Benefits AI your employees can actually trust.
A generic AI chatbot will answer almost anything, including the things it does not know. Tobie is different by design. It answers only from the benefits documents you approve, cites every answer, and tells people when it does not know instead of guessing. Bounded, cited, and reviewable.
For benefits, a confident wrong answer is a real liability.
An employee benefits AI assistant is only useful if people can rely on it. A general chatbot is built to always have an answer, which is exactly the wrong instinct when the subject is someone's coverage.
It fills gaps with guesses
A generic AI chatbot answers from everything it has ever read. When your specific plan details are not in front of it, it produces a confident, plausible answer anyway. That is a hallucination, and on benefits it can be costly.
It answers off-policy
Ask a general assistant about benefits and it will happily speak beyond your plan: national averages, other employers' rules, advice you never approved. Employees cannot tell which part actually applies to them.
No one can see what it said
When the assistant gets it wrong, HR usually finds out from the fallout, not from a record. Without a log and a source, there is no way to see what employees were told or to correct it.
Bounded to your documents, accountable by default.
Source-grounded AI does not answer from general knowledge. It answers from the content you approve, shows its work, and hands off when a question falls outside what it has been given.
It reads only your guide
Tobie answers from the benefits documents you load and approve, and nothing else. Your approved content is the only source of truth it draws on.
It answers with a citation
Employees get a clear, plain-English answer with a citation back to where it came from, so they can see the basis for what they were told.
If it is not in your docs, it says so
When the answer is not in your approved documents, Tobie does not invent one. It tells the employee plainly and points them to a human.
Every exchange is logged
Each question and answer is logged for HR to review, with the source citation attached, so your team can see and correct the record.
An HR AI assistant that stays in bounds.
The same deliberate choices that keep answers grounded also keep your data protected and your team in control. Here is what is built in.
Built on Anthropic's Claude
Tobie runs on Anthropic's Claude, a model built with safety as a first principle, and pairs it with strict source grounding.
Never trains on your data
Your content is never used to train models, and conversation data carries a short 7-day retention window.
Tenant isolation
Each customer's data is separated at the database level using Postgres row-level security. Your data is logically yours alone.
Role-based admin
Role-based admin controls who can see and change what, so the source of truth stays in the right hands.
Full audit trail
Every answer is logged and reviewable, with its source citation attached, so HR can see exactly what employees were told.
You approve the source
Tobie answers only from documents you load and approve, so the content behind every answer is content you signed off on.
A straight answer, when they need it.
All of the guardrails sit behind the scenes. What employees feel is simply a benefits AI chatbot that is easy to use and gives them answers they can trust.
On their phone
Mobile-first, so people can ask about their coverage wherever they are, without digging through a PDF.
Any hour
Available around the clock, so a question at night gets a trustworthy answer instead of waiting on an HR reply.
In plain English
No benefits jargon. Tobie translates premiums, deductibles, and the rest into the words people actually use.
Every answer, accounted for.
This is what your employees actually see: plain-English answers, the questions they ask most, and a citation on every response.
Answers they trust, with the receipts.
Tobie answers from your documents in plain language, cites where each answer comes from, and hands off to a person when it should.
- Plain-English answersCoverage, costs, and enrollment explained the way a person would explain them, without the jargon.
- Suggested questionsThe real things people ask, surfaced up front so they know where to start.
- Escalate to a humanWhen a question needs a person, Tobie routes it to your team instead of guessing.
- Always cites the sourceEvery answer points back to the plan document it came from. No hallucinations.
How the assistant actually works.
Is Tobie just ChatGPT for benefits?
No. A general chatbot answers from everything it has ever read, so it will confidently fill gaps with plausible guesses. Tobie is source-grounded: it answers only from the benefits documents you approve, cites where each answer came from, and says so when something is not covered. It is a bounded, accountable assistant for your plan, not an open-ended chatbot.
How does Tobie avoid hallucinations or making things up?
Tobie is constrained to your approved source content. Instead of generating an answer from general knowledge, it retrieves the relevant passages from your benefits documents and answers from those, with a citation. When the documents do not contain the answer, it does not invent one. It tells the employee it does not have that detail and points them to a human.
Where do the answers come from?
Only from the documents you approve: your benefits guide and the materials you load and sign off on. Nothing else is treated as a source of truth. Because answers are drawn from that approved set, each one can carry a citation back to where it came from, so employees and HR can see the basis for every response.
Is our data used to train AI models?
No. Tobie is built on Anthropic’s Claude, and your content is never used to train models. Conversation data carries a short 7-day retention window, and each customer’s data is isolated at the database level using Postgres row-level security, so your information stays yours.
What happens when Tobie does not know the answer?
It says so plainly and points the employee to a human, rather than guessing. A clear handoff is a feature, not a failure: for benefits, a confident wrong answer is a real liability, so Tobie is built to defer when a question falls outside your approved documents.
Can HR see and control what the AI tells employees?
Yes. Every answer Tobie gives is logged and reviewable, with the source citation attached, so your team can see exactly what employees were told. Role-based admin controls who can see and change what, and because answers come only from documents you approve, HR controls the source of truth itself.
See it answer from your guide.
Start with a Guide Review. We turn your benefits guide into a source-grounded assistant, so your employees get answers they can trust and your team keeps control of the message.