The ProSearch Pulse Volume 20

A growing number of firms are adopting labels like “AI-native,” “AI-powered,” and “AI-enabled,” but a new report from Law.com notes there is still no shared industry definition for what those terms mean in day-to-day legal operations. Leaders at Soxton and General Legal point to a practical standard: where AI sits in the workflow, how quality is managed, and what human review happens before anything becomes client-facing.

Key insights

• Soxton founder Logan Brown describes an AI-first target of using AI to get work to “98% of the way,” with lawyers reviewing anything client-facing so AI output never goes directly to clients.
• General Legal co-founder and CTO Javed Qadrud-Din says AI takes a first cut on incoming work, then lawyers remove, add, and modify content before it goes out.
• “AI-native” can mean very different operating models, including hybrid firms tied to legal tech companies versus standalone law firms, different degrees of AI autonomy, and reliance on off-the-shelf tools versus proprietary systems.
• Early activity is concentrated in commercial and transactional work, where tasks repeat. Litigation is harder to automate, though AI use for complaints and case identification is growing.

For legal teams sorting signal from branding, the most useful move is to press for specifics: what is automated, what is reviewed, and what governance exists around confidentiality, auditability, and supervision. As more firms experiment with fixed fees and leaner staffing, the advantage will go to teams that turn legal work into decision-grade outputs while keeping controls tight enough to earn trust.

A major new front in AI liability is forming in San Francisco. In February 2026, the California Superior Court of San Francisco County coordinated twelve cases against OpenAI into In re: ChatGPT Product Liability Cases, JCCP No. 5431, putting scattered complaints on a single, high-stakes track.

Key insights

• Plaintiffs frame the alleged harms as product-design failures, including a claimed “sycophantic design” that can affirm vulnerable users and worsen mental health outcomes.
• The coordinated complaints go beyond accuracy disputes to alleged psychological harm, including reinforcing delusions, endorsing suicidal ideation, and providing self-harm information.
• Four legal questions look set to drive appellate pressure points:
• Whether a chatbot is treated as a product or a service
• Whether Section 230 applies
• Whether the First Amendment protects AI outputs
• How design-defect tests apply to LLMs
• Discovery is the central battleground. Plaintiffs want internal safety evaluations, red-teaming records, model cards, and documentation of known risks.

For organizations building or deploying generative AI, JCCP No. 5431 reads like an early blueprint for how courts may examine safety testing, product warnings, and design choices. The outcome could shape how AI systems are documented, risk-assessed, and defended long before any claim reaches a jury.

As AI speeds up how data gets processed and reused, a long-standing gap is getting louder. Legal guidance is written for humans. IT systems need deterministic rules. In a Towards Data Science piece, Corné Potgieter argues that GDPR-era ambiguity now creates operational risk, especially as autonomous agents trigger actions without human review.

What’s shifting in practice

GDPR, introduced in 2016, marked an inflection point. Legal intent often shows up as narrative, context-heavy language. Data and engineering teams still have to translate that intent into controls that can be implemented, tested, and audited. As more workflows become automated and agent-driven, that translation layer turns into a pressure point.

Governance can’t live only in policy documents. It has to compile into rules that systems can enforce consistently.

Key insights

• GDPR’s 2016 arrival widened a recurring mismatch. Policies written for human interpretation increasingly fail to map cleanly into implementable requirements for data and engineering teams.
• The proposed governance workflow uses three phases:
• PREP, standards and a catalog
• MAP, structured change requests spanning Business, IT, and Legal
• RUN, continuous enforcement and monitoring
• Data contracts attached to output ports can encode purpose limits, retention rules, and transformation constraints. One example permits product enhancement, bans marketing and profiling, requires pseudonymisation for AI use, and caps retention at three years.
• In a cited automated governance study, a system reviewed 110 access requests in real time, matched every expert-flagged issue, generated 3.6x more warnings, and experts later confirmed 80% of those warnings as valid.

Why it matters for partners operating at AI speed

The upside is speed with traceability. Routine checks become repeatable, and scarce legal judgment stays focused on the truly ambiguous edges where context still matters.

The downside is governance theater. Polished AI-generated rationales can inflate confidence faster than comprehension. Accountability weakens when systems are already moving without a human in the loop.

Potgieter predicts near-term incidents involving agentic AI, followed by tougher regulatory scrutiny around 2027/2028. That timeline raises the bar for demonstrable, auditable controls that hold up under review.

This maps closely to what we see when partners try to transform legal data into strategic insight. The AI + HI model fits here. Machines enforce deterministic guardrails at scale. People provide judgment where the rules run out. Early planning, shared standards, and clear data contracts separate automation you can trust from automation that only looks compliant on paper.