You're benchmarking against the 2%

Everyone feels behind. That's the one thing the headlines reliably deliver: an autonomous agent fleet over here, a whole function automated over there, a usage chart that just goes vertical. Read enough of it and you assume the rest of your industry is three levels ahead of you.

Notion put numbers to that feeling this week. 88% of organizations are still using AI as a personal productivity tool, drafting and summarizing and the occasional brainstorm. Only 2% have agents running critical processes end to end. The companies in the keynotes and the benchmark threads are real, but they are that 2%. Benchmark yourself against the discourse and you're benchmarking against the 2%.

On the other hand, many companies in that 2% are sharing their experiences publicly and it's worth paying attention to. What separates the front-runners isn't more AI, and it isn't a better model. It's rewiring how your org operates coupled with a whole lot of unglamorous work to make that new operating model effective. They're sharing the blueprint that everyone is going to need to follow.

Here's what we're reading this week from the 2%:

The Great Renovation

Notion asked the leaders who purchase AI tooling and the people who actually use it where their org stands, and the two groups describe different companies. Leaders are twice as confident in their org's AI as the employees using it every day. That gap is the real finding: a strategy built on the executive view is planning for a company that doesn't exist below the leadership line.

Less AI for the visible work of an individual, more AI for the connective tissue of an organization.

The part to bring to your team is what separates the mature orgs, because it's none of the things adoption programs obsess over. Training, policies, standardized tools: everyone invests in those, so they sort no one. The moves that actually track with maturity are wiring AI into the systems work already runs on, building governance, and measuring impact with metrics that mean something. The thing separating the 2% is the plumbing, not the water. So the good news is you're not as behind as you might think. The bad news is that getting ahead takes hard work, not just tools and training.

OpenAI on how agents are transforming work

OpenAI's research team published its Codex adoption data, and while learning how much OpenAI is using Codex isn't surprising, they provide rich detail on how work is actually changing. They're seeing growth beyond engineering: legal, finance, and recruiting flipped to agents as their primary tool this spring, and the tasks got longer, with nearly a quarter of Codex requests now estimated at over an hour of human work. The real leading indicator is non-engineers handing off multi-hour work, not engineers shipping more diffs. As the Notion report indicates, you don't have to be here yet but now you have a glimpse of where you're headed.

Block rolls out BuilderBot

Block also shared some details around how they're actually putting AI to use in production. BuilderBot is an orchestration layer that lives in Slack and runs agents across the whole codebase, and Block says it now merges around 15% of their production code changes.

BuilderBot is walled off from customer data and PII, which is governance, and it lets an engineer ship into a service they've never opened, which is integration earning its keep. 15% of production code through agents is a governance-and-integration story, not a model story. They're buying the model off the shelf just like everyone else; the year of work was everything wrapped around it.

We are now factory engineers, not product engineers

Warp's Zach Lloyd published the memo he sent his team about what he sees as the future of work. It echoes what a lot of the most AI-pilled folks are saying: an engineer's job isn't to write code, or even to build the product. It's to build the factory that builds the product. Features shipped becomes a failure metric; what counts is the share of changes that ship with no human in the loop, over what they cost.

The day-to-day job of the engineer is not to build that product directly, it's to build the thing that builds that product.

To his credit he shows the seams: Warp's own factory is "half-working" by his own account, with 1,300 issues queued that no agent has picked up. The line worth sitting with is that software production is moving from R&D to COGS, which quietly ends the era of unlimited token budgets per engineer. At the frontier the unit of work is the system, and the metric is cost per shipped change, not changes shipped. That's the L2-to-L3 jump in a sentence, the one Notion says almost no one has made.

Teaching agents product design at Vercel

If the others are what the frontier built, Vercel is the how. Coding agents generate working UI fast, but they don't know why one pattern became your standard, so they ship things that are technically fine and completely off-brand. Vercel's fix is to treat product decisions like code: a skill that gives agents the reasoning behind judgment calls, linters for the mechanical rules, and a review loop that pulls evidence from Slack, Figma, and GitHub to keep the guidance live.

Encoding institutional judgment, not just code, is what makes a factory ship things you'd sign off on. It's the least glamorous item on the whole list and probably the highest leverage.

One thing to try this week

Anthropic replaced Claude in Slack this week with Claude Tag: one persistent Claude per channel, visible to the whole team, with scoped access to your tools and data, working async over hours instead of one reply at a time. For most orgs it's now the easiest way to drop an agent where high-context work already happens.

Don't sprinkle it everywhere. Pick one channel where the work is genuinely high-context and async, and wire in scoped read access:

• Incidents / on-call. Give it your logs, runbooks, and past incident threads. It triages and surfaces the closest prior incident while humans fight the fire.
• Customer escalations. With account history and docs connected, it does the first-pass research on each ticket before a human picks it up.
• PRD / design review. Drop it where specs land, with your past decisions in context, and have it flag where a proposal quietly contradicts something you already decided.

Choose the channel by the shape of the work, not just the team that's most eager.

The distance between you and the 2% turns out to be mostly unglamorous work. That's good news: unglamorous work is the kind you can start on Monday.

— Claire + Zach