TensorFrog.it
PLEN

Capabilities · Products & integrations

Chatbot delivery — conversation, knowledge, human handover and API

The scope is a chatbot that answers in agreed topics, reaches company knowledge and systems, and hands the conversation to an operator when the automaton is not enough. Other systems call it through an API. The widget sits on the website. TensorFrog delivers software and integrations — it does not run a helpline and is not an SMS or mail operator.

Chatbots

When a chatbot is a process module

The site, mail and the phone answer the same questions in three places. The agent cannot see what the client already heard. The FAQ on the site ages beside the inbox. A chatbot makes sense when there is one topic scope, one source of truth and a clear line: what the bot does not decide. First an assessment: whether a widget with an answer base is enough, or whether CRM, tickets and mail must be wired. A decoration in the corner of the page with no log and no handover is not this scope.

How we will judge the effect

We agree the goal and how we will measure it before work starts. We do not promise percentage savings without a baseline.

  • Share of conversations closed in the bot's scope — without handover when the topic was in the dictionary.
  • Time from a request for a human to a case card that already holds the conversation history.
  • Share of answers grounded in a named source — a document or a system record, not a model guess.

Conversation flows and answer scope

A conversation has an intent, a step and an end. Typical first paths: order status, opening hours, a document from the register, a complaint, handover to a human. Outside scope the bot refuses and offers an operator — it does not guess. A template and a rule where an error must not pass (an amount, a payment status). A model where the question is free-form and the answer still lines up with a record in the system. Language and tone are set by the company. A path change goes through sign-off before production.

Knowledge grounding

An answer rests on a named set: the site, a policy, a chosen document type, a CRM record or a ticket status — when that is in scope. A quote or a source id when it can be shown. No source means no certain answer, not a guess. Documents from the document register or from a base the company names. A public model does not receive the protected layer. A hallucination goes to the log and to a path fix — it is not treated as creativity.

Human handover

Handover threshold: the client asks for a human, sentiment, a topic outside scope, a financial decision or no certain source. Handover opens a ticket or attaches to a case: conversation text, intent, client id. The operator sees the history before answering. The bot does not close a complaint and does not promise a refund. The ticket desk and mail stay where the human works — the chatbot is an intake, not a second inbox.

API integrations — CRM, tickets, mail, payments

CRM, order, tickets, mail and payments — they call and are called through an API contract: event type, session id, client, intent, field data. Order or payment status the bot shows from the system, not from the model's memory. PayU, Przelewy24 and ING imoje stay the operators; the payment layer is on the payments page. Mail and SMS — when in scope — go through the existing correspondence platform, not through a separate chatbot gateway. Each system: its own key, limit and log.

Channel: a web widget

The first channel is a widget on the website: a conversation window, a session id, handover to an operator. No iOS app is built. Android — only when a later stage provides for it, in the application scope. WhatsApp Business API — when the company holds its own account and names it after discovery; TensorFrog is not a messenger operator. Mail and SMS from the chatbot do not start from a separate inbox: they go through the mail system when that is in scope.

Local or hosted model

When the conversation holds personal data, secrets or regulated content, inference runs on a local model — the same rule TensorFrog applies to sensitive code and protected data. A public cloud model does not receive that layer. Infrastructure hosting is not the same as model insight. A ready API, an adaptation or a domain model — the decision follows cost, data and risk. Training and model upkeep are on the AI-models page.

First stage

One topic, a website widget, one answer base or one API (for example order status) and handover to an operator. A session log: intent, source, whether it went to a human. History retention follows an agreed policy. Commercial information in mail or SMS needs consent per channel (PKE); clause wording stays with the company and its counsel. TensorFrog is not an SMS operator. A second channel, a second system or a local model for protected data — a separate quote after sign-off.

Related areas

AI agentsAI modelsTicketingMail and SMSAPI
Products & integrations

Describe your business need — we will propose the right technology and the next work stages. Contact: [email protected]

Go to contactView all capabilities