AI @ Work
Friday, May 29, 2026
Issue #008
This week: how to build an AI that briefs you every morning before you open a single app. Claude gets its most important upgrade in months — and honesty is the headline feature. Robinhood lets AI agents make real trades. YouTube starts automatically flagging AI content. And a Cisco study reveals a serious gap in how we think about AI security. All translated into what it means for your work.
★ This Week’s Big Story
Why this matters to you
Stop being the glue. Here’s how to build an AI Chief of Staff that briefs you, preps you, and tracks what you’d otherwise forget.
Every morning, most people using AI go through the same loop. Check Calendar. Check email. Check Slack. Open a meeting transcript. Then open the AI tool and ask it to help plan the day.
The AI was useful at every step — but you were still the person connecting everything together. AI gave the output. You moved it into the rest of your work. That’s the human-middleman problem, and writer David Perrell at AI Maker this week published the most practical guide we’ve seen for solving it.
The solution is a four-layer system he calls the AI Chief of Staff:
Source layer — tells the AI where to look: Calendar, Gmail, Drive, meeting transcripts, wherever your work actually lives
Context layer — tells the AI what matters: who you are, what you’re building, who matters, what you promised last week
Skill layer — reusable jobs for daily briefs, meeting prep, and follow-up drafts
Automation layer — scheduled runs so the useful work shows up without you remembering to ask
The result: instead of opening six apps before you know what matters, the brief is already waiting. You review it, make the decisions, and approve anything sensitive — but you’re no longer initiating every mechanical step.
You can build this in Claude Cowork, Claude Code, or Codex. The context layer is the part most people skip — and it’s the part that makes everything else actually work. | Operations | Executive & Strategy | All Readers |
Source: AI Maker (aimaker.substack.com), May 28, 2026
AI Platforms
Claude / Anthropic
If you use Claude for research, analysis, or any work where accuracy matters, this week’s upgrade is the one that changes the daily experience.
Claude Opus 4.8 Launches With a Headline Feature You Don’t Usually See: Honesty
Anthropic launched Claude Opus 4.8 this week with performance improvements across coding, reasoning, and agentic tasks — but the most important upgrade isn’t a benchmark score. It’s that the model is now significantly more likely to tell you what it doesn’t know.
Early testers including Shopify, Cursor, and Databricks found that Opus 4.8 flags uncertainties rather than asserting knowledge it doesn’t have, and catches its own code errors at four times the rate of the previous version. For anyone using Claude to support real decisions, a model that says “I’m not sure” instead of confidently making something up is measurably more useful.
Two new capabilities also launched:
Dynamic Workflows — Claude can now run hundreds of parallel subagents on a single complex task in one session
Effort Control — lets you dial between speed and depth depending on what the task needs
Anthropic also confirmed a new, even more capable model class is coming in the weeks ahead. | Operations | Executive & Strategy | IT & Security |
Source: The Deep View, May 29, 2026
ChatGPT / OpenAI
If you manage investments or work in financial services, this is the week AI moved from advising to acting.
Robinhood Now Lets AI Agents Make Real Trades — Not Just Recommendations
Robinhood announced this week that users can connect AI agents directly to their brokerage accounts to execute trades automatically. The agents — built using OpenAI’s infrastructure — can monitor conditions, execute orders, and manage positions without the user approving each transaction.
This is a meaningful threshold. AI moving money in the real world, not just suggesting what to do with it.
For retail investors using AI as a research tool, this is the logical next step. For financial services professionals, the questions that follow are worth tracking: how do regulators respond, and what liability do firms carry when AI executes trades on a client’s behalf? Robinhood’s framing — democratizing access to automated strategies previously available only to institutional investors — is compelling. The regulatory conversation is just getting started. | Finance | Legal & Compliance | Executive & Strategy |
Source: The Neuron, May 28, 2026
Gemini / Google
If your team produces video content — or is considering AI-generated video — YouTube just changed how that content gets labeled, and the change is hard to ignore.
YouTube Is Now Automatically Labeling AI-Generated Content — and the Label Is Much Harder to Miss
YouTube rolled out a significant update to its AI disclosure rules this week. The changes are visible:
For regular videos, the AI label now appears directly under the player, above the description — previously it was buried inside the description
For Shorts, the label appears as an overlay on the video itself
More significantly, YouTube is now automatically detecting and applying labels to AI-generated content — not relying on creators to self-disclose. Labels tied to YouTube’s own AI tools (Veo, Dream Screen) or confirmed by metadata cannot be removed.
For teams using AI in content production, disclosure is no longer optional or easy to bury. It’s the first thing the audience sees. | Marketing | Creative | Legal & Compliance
Source: Mindstream, May 28, 2026
Apple
If you’ve been waiting for your iPhone to actually function as an AI work tool — this is the most promising signal yet that it’s coming.
iOS 27 Siri Redesign Leaked: Dynamic Island Integration, a Chatbot Interface, and Apple’s Most Ambitious AI Push
Bloomberg published leaked iOS 27 designs this week showing Apple’s most significant Siri redesign since 2011. Two changes stand out.
First, Siri moves into the Dynamic Island at the top of the screen — becoming a persistent AI presence rather than something you summon and dismiss. Second, a separate chatbot-style app will support the kind of extended, back-and-forth AI work sessions that currently require a third-party app.
Apple’s full announcement is expected at its Worldwide Developers Conference in June.
The caveat: Apple’s AI history includes the $250 million Siri settlement we covered in Issue #006 — announcements and delivery don’t always land at the same time. But the scale of this redesign suggests Apple understands it can no longer afford to be the laggard in this race. | Operations | All Readers |
Source: TLDR Newsletter · Bloomberg, May 29, 2026
Cross-Platform / Security
If your organization uses AI agents — or is planning to deploy them — this is the most important security finding of the week.
Cisco Tested 15 AI Models From Every Major Lab. Not One Is Safe From Multi-Turn Attacks.
Cisco’s AI threat research team this week challenged how organizations think about AI security. The study compared how 15 models from OpenAI, Anthropic, Google, Amazon, and xAI respond to single-turn attacks (standard benchmark tests) versus multi-turn attacks (an adversary iterating across multiple exchanges).
The gap was significant. Vulnerability rates rose 71% after just five conversation turns. Some highlights:
Claude had the lowest single-turn vulnerability rates of any model tested
Even Claude’s vulnerability rose substantially under multi-turn conditions
GPT-5.4 showed a 9x increase in attack success rate moving from single to multi-turn
The conclusion: the benchmark on your vendor’s model card tells you how the model performs in a controlled test. It does not tell you how it performs when someone is actively trying to break it.
For organizations deploying AI agents — tools that access files, execute tasks, and act on behalf of users — defense in depth is now a baseline requirement: runtime guardrails, input/output monitoring, and real-world red-teaming, not just pre-deployment benchmarks. | IT & Security | Legal & Compliance | Executive & Strategy
Source: The Deep View, May 28, 2026 · Cisco Research
Tools & Workflow — Try These This Week
For anyone whose prompts keep producing mediocre results — the problem is probably not the AI.
Claude Rewards Better Prompts More Than Ever. Here’s the Practical Upgrade.
Cyber Corsairs published a sharp observation this week: as Claude has become more capable, the gap between a lazy prompt and a well-constructed one has widened, not narrowed. A vague ask now produces a noticeably more generic result — because the model has the capacity to go much deeper if you give it direction.
The fix is to treat your prompt like a brief to a talented colleague. Four elements that consistently improve output:
Context — who you are and what you’re working on
Goal — what you’re trying to accomplish
Success criteria — what a good result looks like
Prior attempts — what you’ve already tried
The model’s capability has grown. The question is whether your prompts are growing with it. | All Readers | Operations |
Source: Cyber Corsairs, May 27, 2026
For anyone who accepts AI’s first answer as the final one — there’s a better way.
1 Prompt. 3 Paths. A Probability on Each. How to Use AI for Options, Not Just Answers.
AI Business Insights shared a framework this week built around a simple shift: instead of “what should I do about X?” ask “give me three different approaches to X, and tell me the probability you’d assign to each being the right call.”
The three-path prompt forces the AI to surface assumptions, trade-offs, and edge cases it would otherwise skip in service of one clean answer. The probability framing reveals how confident the model actually is.
When all three paths get similar probabilities, the situation is genuinely uncertain — more of your judgment is needed. When one path is rated 80% and the others 10% each, the model has a real view and you can interrogate it.
Works for decisions, strategies, email approaches, project plans — anything where the first answer shouldn’t automatically be the last word. | Executive & Strategy | Operations | All Readers |
Source: AI Business Insights, May 28, 2026
Worth Watching
The AI industry’s valuation numbers this week were extraordinary, even by recent standards. Anthropic raised $65 billion at a $965 billion valuation — passing OpenAI to become the most valuable AI company in the world. The OpenAI IPO is reportedly targeting September at a $300 billion valuation. SK Hynix and Micron, two of the chip suppliers powering all of this, both crossed $1 trillion in market capitalization on AI demand alone. One number worth holding alongside all of those: the tools discussed in this newsletter start at $20 per month. | Finance | Executive & Strategy |
DuckDuckGo downloads are up 30% as users push back against Google’s AI search overhaul. The privacy-focused search engine is seeing its strongest growth in years — driven not by a new feature but by frustration with what AI has done to Google results. When search starts feeling less like finding information and more like reading a summary someone else chose for you, people look for alternatives. | Marketing | Executive & Strategy |
Glean hit $300 million in annualized revenue this week — up from $100 million just 15 months ago. Glean indexes a company’s internal data across Slack, Drive, Notion, Salesforce, and dozens of other tools, then lets employees search across all of it in plain language. The growth rate is one of the steepest in enterprise software history and suggests that internal knowledge search — finding what your own organization already knows — is one of the clearest near-term AI use cases. | IT & Security | Operations | Executive & Strategy |
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.