Be transparent about the model and data. Answers from your docs. Team replies in Slack or Telegram.

The support questions that carry trust

A customer may ask whether a prompt is stored, whether a feature uses a particular model, or whether generated output is reviewed by a person. A generic chatbot answer can damage trust if it overstates your retention or safety guarantees. The support assistant needs a source that your product and privacy owners have approved, and a visible route to a person for questions that the page does not settle.

Examples include “Which model powers this feature?”, “Are our conversations used for training?”, “Where is support data hosted?”, and “Why did the assistant return this result?” The first three may have a public policy answer. The last may depend on the prompt, model version, and account settings. Keep product facts separate from an investigation that needs access to the customer’s environment.

• Publish model, retention, and data handling statements with an owner and review date.
• Make limitations and escalation routes clear next to AI feature docs.
• Do not treat an explanation of expected behavior as an investigation of one output.

A restrained support model for AI products

MivoChat uses Claude through Anthropic’s API, is hosted in Frankfurt, and does not train on your data. For your own product’s model choice, retention policy, and safety behavior, connect the canonical pages rather than asking the assistant to infer company policy. It can answer only what the connected source supports. If it cannot establish the answer, it should say “I don’t know” and hand off.

The teammate receives the unresolved conversation in a Slack thread or Telegram topic. They can take over, review the exact visitor message, and respond in the channel while the customer stays in the website widget. Once the concern is addressed, the teammate can hand the conversation back to AI or close it. This gives product and trust teams a place to own sensitive questions without turning every routine policy lookup into a meeting.

Understand the allowance before forecasting volume

Starter is $29/month and includes 1,000 AI replies, 3 seats, and one site. Pro is $99/month with 10,000 AI replies, 10 seats, and 5 sites. MivoChat counts each AI reply against the limit. The default estimate is 3 replies per AI-handled conversation, so Starter’s allowance equates to about 333 such conversations.

At $0.99 per resolved conversation, Fin would charge about $329.67 for 333 resolved conversations, before seats. This comparison uses the default MivoChat conversation length; a short exchange may use fewer replies and a long one more. MivoChat has no per-resolution fee, and replies above an allowance do not trigger a surprise model call: they are handed to a person. Paid billing is coming soon, and the no-card trial lasts 14 days.

Connect the public product and trust center

Place the widget on the public product docs, model overview, privacy page, and troubleshooting guide. For a Next.js App Router site, add it through the shared root layout using the supported script placement, using next/script. Keep a clear distinction between public product education and authenticated application areas. A standard site install uses: <script src="https://mivochat.app/widget.js" data-site="YOUR_SITE_KEY" async></script>.

Test language that could be interpreted as a guarantee. Ask whether data is used for training, whether a specific model powers a feature, and what happens to an unusual output. Compare every answer with the approved page. If a policy is absent, do not rely on the bot to invent one; publish the policy or let the human team answer. Revisit the connected pages when model vendors, subprocessors, or product behavior change.

Keep sensitive investigations human

MivoChat does not read private prompts or account data, inspect model traces, change settings, delete a user’s records, or take product actions. The support assistant is a first response based on connected pages, not an observability tool or a privacy decision-maker.

Route suspected safety incidents, data deletion requests, unexpected model behavior, and contractual questions to someone who can investigate. Do not ask a visitor to paste confidential prompts or personal data into a public widget. The assistant can acknowledge that it does not know and pass the thread along; a named teammate still needs to own the follow-up.