The ProSearch Pulse Volume 28

A legal matter is a business problem moving across systems: scope changes, collections expand, review decisions accumulate, invoices land, deadlines shift, and strategic calls get made in emails, meetings, and project plans. The challenge is managing that business record as one connected picture, even when the information lives in many places.

Moving fact development earlier was the argument in Joe's first blog post on Agentic AI. The next opportunity is broader: using agents to reason across the business of the matter itself. Not just what the documents say, but what the work is costing, what scope is changing, what decisions have already been made, and where burden or risk is beginning to concentrate.

That is the problem agents can help solve. The real corpus is the matter: documents, review calls, production history, collection inventories, budgets, invoices, staffing plans, deadlines, client context, and prior matter knowledge. The value is not pulling everything into one platform. The value is giving counsel and legal operations a controlled way to ask matter-level questions across all of it.

A repeatable playbook gives that work a process. Now consider what the process needs to execute. A rule that says to revisit collection scope when a new business relationship appears needs both the collection history and the client's organizational context. A rule that says to assess what another round of investigation is buying needs findings and spend. The instructions can be sound while the information needed to follow them sits somewhere the agent cannot reach.

iManage has announced a new plug-in integration with ChatGPT Enterprise that lets knowledge workers access governed iManage content directly inside ChatGPT. The promise is control. Existing permissions, ethical walls, and audit trails stay intact while users search, review, organize, and save updated work back into iManage.

Key insights

• iManage is pitching this as a practical on-ramp for generative AI in confidentiality-sensitive workflows, where auditability and ethical wall enforcement are table stakes.
• Deployment is framed as low friction. Customers can enable the integration inside ChatGPT Enterprise without manually setting up and configuring a separate connection.
• The integration is built on iManage Context Fabric, which links signals across content and activity data so ChatGPT Enterprise can surface related prior work that is more relevant and more verifiable.
• Legal is the flagship use case. iManage points to contract review scenarios that pull the latest agreement, correspondence, and amendments, then save revisions back into iManage as a new version with full history preserved.

For many legal teams, the gating factor in operationalizing AI is governed access to institutional knowledge, not raw model capability. When permissions, matter boundaries, and version history are non-negotiable, an integration that keeps controls intact and reduces setup effort can move AI from pilot curiosity into everyday workflow.

That is the kind of connection Joe Pirrotta is asking for in this issue. iManage says the plug-in keeps permissions, ethical walls, and audit trails in place, and that users can search documents from a matter and save a new version back to iManage. The matter file stays in its system of record, and the assistant works through approved access.

OpenAI unveiled Astra for Law, a legal-focused configuration of its latest model, GPT-6 Astra, paired with a dedicated Legal Search Index and workflow tooling. The company is positioning it as a foundation legal teams and legal tech vendors can co-create on across research, drafting, and deal support products, beginning with a Trusted Access rollout.

Key insights

• The primary differentiator is Astra for Law’s Legal Search Index, spanning U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs sourced from CourtListener via the Free Law Project. It’s a jurisdiction-specific index with fewer dead ends and more reliable citations.
• In tests on 200 U.S. legal research questions from Vals AI’s Legal Research Bench, Astra for Law reached 54% overall correctness at the highest reasoning effort, compared with 38.7% for GPT-6 Astra using web search alone. The retrieval layer drives the gain.
• Coverage and retrieval improved alongside correctness: Astra for Law found 24% more reference cases on case-law questions and retrieved up to 54% more relevant passages from the correct opinions in an audited set. Legal teams see fewer missing authorities and less time spent backtracking.
• OpenAI is building an ecosystem around extensibility: 26 vendor plugins, including Thomson Reuters, Harvey, Legora, and iManage, plus nine community plugins and 47 custom skills, alongside general availability of ChatGPT for Word. The practical story is integration gravity. The model gets more valuable as it plugs into drafting and knowledge workflows.

For large firms, the competitive question is whether domain-specific retrieval, integrations, and governance controls reduce risk while improving research completeness. OpenAI’s Trusted Access program leans on enterprise assurances, including zero data retention for the API and a promise that ChatGPT Enterprise usage is excluded from human review, as firms assess readiness for high-stakes work where citation quality and defensible process matter.

The plugin list is where this story meets Joe Pirrotta's argument in this issue. OpenAI launched 26 vendor plugins, including iManage, so the assistant can reach systems where matter work already lives.

A federal judge in Delaware declined to approve a proposal to use a unilaterally designed AI model to identify “Recent Graduate” job postings for collective notice in an age-discrimination matter, Goldstein v. RTX Corp.. The court directed the parties to use a cooperative approach and ordered a meet-and-confer on search terms within 14 days.

Key insights

• The court required a collaborative, negotiated process for search terms instead of unilateral training of an AI model to classify postings.
• The court adopted RTX’s approach of filtering roles using specific Workday database fields, reflecting comfort with structured data constraints as an initial boundary.
• The court rejected the AI model proposal even though defendants argued Workday fields could not cleanly isolate “Recent Graduate Positions,” and flagged the risk of under-inclusion.
• The court adopted plaintiffs’ search-term approach in part, and rejected their request to compel production of a 30% sample of postings as unsupported and too close to discovery for this stage.

For legal teams navigating eDiscovery in HR and ATS systems, the ruling provides a clear signal. Courts may accept field-based filtering as a practical starting point, while requiring transparency, joint testing, and an auditable protocol before AI-driven classification determines who receives notice.

At ProSearch, this matches what we see when partners turn legal data into usable insight. Early planning, clear data boundaries, and an AI + HI model, where human judgment validates machine classification, produce search workflows that can be challenged, refined, and defended.