End User AI Chat is an AI support assistant inside the WorkBright app. It answers common onboarding questions from new hires, employees, and authorized representatives while they complete onboarding. Users get answers right away, without waiting on an email reply or leaving the app.
This article explains what the assistant does, how it handles data, how WorkBright keeps its answers from being read as legal advice or employer policy, and the security controls behind it.
The feature runs on Intercom's Fin AI Agent, the same chat technology behind WorkBright's admin support chat. Answers use retrieval augmented generation (RAG), which bases every response on WorkBright's approved knowledge content and USCIS Form I-9 guidance.
What the assistant does and does not do
The assistant gives procedural guidance for completing onboarding in WorkBright. It does not make decisions for your organization.
What it does:
Answers common onboarding questions. This includes document guidance, Section 1 field help, Section 2 support for authorized representatives, process and status questions, and next steps.
Tells users it is AI. The assistant identifies itself as an AI agent at the start of the conversation.
Warns users not to share sensitive information. A disclaimer tells users not to enter personal or sensitive data in the chat.
Escalates instead of guessing. Questions that are unclear, policy-specific, or legally sensitive go to your organization's designated contact.
What it does not do:
It does not give legal, financial, or medical advice.
It does not make eligibility, creditworthiness, or similar determinations.
It does not speak for your organization on policy questions.
Guidance on how to complete Form I-9, based on USCIS instructions, is not the same as legal advice about a specific situation. The assistant only provides the first.
How the assistant answers questions
The assistant only answers from content WorkBright has approved. When it can't find a supported answer, it says so and can hand the conversation to a person.
What happens when a user asks a question
The question goes to Intercom's backend, which gathers the relevant context.
Intercom's machine learning service combines the question with content from the WorkBright Knowledge Hub.
The combined prompt goes to a third-party large language model (LLM) provider, which generates a response.
Intercom checks the response, then sends the answer to the user in the chat.
WorkBright does not train or fine-tune the underlying LLMs.
Safety checks before a response is sent
According to Intercom, Fin runs these checks on every question:
Rewrites the question for clarity.
Pulls content from the knowledge sources WorkBright has turned on.
Ranks that content by relevance.
Writes an answer based on that content.
Runs a final check for accuracy and safety.
If the answer doesn't meet Fin's confidence thresholds, the assistant sends the user to a person instead of guessing. Intercom reports a hallucination rate below 1% for Fin.
What "WorkBright content" includes
WorkBright content is the knowledge WorkBright selects and turns on in the Intercom Knowledge Hub. The assistant can only answer from this content. It includes:
WorkBright platform guidance. Help center articles on using WorkBright, such as completing onboarding steps, filling out and resubmitting forms, uploading documents, account access and one-time codes, and kiosk mode.
Process and how-to content. Snippets and process documentation that WorkBright teammates add to the Knowledge Hub.
USCIS Form I-9 guidance. The Form I-9 instructions and the USCIS employer handbook, as selected by WorkBright.
WorkBright content does not include your organization's own policy documents, any user's personal data, or the language model's general training data. When a source changes, the content is reloaded so the assistant answers from current material.
What gets escalated to your team
The assistant sends questions that need human judgment to your organization's designated contact. It does not answer them. These include:
E-Verify Tentative Nonconfirmations (TNCs) and Final Nonconfirmations (FNCs)
Legal exceptions
Reverifications
Workflows specific to your organization
Any request that requires a user's personal details
Your HR team keeps final authority over I-9 and onboarding decisions.
How end user data is collected and stored
Third-party AI providers do not keep chat messages or use them to train models. User messages and WorkBright content are the input used to generate a response, and the response is the output. Both are treated as Customer Data.
Personal data in messages. If a user includes personal data in a message, the AI provider processes it only to generate the response. The provider keeps nothing after the response is returned. Intercom keeps the conversation record for logging.
Zero Data Retention. Intercom's AI providers delete inputs and outputs from their systems once the response is generated.
No training on customer data. Intercom's contracts prohibit its AI providers from using inputs or outputs to train or improve their models or services.
Safety classifiers do not store data. AI providers may run safety classifiers and review prompt metadata to enforce their usage policies. Per Intercom, these processes do not store Customer Data or review personal data in the prompt.
Access controls and logging. Access to LLM environments is restricted and logged. Intercom treats its LLM providers as Tier 1 subprocessors, with periodic user access reviews and annual audits under ISO standards.
No pre-processing redaction. Intercom does not currently remove personal data before it reaches the AI provider. WorkBright covers this with the in-chat disclaimer, escalation of requests involving personal details, and Zero Data Retention with no model training.
Conversation logs on request. WorkBright can provide your organization's chat history when you ask for it.
How WorkBright keeps answers separate from policy and legal advice
The assistant is set up so users don't mistake its answers for your organization's official policy or for legal advice.
AI disclosure. Fin states that it is AI at the start of each conversation. WorkBright legal reviewed the disclosure language. This aligns with the transparency requirement in Article 50 of the EU AI Act.
Confirms it is AI when asked. If a user asks whether they are talking to a person, the assistant says it is AI. This has been tested with many different phrasings. As Intercom notes, no system can guarantee the right response to every prompt, so human oversight and the full conversation log remain available.
Procedural guidance only. Answers come from WorkBright content and USCIS guidance. Policy-specific and legally sensitive questions go to your organization.
Terms and privacy. WorkBright updated its end user license agreement to cover this feature. Existing customers accept the updated terms the next time they log in. The chat links to WorkBright's Privacy Policy.
Security controls
Intercom protects Fin against common LLM threats, including the relevant items in the OWASP Top 10 for LLM applications. The details below come from Intercom's Fin AI security materials and public documentation.
LLM threat | How it is addressed |
Prompt injection | User input is handled through separate, limited LLM calls with defensive prompts, so it is not passed directly to the model. Intercom monitors for new injection techniques and moves to stronger model versions. Few attempts have succeeded, and those were fixed quickly. |
Hallucinations | Answers are limited to approved content through RAG. Newer models are less prone to hallucination, and Intercom assesses how any hallucination could affect later behavior. An accurate knowledge base lowers the risk further. |
Data leakage | Intercom assumes anything sent in a prompt could appear in a response, so it only sends data the parties are already authorized to see. Access controls, logging, and Zero Data Retention apply. |
Data poisoning | WorkBright does not train the LLMs. Knowledge base, user, and conversation data are protected with data validation, limited prompt context, and anomaly detection. |
Evasion attacks | Production models are scanned for LLM vulnerabilities. Prompts are designed to resist manipulation, and safeguards are updated as new techniques appear. |
Insecure integrations | APIs use secure defaults, including a secure proxy to LLM providers and customer controls such as Identity Verification. Intercom runs regular penetration tests and security reviews. |
Testing
Intercom tests every model and prompt change before release. Testing includes:
Backtesting in a secure development environment to confirm answers stay based on the provided content
Automated and manual quality checks
A/B testing on a small share of traffic before wider release
After release, Intercom's engineering and machine learning teams monitor production. Intercom also runs red team testing, including jailbreak attempts, reviews of how a hallucination could affect later behavior, and vulnerability scans. A model that fails Intercom's security criteria is not deployed.
Certifications
The AI layer is covered by Intercom's compliance program. According to Intercom, Fin holds these certifications and compliance standards:
SOC 2 Type II
ISO 27001, ISO 27701, ISO 27018, and ISO 42001
HIPAA
GDPR and CCPA
You can review this documentation as part of your own due diligence at trust.intercom.com.
