[ HEALTHCARE & LEGAL ]
Most website chatbots guess when they do not know. In a clinic or a law firm, that guess becomes something a human has to walk back. Ours answers only from knowledge you approve, shows the source it used, and offers a call when the answer is not there.
Cited
every factual answer
Logged
question, answer, sources
Isolated
deployed per client
You can try it right now. The assistant in the corner of this page is the same product, running on our own content as the reference install. Ask it what we charge, or whether we are HIPAA compliant, and watch where it refuses to guess.
[ WHY THE USUAL ONES DON'T FIT ]
In most industries an inaccurate chatbot reply is an annoyance. In yours it is a liability, a compliance conversation, or a promise your team never made.
Asked whether you are compliant, whether you take a certain insurance, or whether you handle a certain matter type, a generic bot will produce a plausible answer. Plausible is not the same as true, and the visitor cannot tell the difference.
The moment a bot names a number for a specific situation, that number is what the visitor heard. Ours gives documented ranges and hands anything situation-specific to a person.
When something goes wrong with an ordinary widget, there is no record of what was asked, what was answered, or what it drew on. You find out from the person who was told the wrong thing.
[ WHAT WE BUILD ]
You decide what it is allowed to know. Outside that, it does not improvise — it says the question deserves a real answer and offers a call. The refusal is a feature, and it is the behaviour we tune hardest.
Each answer points at the passage it came from, so a visitor, a partner, or a compliance officer can verify it immediately rather than taking the machine's word for it.
Question, answer, sources cited, and timing — recorded. When someone asks what the assistant told a patient or a prospective client last Tuesday, there is a record instead of a shrug.
We build a set of real questions with known-correct answers and score the assistant against it before launch and after every change. Accuracy stops being a feeling and becomes a number you can watch.
It runs as your deployment rather than a shared service, so your knowledge and your logs stay yours, and it installs with a single script tag on whatever your site is built with. On the model side we configure training-disabled and zero data retention, with a business associate agreement in the chain where healthcare requires one. We will not tell you "nothing leaves" — with any hosted model the content reaches the provider for inference, and the defensible posture is the contract, not a slogan.
[ WHERE IT GOES NEXT ]
The version that answers from your published knowledge is the one worth launching first, because it is the one you can verify. Once it is reliably right, the same assistant can reach further.
We maintain open-source connectors for the systems your practice already runs on, so the assistant can be extended to answer from real records under the same logging and permissions.
The same knowledge layer can answer in Slack, Teams or email for your staff, rather than only on the website for visitors. Internal questions are often the ones eating the most time.
For teams whose real problem is the underlying system rather than the front door, the assistant is one piece of a larger build.
[ FAQ ]
An off-the-shelf chatbot will happily invent an answer when it does not know one. In a regulated business that is the whole risk: a bot that guesses at whether you are HIPAA compliant, or quotes a price you never agreed to, creates a problem a human then has to walk back.
This assistant answers only from a knowledge base you approve. When the answer is not in there, it says so and offers a call instead of guessing.
Two mechanisms. Every factual answer carries a citation pointing at the exact source passage it came from, so anyone can check it in one click. And every exchange is written to an audit log: the question, the answer, the sources cited, and how long it took.
We also score the assistant against a set of real questions with known-correct answers before launch and after every change, so "is it still accurate" is a measurement rather than an opinion.
The honest answer starts with what it is not: this is a website assistant working from your approved public knowledge, not a system that should be handling protected health information in a chat box. We design the boundary deliberately, and we wrote up that reasoning in building a HIPAA-conscious chatbot without PHI.
Where sensitive data is genuinely in scope, the defensible configuration is a business associate agreement plus a model configured with training disabled and zero data retention. Not claims that "nothing leaves" — with any hosted model, the content reaches the provider for inference. We will tell you that plainly rather than sell you a comfortable story.
It is deployed per client rather than as one shared multi-tenant service, so your knowledge base and your conversation logs are isolated to your deployment. It installs on your site with a single script tag and does not care what your site is built with.
It depends on the size of your knowledge base, how many questions it needs to answer correctly, and whether it needs to reach into a live system rather than static content. You get a fixed quote after one scope call, and the quote is the price.
That is the natural next step rather than the starting point. We maintain open-source connectors for Clio, MyCase, Filevine and IntakeQ/PracticeQ, so the same assistant can be extended to answer from live records once the static version is proven.
We recommend earning that step: get the assistant reliably right on approved content first, then widen what it can reach.
[ GET ONE ON YOUR SITE ]
Try to make the assistant on this page say something it should not. Then book 30 minutes with a co-founder and we will scope what yours would need to know, what it must never answer, and what a fixed quote looks like.
