AI @ Work
June 5, 2026
Issue #9
This week: why your AI strategy is probably failing — and it’s not the tools. Anthropic’s next model is leaking and it sounds bigger than anything they’ve shipped publicly. ChatGPT just gave Codex’s power to everyone, not just developers. Google Pinpoint opened its doors to the public. And Uber found out the hard way what happens when you don’t set AI spending limits. All translated into what it means for your work.
---
★ This Week’s Big Story
Your AI strategy has a trust problem, not a tooling problem. Agency > Agents.
If your organization has invested in AI tools and isn’t seeing the productivity gains you expected, Elena Verna’s piece this week — cross-posted by Lenny Rachitsky to Lenny’s Newsletter — may be the most useful thing you read today.
Elena Verna makes the most important argument about AI adoption we’ve seen this year — and it has nothing to do with which model you’re using. Her thesis: most companies already have the AI tools they need. The blocker is the organizational structure that prevents employees from actually using them.
The pattern will feel familiar. Information gated by titles — when employees get promoted to manager, they gain access to Slack channels and data they couldn’t see the day before. Decisions routed upward for approval, creating a loop where decisions are slow because they’re expensive, and expensive because they’re slow. Employees trained for years to wait for permission who, given autonomy, don’t know what to do with it.
The contrast she points to is Anthropic’s own structure. Anthropic classifies everyone as either “member of technical staff” or “member of non-technical staff” — specifically to prevent titles from becoming permission structures that decide who gets context, who gets heard, and who gets to move.
The line that cuts deepest: “Agents don’t have agency. They wait to be told what to do. High-agency employees do the opposite — they find the signal, make the call, and push the work forward.” Every company is talking about deploying AI agents while keeping human employees in structures designed to minimize their agency.
Her practical recommendation for executives: don’t try to restructure the whole organization. Prototype it. Pick one R&D or innovation team, give them a flat structure and aggressive targets, put your most high-agency people in it, and see what happens.
Executive & Strategy | HR & People | All Readers
Source: Lenny’s Newsletter (cross-post from Elena’s Growth Scoop), June 5, 2026
AI Platforms
Claude / Anthropic
If you thought Claude Opus 4.8 was the end of Anthropic’s current release cycle, the inbox this week suggests otherwise.
Anthropic “Oceanus” Is Leaking — and It Sounds Bigger Than Anything They’ve Shipped to the Public
TLDR AI reported this week that Anthropic appears to be gearing up for the public launch of a new model called Oceanus — described as a version of Mythos that surpasses even Mythos Preview, which is currently restricted to vetted organizations through Project Glasswing.
The timing is notable. Anthropic simultaneously issued a public warning that recursive self-improvement — AI systems helping to build better versions of themselves — may arrive sooner than most organizations expect.
Nothing has shipped yet and no release date has been announced. But when a model is being discussed by name in public channels before an official announcement, the release window is typically close.
Executive & Strategy | IT & Security
Source: TLDR AI, June 5, 2026
ChatGPT / OpenAI
If you’ve heard of Codex but assumed it was only for developers, OpenAI just changed that assumption — deliberately and completely.
Codex Is Now Inside ChatGPT — and OpenAI Is Bringing Its Power to Everyone, Not Just Coders
OpenAI this week folded Codex directly into ChatGPT, collapsing the wall between its developer-focused agent and its consumer product. The goal, according to Alex Embiricos, head of enterprise product, is to bring Codex’s capabilities “to everyone.”
Three practical use cases for non-developers from OpenAI’s press briefing:
Sending messages — connect your email or Slack, describe what you want to say, and Codex drafts it, finds the recipient, and sends it with a single approval
Calendar briefings — connect your calendar and ask Codex questions about upcoming events or surface specific information on demand
Daily automations — set a recurring workflow: “Every morning at 8 AM, give me a summary of today’s meetings and the most important action items from my inbox”
OpenAI also launched six new role-specific Codex plugins targeting creative production, sales, and public equity investing.
Operations | Executive & Strategy | All Readers
Source: The Deep View, June 5, 2026
Gemini / Google
If you work with large document collections — contracts, research files, email archives, case files, meeting transcripts — Google just opened the tool that helped journalists win Pulitzers. It’s free.
Google Pinpoint Opens to Everyone — The Free Document Intelligence Tool Most People Have Never Heard Of
Google Pinpoint has been restricted to journalists and academics since launch. This week, Google opened it to the public — for free.
Pinpoint is a document intelligence system built for massive collections. Upload up to 200,000 files per collection — PDFs, emails, audio recordings, handwritten notes, scans, spreadsheets — and search across all of it in plain language. It transcribes audio in over 100 languages, makes handwritten documents fully searchable, and extracts data to spreadsheets across up to 100 documents at once.
The comparison that matters: NotebookLM synthesizes and generates — it creates podcasts, slides, and reports from your documents. Pinpoint organizes and discovers — it finds the needle in a 200,000-document haystack. Use Pinpoint to surface and sort. Move the most relevant files into NotebookLM to create outputs.
Available now at pinpoint.google.com. Files are not used to train AI models and remain private unless you choose to publish.
Legal & Compliance | Operations | All Readers
Source: Wonder Tools (Jeremy Caplan), June 5, 2026
Cross-Platform
If your organization is expanding AI tool access without a spending framework in place, Uber’s experience this week is the cautionary tale worth reading before your own invoice arrives.
Uber Capped AI Spending at $1,500 Per Employee Per Month After Costs Spiraled — Here’s What to Do Before That Happens to You
Uber implemented a $1,500 monthly cap per AI coding tool per employee after costs rapidly exceeded expectations. Per-seat subscriptions compound fast when a large organization enables them broadly and usage accelerates.
The lesson isn’t that Uber made a mistake by deploying AI broadly. The lesson is sequencing: deploy broadly, measure intensively, cap deliberately.
Three things to do now:
Get a dashboard showing real-time AI tool spend by team — most vendors offer this in enterprise settings
Set a soft alert at 70% of your expected monthly budget so you see the trend before it becomes a problem
Establish a per-employee cap that requires approval to exceed, not one that blocks work automatically
The goal isn’t to restrict AI use. It’s to make the spend visible before the invoice is the first signal something changed.
Finance | IT & Security | Executive & Strategy
Source: TLDR, June 4, 2026
Tools & Workflow — Try These This Week
For anyone who has handed a task to an AI agent and watched it go sideways — the problem is almost always in how the task was described.
Stop Babysitting Your AI Agents: The /goal Framework for Tasks That Actually Finish
AI Maker published a practical framework this week built around a simple observation: most AI agent failures happen not because the agent is incapable, but because the task was too vague to complete without constant course correction.
The /goal framework structures every agent task with four components before it runs:
Goal — the specific outcome you want, stated as a finished result (”a formatted spreadsheet of vendor contacts sorted by region” not “organize the vendors”)
Constraints — what the agent should not do, what requires your approval before proceeding
Context — what the agent needs to know about your situation that it can’t infer from the task alone
Done signal — how the agent should know when it’s finished, so it stops rather than continuing to improve indefinitely
The done signal is the part most people skip — and the part that most often causes agents to over-generate or loop.
Operations | Executive & Strategy | All Readers
Source: AI Maker (aimaker.substack.com), June 4, 2026
For anyone who uses Claude and hasn’t touched the Effort Control setting — almost nobody has, and the difference is immediately noticeable.
Claude Has a Thinking Dial. Almost Nobody Has Touched It. Here’s What It Does.
AI Business Insights flagged this week that Claude’s Effort Control feature — which lets you tell Claude how deeply to reason before responding — is nearly invisible to most users, and the difference between settings on a complex task is dramatic.
The practical guide: for simple, fast tasks — drafting an email, reformatting a document, a quick summary — keep Effort Control low. Speed is the point. For tasks requiring careful judgment — analyzing a contract, evaluating a strategy, reviewing financial projections — turn it up. The output quality at high Effort Control on complex tasks is noticeably different from the default.
In Claude, Effort Control appears in the settings menu for the current conversation. You can also signal it directly in your prompt: “Think carefully before responding” or “Give me your quick take” both work as informal triggers.
All Readers | Operations
Source: AI Business Insights, June 4, 2026
Worth Watching
Anthropic confidentially filed for an IPO this week — and could reach public markets before OpenAI. A public Anthropic would be structurally different: quarterly disclosures, public scrutiny of spending, and shareholder pressure alongside its safety mission. | Finance | Executive & Strategy |
SpaceX priced its IPO at $135 per share, valuing the company at $1.77 trillion. SpaceX is also the company powering Claude’s compute at $1.25 billion per month. An IPO makes Elon Musk’s relationship with Anthropic more public and more scrutinized — two courtroom opponents linked by $45 billion worth of compute contracts. | Finance | Executive & Strategy |
Google Gemini was successfully hijacked via WhatsApp in a documented prompt injection attack. A malicious message manipulated Gemini into taking unintended actions without the user’s knowledge. This is not a Gemini-specific vulnerability — it’s an architectural issue with any AI assistant connected to external inputs. If you have AI agents reading your messages, this is the attack vector to understand. | IT & Security | Legal & Compliance |
Bot internet traffic overtook human internet traffic for the first time this week. The majority of web traffic is now AI crawlers and automated systems — not people. For marketing teams optimizing for human readers, the audience has shifted in ways that aren’t visible in most analytics dashboards. | Marketing | IT & Security |
Published by Independent Intelligence — an independent AI newsletter for people who use AI at work, not just read about it. Five newsletters, one Substack. Friday general edition + Wednesday deep-dives on Claude, ChatGPT, Gemini, and Perplexity. Forward freely.