News You Can Use

Reasoning Models Dominated - the release and wide adoption of OpenAI's o1 and DeepSeek-R1 confirmed the shift from massive pre-training (GPT-5 hype) to inference-time compute ("reasoning"). Niche over Generic - as predicted, "God models" plateaued. The market pivoted to smaller, distilled models (e.g. o1-mini) and RAG workflows that leverage SME knowledge rather than relying solely on model parameter size.

Extraction Worked, End-to-End Lagged - GenAI became standard for structured data extraction in diligence (e.g. Harvey, Legora, Hebbia). However, the "end-to-end" automated deal execution faced friction due to integration costs and reliability issues. Client Push - clients increasingly demanded data-backed insights from these tools, moving beyond simple "review" to "structured analysis" for decision-making.

Budget Shift - the GC Pulse 2025 report showed in-house tech adoption jumped (44% frequent usage), with 80% of departments allocating up to 20% of budgets to tech. Microsoft Ecosystem - in-house teams bypassed expensive legal-specific SaaS for "good enough" internal builds using Microsoft Copilot Studio and Power Platform, forcing law firms to compete on implementation advice rather than just tool access.

Fragmented Standards - efforts like the Allens AI Benchmark and Vals AI picked up speed, but a unified industry standard failed to materialise. Vendors remained opaque about training data. Focus on "Reasoning" - metrics shifted from "speed/context window" to "reasoning accuracy" and "citation hallucinations" (where models like o1 scored ~1.6/3, still requiring human verification).

Consolidation Reality - the market saw significant M&A activity, such as Cleary Gottlieb acquiring Springbok AI (acqui-hire) and the consolidation of struggling startups (e.g. Robin AI's challenges/restructuring). Wrapper Fatigue - "thin wrapper" startups died out. Firms solidified around a hybrid model: buying platforms (Google/Microsoft) for infrastructure and building bespoke workflows for practice-specific logic.

Fragile Autonomy - while 2025 was touted as the "Year of the Agent" (99% of devs exploring it), production reliability remained a blocker. Agents worked well for retrieval (RAG) but struggled with complex, autonomous decision loops without human intervention. Orchestration - "agents" often ended up being rebranded rigid workflows (orchestration) rather than true autonomous entities.

Risk Maturity - the conversation successfully moved away from "will OpenAI train on my data?" (solved by Enterprise/Zero-retention terms) to "can we trust the reasoning?" Verification - with the rise of reasoning models, the risk focus became verification of outputs and citing sources, as even advanced models continued to hallucinate precedents.