How Shlix IBS uses AI
Shlix IBS uses AI language models to read and sort the requests that arrive in an insurance agency's service inbox. It drafts work from the agency's own records. It does not give insurance advice, bind or change coverage, or make decisions a licensed person should make. This page explains where AI is used, what it sees and how people stay in control.
1. Which models we use
- Anthropic (Claude), through its commercial API, to read and classify agency requests and read the documents attached to them.
- Groq, through its commercial API, for demonstrations on sample (invented) data only. We do not send agency data to Groq.
The full list of companies that handle data for us is on the sub-processor page. We tell agencies before adding a new AI provider that would see agency data.
2. What the AI sees
Only what a task needs: an email's subject, sender and body up to a fixed length, and the attachments needed to complete the request. Values on a certificate or ID card come from the agency's records and the requester's sample, not from the model's imagination, and each value shows where it came from.
3. Training
We do not use agency data to train AI models, and we do not let one agency's data serve another. Anthropic's commercial terms provide that it may not train its models on customer content sent through its commercial API. Groq is used only with sample data, so no agency data reaches it.
4. Confidence and the Exceptions queue
For every email, the model gives a probability that it is agency work and a confidence in what kind of request it is. The agency sets the threshold (0.8 by default):
- Clearly agency work becomes a task, routed to the right person with a due time.
- Clearly personal email is dropped in memory. Nothing is stored, not even the subject or sender.
- Anything in between goes to the agency owner to decide. Only the sender and subject are shown, and they are deleted once the owner decides.
- A request whose type the model is unsure of goes to the Exceptions queue for a person, instead of being guessed.
- If the customer, the policy or the requester cannot be matched, the automation stops and asks a person.
5. People stay in the loop
Each request type has its own approval setting. Certificates, and anything the agency marks as needing an account manager, wait for one, and the person who prepared the work can never approve it. Owners can switch customer email off for the whole agency. The agency's licensed staff remain responsible for what is sent; see section 5 of our terms.
6. When the AI is wrong
- An approver can send any draft back with a note, edit it, or reject it. Nothing that needs approval goes out without one.
- Every step is kept in the task's audit trail, with what the AI found and who approved, so a mistake can be traced.
- Agencies can report a wrong result from the task itself or through support. We review it, adjust the request type, routing or wording rules with the agency, and tell them what changed.
7. What we do not do with AI
- We do not quote an automation rate before running on your requests. Your first month measures it.
- We do not use AI to write testimonials, reviews or customer stories (see our testimonial policy).
- We do not use AI to make eligibility, underwriting, pricing or claims decisions.