News You Can Use

Edition 24 · 15th - 31st Aug 2025

News You Can Use

Deep Dives

Three stories worth sitting with

PromptQL - Being 'Confidently Wrong' is holding AI back

PromptQL - Being 'Confidently Wrong' is holding AI back

What
The article warns that AI's biggest problem is not messy data or bad models -- it is being wrong with confidence. "Confidently wrong" systems mean users spend more time checking the AI than benefitting from it. Even a small error rate (like 10%) in multi-step workflows often results in two out of three attempts failing. The article suggests that AI should instead admit uncertainty when needed and get smarter over time, through a learning loop that combines machine confidence with human feedback.
So what
This insight hits home: to gain trust and increase adoption, we should prioritise systems that can indicate when they are unsure, learn from corrections to improve over time, and help users level up rather than relying on blind trust. For example, positive feedback from lawyers testing playbooks on Clausebuddy highlighted that it flags clauses that it lacked information on and should be looked into manually.

Artificial Lawyer - Stop Calling Workflows 'Agents'

Artificial Lawyer - Stop Calling Workflows 'Agents'

What
The article pushes back on the trend of labelling any automated workflow as an "AI agent". It argues that true agentic AI should operate end-to-end, meaning no human intervention between steps. A genuine agentic system could receive a request, follow workflow seamlessly, handle exceptions, and reroute actions when needed -- all autonomously.
So what
When building and promoting AI workflows, accuracy in labels matters. Calling something "agentic" sets expectations -- autonomy, continuity, and decision-making. If experimenting with multi-step task flows, design and build for true autonomy, otherwise address them as "assisted workflows".

Worth Reading

Everything else worth a click

NVIDIA - SLM Agents

NVIDIA research arguing small language models are sufficiently powerful and far more economical than large ones for specialised agent tasks. Make the heterogeneous combination of SLMs and frontier LLMs the default for production agents.

GenAI Report 2025

Annual GenAI market report covering adoption across professional services, vendors, use cases and spend. Useful long-form market scan.

Learning News - AI as Meeting Facilitator

Pieces like Funware are moving beyond transcription to actively facilitate meetings - managing energy, shaping dynamics, guiding discussion. Interesting angle for L&D and internal training.

PromptQL - Being Confidently Wrong

Makes the case that AI's core trust problem is confident error. Even a 10% error rate per step breaks multi-step workflows, so systems need to flag uncertainty and learn from corrections.

Gunderson Podcast (S2 E7 - Lessons from the Edge of Tomorrow)

Gunderson Dettmer's podcast on what they're learning from running at the edge of AI practice. Worth a listen for practical takes on how an AI-forward firm operates day to day.

Legal Tech Trends Newsletter #45

Peter Duffy covers Factor's Sensemaker Academy, updates on Harvey and Legora, fresh funding and M&A, and PE activity across legal services. Five linked podcasts to work through.

Above the Law - AI and Legal Impact (Part 1)

Stephen Embry on AI's "equalising effect" - underperforming lawyers catch up to top performers, quality gaps narrow, and differentiation shifts to human judgment and client context.

Above the Law - AI and Legal Impact (Part 2)

Embry's follow-up on how AI can democratise legal knowledge while still needing human experts to validate. Frames AI as a mentor to juniors that challenges and critiques rather than hands over answers.

Limitations of Embedding Based Retrieval (PDF)

[Internal AG resource] Paper on where pure embedding retrieval fails for legal-style queries (compositional, multi-hop, conditional). Good grounding if you're evaluating or building RAG systems.

Juro and Wordsmith Strategic Partnership

Juro (CLM) and Wordsmith (legal AI agents) integrate via MCP. Juro's 550+ clients get access to Wordsmith's agent workspace, and Wordsmith users get Juro's contract data extraction.