At a glance
| Industry | Professional membership association, legal |
| Membership | Approximately 15,000 lawyers across 31 practice sections |
| Objective | Make a decade of scattered CLE, publication and practice content findable, without giving legal advice or giving away paid content |
| Key use cases | Member self-service, comprehensive legal research across the association’s own content, CLE discovery, practice management guidance, lawyer referral, staff enablement |
| Solution | CustomGPT.ai enterprise. Nine agents, a custom crawler for the CLE catalog, own-content-only retrieval, SSO-gated deployment |
| Corpus | 634 million words made searchable |
| Security | SSO at the association’s own identity provider, JWT-signed embed inside the paywall, SOC 2 validated by the customer’s auditor |
| Shared | Anonymously, at the customer’s request |
The challenge
A solo practitioner with a client on hold, hunting the business-law form that lets her close today. A litigator mid-matter checking whether opposing counsel is right about which county governs. A member three weeks into their first year who does not yet know the association already owns the answer.
All three are after material the association spent years building. None of it needed to be created or bought. It needed to become reachable, and it sat across six content systems with nothing in common, each with its own search box, none of which searched the others.
The solution
The evaluation was led by the association’s practice-management director and its director of continuing legal education, with IT and marketing involved early. The criteria were unusually precise, because the buyer was a practicing business attorney who reads contracts for a living.
- Answer only from association content. Retrieval is grounded in the uploaded corpus, and the assistant says it does not know rather than improvising. For an audience of lawyers, the refusal behavior had to be as reliable as the answers.
- Respect the paywall. Section materials and CLE recordings are paid products. The assistant answers generally, then routes the member to the gated page, so search became an upsell instead of a leak.
- Reach what other tools skip. An authenticated crawler was built for content behind the member login. Engineering wrote a dedicated connector for a CLE platform with no usable export. A custom pipeline transcribed and ingested a video library with no transcripts.
- Confidentiality, provably. Nothing uploaded trains a third-party model, no member question reaches a public generative AI service, and no member exists in CustomGPT at all.
The capability that decided it
Most tools handle paid content by excluding it, which leaves the assistant weakest where the association’s content is strongest. We built the other option instead. The viewer opens the page an answer came from, highlights the passage and blurs the rest, so the citation leads to the page that sells access.
Seventy-two business-law forms were run in front of the executive sponsor before launch. Watching the citations point at the paywall rather than around it was the moment the business case made itself. That viewer is now standard for every enterprise customer.
The results
| Measure | First Eight Weeks |
| Member questions answered | 1,940 |
| Peak questions in a single week | 525 |
| Answer rate | 92.4% |
| Arriving through the site widget | 91% |
| Days with member activity | 54 of 54 |
Nine times in ten the association already had the answer on file. Nine in ten of those questions arrive through the widget embedded in the association’s own site.
There has not been a single day since launch without a member question, with no reminder campaign behind it.
Members also invented a use case nobody planned for. As the association’s director of continuing legal education put it: “We had one user last week that actually uploaded a pleading, and used the tool to dispute opposing counsel’s view on which county a minor resided in.”
Nine agents run in production today on one contract: a member assistant, a research agent, a CLE agent, an events agent, a referral agent and internal staff tools. The association has already renewed.
Why it worked
- Grounding that holds under cross-examination. Own-content-only retrieval, with a persona the association wrote itself. For lawyers, being reliably unwilling to guess mattered as much as being right.
- Citations that respect the business model. The assistant answers from gated material at a level the association controls, then routes the member to the page that sells it.
- One contract, as many assistants as the work needs. Nine agents, no new line item and no fresh negotiation for any of them.
- The hard parts were ours. The firewall, the CMS with no export and the untranscribed video were solved by our engineers rather than scoped out of the project.
Frequently asked questions
What is a legal research AI tool for a bar association?
It is an assistant grounded only in the association’s own content, so members can ask a question in plain language and get an answer drawn from the association’s CLE recordings, practice guidance, forms and publications, with the source cited. It does not search the open web and it does not give legal advice.
How accurate is it?
This deployment answered 92.4% of real member questions from association content in its first eight weeks. The number that makes that meaningful is the refusal behavior: when the corpus does not cover a question, the assistant says so rather than improvising.
Does it give legal advice?
No. It reports what the association has published and cites where the answer came from. Anything outside that corpus is refused. The distinction between reporting published material and advising on a matter is enforced by restricting retrieval rather than by instructing a general model to behave.
How does it handle content that members pay for?
It answers from gated material at a level the association controls, then sends the member to the page that sells access. The viewer opens the source page, highlights the passage the answer came from and blurs the rest, so paid content drives a purchase rather than leaking.
Can it reach content behind a member login?
Yes. The most valuable material here sat behind a member login no standard crawler could reach, so an authenticated crawler was built for it. A CLE platform with no usable export got a dedicated connector, and a video library with no transcripts got a custom transcription pipeline.
How is member confidentiality protected?
Nothing uploaded is used to train a third-party model, and no member question or uploaded document reaches a public generative AI service. Access runs through the association’s own identity provider, the assistant is embedded inside the paywall and signed with JWT, and no member exists as a user in CustomGPT. This deployment was SOC 2 validated by the association’s own auditor.
Who can use it, and how do they reach it?
Members only, through single sign-on at the association’s identity provider, with membership status deciding access. Ninety-one percent of questions arrive through the widget embedded in the association’s own site, so members are answered on the page they were already on.
Can one association run more than one assistant?
Yes, on a single contract. This association runs nine: separate agents for members, section research, practice management, events, referral and internal staff, each scoped to its own content. Adding one is a configuration rather than a new procurement.
How long does a deployment like this take to stand up?
The build work here was the content, not the software: an authenticated crawler, a CLE connector and a transcription pipeline. The association also ran adversarial testing in staging against the real corpus, and reviewed every source and sitemap by hand, before anything was promised to members.
What does it cost to add the second assistant?
Nothing beyond the existing contract. The association added nine agents without a new line item or a fresh negotiation for any of them, which is the difference between a platform and a project.
Figures from CustomGPT.ai product usage records. Shared anonymously at the customer’s request.
