The AI chat widget nobody documented

AI chat widgets can look like a simple way to answer questions, qualify leads or help visitors find the right page. But the conversation may include names, email addresses, account details, business plans or descriptions of sensitive problems. If the site does not explain where that input goes, visitors cannot make an informed choice about what to share.

In OrangeStealth’s monthly refresh dated 20260929, 4.1% of the US B2B SaaS 5-50 population, with a sample size of N = 1094, had an AI chat interface without a clear disclosure about how visitor input was handled (n=45). This is a public documentation gap. It is not proof that any company misused data, suffered a breach or exposed information.

The missing explanation matters

A conversational interface feels private. Visitors may write to it as if they were speaking with an employee who understands the company’s confidentiality rules. Behind the interface, however, several systems may be involved: the website operator, a chat vendor, an AI model provider, analytics services and a customer relationship platform.

The important questions are practical. Is the conversation retained? Is it linked to an identifiable visitor? Can a vendor use the content to improve a model? Which providers receive it? How long is it stored? Can a user request deletion? The answers may exist in contracts or internal settings, but visitors need a clear public explanation before they submit information.

A vague privacy policy does not necessarily solve the problem. A policy might discuss cookies, contact forms and analytics without identifying the AI chat tool or describing its data flow. Likewise, a short warning inside the widget may tell users not to enter sensitive information while leaving retention, sharing and training practices unexplained.

A hypothetical conversation

Imagine a visitor asking whether a software product can support a confidential customer workflow. To get a useful answer, the visitor pastes part of that workflow into the chat. The widget sends the message to an outside provider, stores a transcript and creates a lead record for follow-up.

None of those steps automatically means the system is unsafe. The problem is that the visitor may not have expected them. The business also may not have documented the same flow consistently across its public notice, vendor agreement, internal data inventory and deletion process.

This kind of mismatch can become a security, privacy and vendor-management issue at the same time. It can also complicate customer questionnaires and third-party risk review because the company cannot readily explain which party receives conversational data and under what controls.

How we measured it

This research used passive, public-data-only observations from the US B2B SaaS 5-50 population, with a sample size of N = 1094. The aggregate was generated on 2026-09-29 as part of monthly refresh 20260929. Results are k-anonymised, and no company is named. We did not log in, submit prompts, test private systems or attempt to determine what happened behind the public interface.

That limitation matters. A missing public statement does not prove that backend safeguards are absent. It shows that a visitor cannot readily verify the handling of chat content from the information presented publicly.

What to actually do about it

• Map the complete conversation path. Identify every service that receives chat content, associated identifiers and generated summaries. Include integrations that copy transcripts into support, analytics or sales systems.
• Check the vendor settings and contract. Confirm retention, model-training use, access controls, deletion support, subprocessors and incident-notification terms. Do not assume the default configuration matches your public claims.
• Add a clear notice at the point of entry. Tell visitors what receives their input, the purpose of processing and what they should avoid sharing. Link to a fuller policy that names the relevant service categories and explains retention and user choices.
• Test your own deletion process. Make sure the team can locate and remove a conversation across the widget, connected platforms and retained exports when policy or applicable obligations require it.
• Review the public surface regularly. The free OrangeStealth public checker can help identify visible issues. For a broader passive review of externally observable exposure, an External Security Posture Assessment can place the finding in context without treating it as proof of compromise.

If an AI chat tool is live on your domain, open it as a first-time visitor and read only what the visitor can see. Then compare that explanation with the actual vendor settings and data flow. Any difference between those two views is the place to start.