AI and handoff architecture

Website chat AI: a safe human handoff in Microsoft Teams

A useful AI needs a reliable exit. Uncertainty, sensitive topics and buying intent should trigger a visible, tested handoff.

Fact-checked: 7 Verified sources

Control the system

Five layers matter more than a model name

  1. 01

    Approved sources

    Version product, help and policy content that the assistant may use.

  2. 02

    Answer policy

    Define uncertainty, prohibited topics, citations and personal-data rules.

  3. 03

    Handoff thresholds

    Transfer buying intent, complaints, sensitive data and repeated uncertainty.

  4. 04

    Teams context

    Pass the question, relevant summary, language and minimum visitor context.

  5. 05

    Quality loop

    Review bad answers, transfers, abandonment and human corrections.

Different product approaches

Own Microsoft stack, packaged AI or an existing bot

ApproachDocumented candidatePrimary check
Own Microsoft AIWebChat with Foundry and Power AutomateAzure, Premium and consumption costs plus operating ownership
Packaged modelsSocial Intents with ChatGPT, Gemini and Claude by plantraining, conversation and privacy limits
Virtual assistantsuWebChatmodel, data flow, translation and handoff
Existing bot platformCaptivate with GPT, Watson and Amazon Lexversion, context transfer, compatibility and price

Sources: 4, 5, 6, 7

Handoff contract

Give the human enough context, not every possible datum

  • original question and a relevant conversation summary
  • sources used or the reason no source was found
  • handoff reason and urgency
  • language, entry page and voluntarily supplied contact details
  • a clear status for visitor and agent
  • the ability to stop the AI and correct the record

Sources: 1, 3, 4

Acceptance test

Test difficult questions before launch

  • outdated prices and features that do not exist
  • ambiguous, multilingual and intentionally false premises
  • personal, medical, legal and abusive content
  • prompt injection and requests for internal information
  • buying intent, cancellation, complaint and urgent escalation
  • knowledge-source, AI or Teams connection failure
FAQ

Frequently asked questions

Can WebChat use an organisation's own AI?

The setup guide documents an approach using Microsoft Foundry, Power Automate Premium and Azure pay-as-you-go. Plan for prerequisites, ownership and extra cost.

What is human handoff?

A controlled transfer from an automated conversation to a person, with relevant context and a clear status for both visitor and agent.

Which AI is best for website chat?

That depends on control, approved knowledge, privacy, handoff, languages, cost and operations. A model brand is not a quality guarantee.

Can AI qualify leads automatically?

It can structure questions, but transparent criteria, data minimisation, human escalation and legal review are required.

Verified sources

Sources and review date

  1. 01
    Einrichten und Verwalten von Livechats

    Microsoft · Fact-checked:

    Open source
    verified
  2. 02
    WebChat für Microsoft Teams

    inwebco GmbH · Fact-checked:

    Open source
    verified
  3. 03
    Wirklich nativ in Microsoft Teams – statt nur eingebettet

    inwebco GmbH · Fact-checked:

    Open source
    verified
  4. 04
    Setting Up WebChat: The Complete Guide for Microsoft Teams

    inwebco GmbH · Fact-checked:

    Open source
    verified
  5. 05
    AI-Powered Live Chat for Microsoft Teams, Slack, and Google Chat

    Social Intents · Fact-checked:

    Open source
    verified
  6. 06
    uWebChat License

    Microsoft Marketplace / Universal IT B.V. · Fact-checked:

    Open source
    review note
  7. 07
    Captivate Chat

    Microsoft Marketplace / Captivate IO Ltd · Fact-checked:

    Open source
    review note