hands over to a person.

Visitor: Where is my order?
Agent: I can check that. What is your order number?
Input checked / Intent: order status / Missing: order number
Visitor: ORD-1042
Agent: Order ORD-1042 has left the warehouse and is due on Thursday. Here is your tracking page.
Route: order lookup as the signed-in customer / Reply checked / Source: order record
Visitor: It is late. Can I get a refund?
Agent: Refunds are decided by a person, so I am passing this to the support team now. They will see everything we covered.
Rule: refunds go to a person / Hand-over
Handed to support team
- Order ORD-1042, due Thursday
- Customer says it is late
- Asked for a refund
- Tracking page already shared

Visitors cannot type here. This is a scripted exchange; the line under each reply shows what ran behind it.

Illustrative demo. Scripted, not a live system.

How we keep assistants safe

An assistant speaks for your business. These limits are set before launch.

Approved content onlyGroundedAnswers come from sources you approve, and the assistant says when it does not know.
Hand-over rules you writePersonComplaints, money, sensitive topics and direct requests go to a person with the whole conversation.
Input and output guardsCheckedPersonal data is masked, injection attempts blocked and replies checked before they are sent.
A test set of real questionsTestedEvery release is run against questions your team actually gets.

Tools we commonly work with

We choose channel, model and help desk wiring to suit your customers. Naming a tool is not a partnership or an endorsement.

Models: Claude
GPT
Gemini
Llama
Mistral
Channels: Website widget
WhatsApp Business Platform
Microsoft Teams
Slack
In-app chat
Help desk and CRM: Zendesk
Intercom
Freshdesk
HubSpot
Salesforce
Retrieval: pgvector
Qdrant
Elasticsearch
Azure AI Search
Voice: Whisper
Deepgram
ElevenLabs
Twilio Voice

Common questions about this work

Tell us what you want built. We will scope it with you.

Start an AI project

Assistants that answer from your content, and know when to hand over

What we build for chat

Customer service assistants

Internal knowledge assistants

Hand-over to people

Several channels

Languages and voice

Analytics and guardrails