Harvey engineers set out three principles for moving from prompt-based features to scalable agent frameworks, covering shared infra, evaluation and safety. Useful if you're thinking about how to industrialise "agents" inside a firm rather than just ship one-off copilots.
Former A&O boss David Morley explains why PE money is flooding into law firms, what investors actually want, and how that collides with partnership culture. Good context for any "should we take outside capital?" discussion.
Do LLMs Truly Understand When a Precedent is Overruled? (PDF)
[Internal AG resource] New benchmark on 236 SCOTUS case pairs finds long-context LLMs still struggle to reliably spot overruling relationships, especially in older cases. Shows era bias, shallow heuristics and "temporal reasoning" failures that matter for serious legal research use cases.
Debevoise dissect why "agents" fall over in messy real workflows and where humans still need to be firmly in the loop. Helpful framing if you're being sold fully autonomous legal workflows.
Long riff on OpenAI's current strategy, model competition and whether MCP is too complex for most users to care about. Worth a listen if you're trying to read the tea leaves on tooling ecosystems and "daily driver" models.
Harvey's new internal "arena" pits models and systems head-to-head on real expert preferences, not just leaderboard benchmarks. Interesting blueprint for how firms might evaluate models and vendors at scale rather than trust marketing slides.
Survey shows in-house teams leaning on gen AI for cost and efficiency, while firms use it more for differentiation and pitch theatre. Useful reality check on who's chasing margin vs marketing.
Follow-up analysis finds the infamous "innovation gap" is narrowing but not gone, with 90%+ of in-house saying innovation matters yet still frustrated with delivery. Good fodder for partner decks about why tinkering isn't enough.
Mabey argues lawyers aren't magically exempt from the automation curve and should stop assuming "this time it's different" for legal. A sharp take on where legal work really is and isn't defensible against AI.
Provocative polemic questioning Harvey's valuation and arguing it's essentially a dressed-up OpenAI front-end. Worth reading as a counterweight to the hype, especially if you're signing big multi-year deals.
Explores how an NHS Copilot pilot is saving 400k hours a month for 100k staff by cutting admin, while flagging the usual blockers of legacy IT and skills. A useful public-sector case study you can actually quote to boards.
Legora's new Portal capability extends its knowledge platform into secure client-facing workspaces, tying together DMS content, portals and intake systems into one view. Useful benchmark for what client portals could look like in an AI era.
White & Case unveils Atlas, a multi-model, ChatGPT-style assistant built on a proprietary AI platform, plus a firm-wide change programme to drive adoption. The kind of coherent AI roadmap clients will soon expect from every global firm.
100-page deep dive on the policy and governance headaches created by AI assistants and agents - from accountability and labour impacts to "how much should we actually delegate?" Ideal for anyone drafting firm or regulator-level AI policies.
OpenAI's latest flagship promises faster, more conversational performance than GPT-5 with better tool use and cheaper tokens, plus simpler ways to build customised assistants. Relevant if you're revisiting your model stack or planning 2026 platform migrations.
Sketches a "grand bargain" where firms offer transparency on AI use, process and pricing in return for volume and partnership from clients. Good stimulus for redesigning panel terms in an AI-first world.
Practical on-ramp for ops, product and legal folks who want to move beyond no-code and into tools like Cursor, v0 and proper engineering workflows. A good confidence booster for "non-technical" innovators who secretly are.
Furlong warns that PE will treat law firms like any other asset - squeezing margins, pushing scale and expecting exits - with big implications for culture and ethics. Essential reading if PE is sniffing around your market.
Concrete patterns that work: be explicit, set roles and constraints, use worked examples, add structure (XML tags), and iterate with evals/guardrails. Useful to standardise team prompts and cut rework.
Shifts agents from brittle tool-call prompts to code that imports MCP tools on demand, filters data before the model, and runs logic off-context. Read if you build enterprise agents - fewer tokens, tighter security/state, clearer execution.
Vigilante Lawyers Expose the Rising Tide of AI Slop in Court Filings (PDF)
[Internal AG resource] Lawyers are publicly cataloguing AI-generated fake citations and other "slop" in filings; sanctions are up, but deterrence is patchy. Use this to justify source-verification checks, AI-use logging, and targeted litigator training.
Google - Context Engineering: Sessions, Memory (PDF)
[Internal AG resource] Google's internal whitepaper on "context engineering" lays out how to design sessions, memory stores and retrieval for long-lived agents. Gold dust for anyone building serious agent frameworks rather than single-shot chats.