An AI will advisor that answers questions and captures leads for a legal services platform.
Published

Most visitors arrive on a guide about wills, read, and leave. A generic chatbot risked wrong legal information, quoted fees, promises of outcomes, and leads handed to sales with no context. Some of the platform's own guide content also contained legal errors, so the assistant could not simply trust whatever it retrieved.
Answer a reader's first questions accurately from reviewed facts, never quote a price or guarantee an outcome, move to a human specialist at the right moment, capture contact details inside the conversation, and give the business full visibility of every answer, source and decision.
What shipped
A helpful first conversation on the guide page
A single script tag adds a launcher to any page. It records where the visitor came from and what they read, offers a nudge and starter questions that fit the guide, and opens the chat in an isolated frame. Readers can type, or hold a button and speak in English, Hindi or mixed speech.
Reviewed facts first, sources second
The assistant always has a reviewed set of will-law and service facts in front of it, each with a citation and a review field for a lawyer to sign off. When a question needs more, it retrieves passages from statutes, service pages and guides. Reviewed facts outrank guides, so a guide that gets the law wrong cannot override them.
Check the answer before the reader sees it
One step classifies the question, writes the answer and judges whether it is time to offer a specialist. A verifier then checks that every claim is supported, that no price or guarantee appears, and that the answer is not personal legal advice. A failed draft is rewritten once, and if it fails again the reader gets a safe reply that offers a specialist.
Hand off to a person, conversationally
Pricing questions are never answered with a number, and disputes, urgency or an explicit request for a call move to a specialist faster. From the third message the assistant can offer a specialist, and by the fourth it always does; the visitor replies with a name, then a number or email. There is no form and no live takeover: a specialist calls back with a short AI brief and a suggested next step.
Give the business a console, not a black box
An inbox shows who needs a call and who is only browsing. Each conversation shows the transcript, the sources and guardrail results under every answer, the page journey, attribution, the brief and the CRM status. Insights, a knowledge screen, a sales-system view and a hash-chained audit log round it out.
What it looked like in action
Representative mockup using anonymized sample data. The interaction patterns reflect the production flows; names, amounts, IDs, and dates are illustrative.
Inside the product
Screens captured from the working system with illustrative sample data. Names, contacts, companies and figures are placeholders.










Business outcomes
Technical work demonstrated
The systems and controls behind the story above.
Source-priority retrieval
Statutes are chunked by section, guides are de-duplicated and stripped of prices, and passages are embedded into a vector store; results are re-ranked so reviewed facts beat statutes, statutes beat service pages, and service pages beat guides.
Single-call classify, answer and judge
Merging routing and answering into one call cut median latency substantially; the static prompt prefix is identical on every call so most input tokens are served from cache.
Code-computed verifier verdicts
The verifier checks faithfulness, price, guarantee and personal-advice rules with the whole reviewed set as cached context, and pass or fail is computed in code from its findings rather than from its own flag.
Deterministic contact capture
Name and number replies are handled by rules without a model call, including names given mid-sentence and replies in Hindi, with a policy for how many asks and how far apart.
Evaluation harness and live scores
More than a hundred scripted cases, including whole conversations, lead capture and an AI grader for legal accuracy, run through the real pipeline and fail on any price leak, guarantee or prompt leak; each live answer is also scored for grounding and reply rules.
Hash-chained audit and PII masking
Audit rows each include the previous hash, so edits or deletions are detectable, and phone numbers and emails are masked before traces leave the application.
Architecture
Answer pipeline
Retrieval over statutes, service pages and guides is re-ranked by source priority, then one model call classifies, answers and judges the hand-off moment, a verifier checks the draft, and a lead policy decides when to ask for contact details.
Conversation and lead records
Visitors, visits, page journeys, conversations, messages with sources and guardrail results, leads, sales-system records and the audit chain are stored in a managed document database with vector search for retrieval.
Console and tracing
The admin console reads the same records, and every turn is traced with retrieval, answer and verifier steps, token counts, cost and latency, so engineers and sales leaders see the same story.
Information, not advice
The assistant explains general law and services; personal situations, disputes and urgency go to a specialist.
No prices, no promises
Pricing is always handed to a specialist and guarantee language is blocked by the verifier.
Reviewed before trusted
Facts carry a citation and a reviewer field, and unreviewed facts are flagged in the console until a lawyer signs them off.
Everything is inspectable
Sources, guardrail outcomes, attribution, briefs and an append-only audit trail are available for every conversation.
Where this proof applies
This case study supports the solution and service pages below because the same workflow, controls, and handoff model are visible in the implementation.