Media is full of stories meant to make us mistrust and hate each other. Here’s a story of a 9/11 hero - a 24yo kid who, trapped in the towers, selflessly decided to save almost a dozen people rather than himself. He found a path to safety and instead chose to go back in to rescue more survivors. On that awful day - and all of the ordinary days - there are countless acts of human good. Of people who put others before themselves. Of the kindness of the human spirit. Let’s drop the rage clickbait and share those.

The jobs apocalypse is postponed. An AI jobs boom is here According to @TheEconomist, AI is actually proving to be a net job creator in the US, easily generating over 1M new positions (from data center construction to AI engineering) to offset back-office layoffs. While routine admin and customer service roles face real disruption, the overall labor market is proving resilient. economist.com/finance-and-ec…

100% of our team is active on AI tools. 71% of our org is active in Claude alone daily. Here's what got us there: 1. We set the expectation that AI fluency is a baseline expectation at Scribe, and a big part of how we move faster and win. Usage spiked every time we celebrated AI wins at All Hands. People saw what was possible, then went and did it themselves and got celebrated for it. 2. We didn't insist on building everything in-house. Buy was often the right call. No tooling to maintain, faster time to value. 3. We gave everyone access with limits by team. While a person knows their own workflow and where AI can help best, we didn't want unintentional token waste. We shared efficiency best practices across the org like delegating to sub agents and watching the session length to make this work.

Companies have years of data that exists nowhere else on the internet. Thomson-1 is the first example. It's trained on decades of proprietary content, technology, and domain expertise no other company can match. The result is a model Thomson Reuters fully controls, without the heavy inference costs of typical frontier models. The model has been trained on less than 10% of Thomson Reuters content so far, and is still on par with the latest frontier models across a range of tasks.

A few observations on AI transformation in the enterprise from recent conversations: * “Agentic” does not necessarily mean fully autonomous. The more durable implementations are combining deterministic workflows with AI steps, especially where reliability and debugging matter. * The real work is selecting the right workflows to automate, integrating fragmented data, setting governance, and helping teams change how they work. * Security and privacy are hard requirements. Clear controls over approved apps, user privacy, and employee review materially affect adoption. * Human-in-the-loop is still the default for consequential outputs. One team put it well: AI is great for the middle-to-middle. Brainstorming and final review should stay with a human. * Companies are becoming more selective after broad experimentation. Teams are moving away from unlimited AI-tool purchasing (no more "let a thousand flowers bloom") toward measurable ROI and business case requirements. * Token efficiency is becoming real operational work: model routing, shorter contexts, sub-agents, and centralized observability layers.

It's the same planning fallacy - we overestimate how much we can do in the short-term, and underestimate how much we can get done long-term. It's so common in tech, it has a name - Amara's Law. People see how sophisticated the models are and get happy eyes about what's possible....and then get hit by the reality of actually making that work in the enterprise. It requires so much more than just great models.

Had a great conversation with @DataFramed about why agents haven't dramatically changed knowledge work (yet) and why companies must harness their specialized intelligence first. The models are PhD-level at this point, but none of that training tells them how your company uniquely works. That's what I mean by specialized intelligence - the accumulated decisions, judgments, and exceptions you've built up to serve customers and ship product your way. It's real IP, but it's still stuck in people's heads instead of being something that you can apply to your AI to build a generational advantage for your business. At Scribe, we're working with 90K+ companies (including 50% of the F500) to help them harness their specialized intelligence and turn it into an asset that keeps compounding. Listen → datacamp.com/podcast/what-do…

We talked to senior enterprise leaders to understand where AI transformation actually stands right now. The same problem kept coming up: they don't know if their AI investments are creating value, or if they're even being pointed at the right problems. → More than half say the #1 reason AI initiatives fail is lack of data on how work actually gets done → Most organizations lack systematic methods for identifying automation opportunities and instead rely on leadership intuition and hunches → A third of leaders say the biggest barrier to moving faster is getting team buy-in and managing change. Employees are interested in AI, but many lack a clear starting point, training, and confidence about what adoption means for their roles.

It feels like we've hit diminishing returns on intelligence for many tasks We may no longer see every product auto-switch to the next frontier model upon release This is great for app-layer builders (many opportunities to bring down COGS!)

AT&T is routing 40% of AI usage to open models, and intends to increase it to 60-70%. Frontier models should not be default for most ordinary enterprise work. This is the exact guidance we've given internally - use frontier models for critical work and route everything else to cheaper models. As @JayaGup10 wrote, "frontier systems should be reserved for the smaller set of tasks where additional intelligence still changes the outcome."

Controlling AI costs is top of mind. At Scribe, we've given the following guidance to our team: 1. Delegate to sub-agents wherever possible. Smaller, faster models handle a lot more than people assume once the task is tightly scoped, and you can give each sub-agent its own model and thinking-effort level for what it's actually doing. 2. Watch session length. Long threads are the number one source of wasted spend, since every new turn replays the entire conversation from the start. 3. Use handoffs. When a thread goes deep on one subproblem, ask the agent to write up a handoff doc summarizing where things stand, then start a fresh thread from that instead of dragging the whole history forward.

The next frontier of enterprise AI is about how you bring general intelligence, what's coming out of the labs and the models, to bear in the specific operating context of a firm. A company's value is it's specialized intelligence or "special sauce." The accumulated set of decisions, judgments, and exceptions that add up to how they uniquely service customers, ship a product, manage exceptions, etc. The models are incredibly intelligent, and have been for quite some time. But to see outcomes, companies have to make this context on how things get done legible for both their people and agents.

A lot of leaders I talk to want to "agentify everything" and transform the whole business. While you can absolutely do that, there are a bunch of quick wins on the way there that don't need any of that. There is massive opportunity in looking at what your best people are doing and replicating that across the team. When I was at McKinsey, if you wanted to make an ops team more efficient, you'd find the person who consistently outperformed everyone else and ask them why. They'd usually pull out a binder of shortcuts they'd built for themselves - never shared, never written up anywhere. That binder became the new standard for the whole team. What's changed is that finding that person and mapping what they actually do used to take a team of consultants weeks. Now, an LLM can watch how work happens across your tools and turn it into a workflow map on its own at a scale no human team could match by hand.

Really interesting analysis! Don't throw unlimited tokens at people who aren't yet fully ready to use them.

Great breakdown. The part to own is your specialized intelligence: how your best people actually do the work. The judgment calls, the unspoken expectations, the workarounds that never made it into a doc.

This matches how our own thinking has shifted. Scribe started as something people used to capture and share how work gets done. Now, we're also building for agents. What an agent needs most is your company's specialized intelligence in a form it can actually use. In our own testing, agents with access to Scribe's structured data performed ~70% better.

100% agree. The optimism I see on here is pretty disconnected from what I actually hear talking to enterprise leaders. Supporting agents means fundamentally changing how you work, and that takes enormous upfront effort: making your specialized intelligence legible and usable across the org instead of leaving it in people's heads.

AI agents are exposing how much of every company was held together by people quietly filling in everything nobody bothered to explain. The work was never as organized as it looked. People just knew what everyone meant.