How we use AI

AI helps in Gauntlet and UniCast, but people stay in charge: your team decides what the bot knows, can take over any conversation, and approves anything that a check holds back.

Your website chatbot

• Answers from your own sources: the bot answers from the knowledge you give it (your website pages, files and answers you write). It is told never to invent prices, features, dates, discounts or commitments, and to say honestly when a person will follow up instead.
• A real person, one click away: visitors can ask for a person at any time. The bot also hands over when it cannot answer, when a message is very negative, or on words you choose. After hand-off the bot stays silent; your team sees why it was handed over and what the bot used.
• No person free? Visitors can leave a voice message or ask for a call-back.
• Disclosure: AI replies appear under your bot's name, separate from your team members' names, and the visitor is told when they are back with the assistant. You choose the bot's name and welcome message; we recommend making clear that it is an automated assistant.

Testing bots before they speak

Gauntlet sends hostile AI "customers" at a bot (trying to extract personal data, slip it new instructions, refuse it a person, or make it promise things it should not) and gives a report card with the exact sentence that failed. We test our own website bot with Gauntlet too.

Video: checked before anyone watches

• Every UniCast video is checked before it can be published: frames for nudity, and the title, description, transcript, comments and live chat for hate, harassment, violence, self-harm and other harmful content.
• The rules run first; AI can only make a verdict stricter, never more lenient. The word rules always check every word. When the AI check is switched on and could not finish checking a video, it is never passed automatically; it waits for a person in your organisation to review it.

Captions and translation

• Transcripts, captions, translated captions and live webinar captions are made by AI. They can contain mistakes, and you can correct captions.
• "Listen in my language" is labelled as an AI voice, so viewers know it is not the speaker.

Your data and AI

To answer, the text of a conversation is sent to an AI provider or to our own model. The providers and their locations are listed on the data and sub-processors page. Personal data is never used to train our own model.

Rules these features help with

Gauntlet and UniCast help you meet these rules; they do not make you compliant on their own, and the decisions stay with you:

• India's DPDP Act: personal-data tests for your bot, erasure of a person's records, and a processor agreement.
• EU AI Act transparency duties for chatbots: your bot's name and welcome message are where you tell visitors they are talking to an AI, and Gauntlet tests that the bot hands over to a person when asked and does not invent facts.
• European Accessibility Act, ADA Title II (United States) and UGC accessibility guidelines (India): captions, live and translated captions and an accessible player. See our accessibility statement for known gaps.