| 月曜日 · 2026年6月1日 |
Issue № 007 |
15 min read |
|
|
The FindSkill Weekly Brief
The Skill
Just for FindSkill Pro members. The AI news that actually matters for your work.
|
|
Pro Members Only
|
A Private Brief
|
Not Published Anywhere
|
|
|
|
Mia
AI Learning Editor · FindSkill.ai
|
Welcome to Issue 007.
I'm Mia. Every Monday I land in your inbox and tell you what the week's AI news actually means for how you work — minus the jargon, the hype, and the "10 things you MUST know" energy. Pro members only, posted nowhere else.
Anthropic had one of those weeks. Opus 4.8 dropped on Wednesday. A funding round that values the company past OpenAI — north of $900 billion — was closing alongside it. And the quiet one I keep circling: KPMG just put Claude in front of all 276,000 of its people. Stack those three up and they say the same thing. AI stopped being a clever tab you open and became something organizations run. Which means the skill that matters next isn't prompting. It's governing.
A reply note. Last week I asked which weekly task you'd grade with a five-line rubric first. A strong batch of you wrote back — client-facing work led by a mile, with invoice approvals and hiring-screen notes close behind. Those rubrics go out this week.
— Mia
|
01 |
This Week in AI
Three stories worth your attention
|
Model Release
Claude Opus 4.8 shipped with a dial for how hard it thinks — and the power to run its own fleet.
On May 28, Anthropic shipped Opus 4.8. Two changes matter more than the benchmark bump (88.6% on SWE-bench Verified, if you're counting). First, effort — a dial, now sitting right next to the model picker, that sets how hard Claude thinks before it answers: High by default, up to Extra and Max, down to Low for quick jobs. Second, dynamic workflows — hand Claude Code a big task and it writes its own orchestration script, spins up tens to hundreds of parallel sub-agents, and checks their work against your tests before handing it back. One developer (@SYMBiEX) watched it spawn 85 agents for a single task without being told a number. As another (@0xCodez) put it: "most people still babysit one Claude window — while the people who get this are handing off entire migrations and walking away."
|
What it means for you
You don't have to write code to feel this. The unit of AI work is shifting from "one chat you supervise" to "a fleet you direct and then verify" — and the lever now in your hands is effort. Open whatever AI you use (Claude, ChatGPT, and Gemini all have a version of this control now), find the effort/reasoning setting, and stop leaving it on Max "just in case." Which one to pick when is literally this week's Term.
|
Business
GitHub Copilot's flat fee dies tomorrow. The meter starts running June 1.
Starting June 1 — tomorrow — every GitHub Copilot plan moves from a flat monthly fee to usage-based AI Credits, where one credit equals one cent. Your seat price doesn't change (Pro stays $10, Business $19 a user). What changes is that your prompt, the response, and the model you pick now spend credits — so a heavy week of AI chat or agent runs can push you past what your plan includes. Plain code completions stay free. The part developers are bracing for: when your credits run out, Copilot used to quietly downgrade you to a slower model and keep going. Now it stops — unless you've deliberately set an overage budget. The mood on developer X, in one widely-shared line: "you will get less, but pay the same price."
|
What it means for you
Before tomorrow, open Copilot's billing settings and look at exactly one number: your additional-spend budget. Leave it at $0 and you cannot be charged a cent over your subscription — Copilot just pauses until the month resets. Raise it on purpose only when a real project needs it. And read the bigger signal past Copilot: flat-rate, all-you-can-eat AI is ending across the board. The meter is coming for every tool you use.
|
Policy
The White House shelved its AI oversight order hours before signing. The guardrails just became your problem.
On May 20, the White House postponed an executive order on AI — hours before the planned signing. And this was already the light-touch version: it would have asked frontier labs to voluntarily share advanced models with the government for up to a 90-day safety review before public release. OpenAI and Anthropic had been at the table negotiating it. President Trump said he "didn't like certain aspects" and worried the order "could have been a blocker" in the AI race with China. Set the politics aside; the practical upshot is the same for everyone. The one piece of federal AI oversight that was on the calendar didn't happen, and nothing has replaced it on the schedule.
|
What it means for you
Don't wait for Washington to define what's safe to do with AI at work — that question now lands on your employer's policy, or squarely on you. If your company has no AI acceptable-use rule, the only real guardrail is your own judgment about what you paste where. That's not a hypothetical: it's this issue's Term and the 15-minute Workflow below, both built to help you see and govern the AI you're already using.
---
|
|
02 |
Term of the Week
The one concept to understand this week
|
| |
|
Term №007
Shadow AI
< shadow AI >
The AI tools people use for real work without their employer's approval or oversight — your personal ChatGPT open in another tab to polish a client email, a free transcription bot quietly sitting in the meeting, a browser extension that "just helps." Useful, everywhere, and completely invisible to whoever's supposed to be guarding the company's data.
|
Think of it like this →
Remember work phones before anyone had a "BYOD policy"? Everyone was already using their personal phone for work email. IT pretended not to notice. It was fine — right up until someone lost a phone with client data on it, and overnight there were rules, passcodes, and a remote-wipe button. Shadow AI is the 2026 version of that exact moment. The tools are already in the building. The only open question is whether your company writes sane rules before something leaks, or scrambles to write them after.
|
|
⚠ Common misconception
"Shadow AI is a security problem you fix by banning the tools." No — banning is what creates shadow AI in the first place. The demand doesn't vanish; it just moves onto personal accounts where you have zero visibility and zero logs. Every serious rollout — KPMG's included — starts by mapping shadow usage, not punishing it, because that map is the clearest signal of where the real demand (and the real value) already lives. You don't stamp it out. You give it a sanctioned, governed path so people stop routing around you.
|
Where you'll hear it: In every IT-security vendor report (the "73% of employees use unapproved AI tools" stat is the new doorbuster). In your HR department's freshly-drafted "AI acceptable use policy." On the first slide of any enterprise rollout deck — it was step one of KPMG's. And increasingly on earnings calls, tucked under "AI governance," as the thing CISOs got blindsided by in 2025.
Lead an AI Rollout for Your Team — the shadow-AI audit is Lesson 2 →
|
|
03 |
Deep Insight
What KPMG decided before handing Claude to 276,000 people — and what to steal at any size
|
There's a number bouncing around AI Twitter this week — KPMG gave Claude to 276,000 employees — and most of the takes stop at the number. The number is the least interesting part. What KPMG decided before anyone logged in is the part that scales all the way down to a team of five.
KPMG is one of the Big Four — 276,000 people across 138 countries, very likely the firm your firm's auditor works for. On May 19 it signed a global alliance with Anthropic to embed Claude directly into Digital Gateway, the platform its staff and clients already use for tax, legal, and private-equity work. It's easy to read that as "huge company buys huge AI" and scroll on. But the research on rollouts this size is brutal, KPMG clearly read it, and the moves they made in response are size-agnostic. Here's what's worth copying.
| | SUBJECT By the numbers: 276,000 employees · 138 countries · one platform (KPMG Digital Gateway, on Azure) · full rollout targeted for September 2026. The backdrop that explains every decision below: MIT's State of AI in Business found 95% of enterprise GenAI pilots delivered no measurable P&L impact, and S&P Global found 42% of companies abandoned most of their AI initiatives in 2025 — up from 17% a year earlier. As one mid-market AI builder (@snow_w_lee) put it this week: "distribution problem: solved. adoption problem: just started." |
What's worth copying isn't the budget — it's the four decisions KPMG made before a single person logged in. Each one inverts the instinct most teams start with.
| i. | They embedded it where work already happens — they didn't bolt on a chatbot. Claude lives inside Digital Gateway, the tool KPMG's people already open every day — not as a separate app with its own login and its own tab to forget about. The default instinct is "buy seats, send the welcome email." KPMG's was "make the AI disappear into the workflow people already use." A tool in a new tab gets tried twice and abandoned; a tool inside the thing you already use gets used. The team-of-five version: put AI inside your existing doc, CRM, or help desk — not in yet another browser tab nobody returns to. | | ii. | They started narrow, not everywhere. The rollout opens with Tax & Legal and private-equity work — specific, high-value, measurable — and then expands. Not "everyone, every workflow, day one." That 95%-of-pilots-fail number is mostly the story of "we gave everyone access and called it transformation." KPMG picked a beachhead instead. Steal it directly: choose the single most repetitive, measurable task your team does, put AI on only that for a month, and prove the number before you widen. | | iii. | They led with governance — before speed. KPMG's CEO framed the whole alliance around responsible AI, trust, and data boundaries, not "10× productivity." Sounds like corporate boilerplate until you see the flip side: IBM found 63% of organizations that suffered an AI-related breach had no AI governance policy at all. Governance isn't the brake on a rollout — it's what lets the rollout survive contact with real client data. And it doesn't require a 40-page document. It requires four things: what's safe to paste (and what isn't), a named owner for each use case, a human-review gate on anything client-, money-, or people-facing, and basic logging so you can see what's happening. | | iv. | They're measuring usage, not seats. The trap inside every "276,000 people now have access" headline is that access isn't adoption. The number that actually predicts whether this works is weekly active use by role six months from now — not the seat count in the press release. The practitioners watching closest said it flat out this week: "how many use it regularly in 6 months?" The steal: from day one, track who genuinely uses the tool every week and what it changed — not how many logins you paid for. |
What KPMG did vs. what kills most rollouts Input · How most rollouts go Buy seats for everyone. Send a launch email and a link to a "prompt tips" doc. Measure success by login count. Treat AI as a software purchase. Result: a spike of curiosity, then the 95% — no measurable impact, quietly shelved inside a year. | | | Output · How KPMG (and the 5%) do it Embed into the daily tool. Start with one high-value workflow. Set four governance guardrails before launch. Measure weekly active use by role at 30, 60, and 90 days. Result: AI disappears into better work for one team, the number gets proven, and then it expands. Boring, disciplined, and it survives. |
The reflex here is "that's KPMG, that's not me." But strip out the 276,000 and the playbook is the same at any size: embed it, start narrow, govern it, measure real use. The team-of-one version of step one is Section 04 — a 15-minute audit of the shadow AI you're already running. Do that this week. You cannot govern what you cannot see.
|
04 |
The Workflow
One way to use AI at your job this week
|
| |
Run the shadow-AI audit on yourself, before your IT department runs it on you.
Open a blank doc. List every AI tool you've personally touched for work in the last two weeks — be honest, the whole point is the stuff that isn't on an approved list. The personal ChatGPT account. The free meeting notetaker. Copilot. The browser extension that summarizes PDFs. The image generator you used for one slide. For each tool, write three quick things: (1) what work data you actually fed it (a client name? a contract? just a vague question?), (2) whether your company has genuinely approved it — "nobody's said no" is not a yes — and (3) what would happen if that data turned up in a breach or a training set. Then sort every tool into three buckets: green (no sensitive data — keep using it), yellow (sensitive but fixable — move to an approved tool, or strip names and numbers first), red (client, financial, or personal data sitting in an unapproved consumer account — stop, or get it approved, today).
Why it works: This is exactly step one of KPMG's 276,000-person rollout — map reality, not announcements — compressed to fifteen minutes and one person. You can't govern what you can't see, and almost everyone is running more shadow AI than they'd guess. The audit isn't about guilt; it's about converting invisible exposure into a short, fixable list. Most people find the same thing: one genuinely red item (usually a personal account doing company work) and a couple of easy yellows. Naming them is 80% of the fix.
|
Do this week
Do the audit Tuesday morning — it really is fifteen minutes. Then fix the single reddest item before lunch. Nine times out of ten that means moving one workflow off your personal login onto the tool your company actually sanctions. One red item closed beats a perfect policy you never wrote.
|
|
|
05 |
The Side Play
One way to make money with AI this week
|
| |
| ◆ Income Idea · Play №007 |
Become the person who writes the AI rules small businesses don't know they need yet.
Every business under ~200 people is quietly fielding the same question — "are we allowed to use ChatGPT for this?" — and has no answer. The federal guardrail everyone assumed was coming just got postponed (Story 3), and the KPMG headlines have every owner wondering if they're behind. That's your opening. Build a productized AI policy + rollout starter kit: (1) a one-page acceptable-use policy template with the green/yellow/red data tiers from this issue — what's safe to paste, what isn't, who owns it, what needs a human check; (2) the shadow-AI audit worksheet from Section 04, scaled to a small team; (3) a one-page 30/60/90 adoption tracker; (4) a 60-minute "walk your team through it" call script. Sell it two ways. As a $79–$199 digital template pack on Gumroad or Lemon Squeezy for owners who'll do it themselves. Or as a $750–$2,500 productized service — "I'll run your AI-policy workshop and customize the kit for your team" — pitched to local SMBs through your Chamber of Commerce, an industry association, or your own LinkedIn network.
Why it works: The demand signal is screaming from three directions at once — the KPMG news, the postponed executive order, and every owner's overflowing inbox. And the work is judgment, not code: you're packaging what "good" looks like for a 40-person firm that can't afford a McKinsey engagement and doesn't need one. Realistic revenue: $500–$4,000/month solo, and meaningfully more if you niche to a single vertical you already understand — dental practices, law firms, marketing agencies — because then the policy examples are specific and the trust is instant. Not a unicorn. Honest numbers, and a roughly two-quarter window before "AI use policy template" is a $12 commodity.
|
Do this week
Write the one-page policy template tonight — steal the four-part structure straight from the Deep Insight, it's already done. List it on Gumroad this week, rough edges and all. Then DM three local business owners you actually know and offer the workshop. The template sells the service; the service funds the polish.
|
|
|
06 |
The Stack
Three tools I'm actually testing this week
|
01 |
Gamma
AI presentations, docs & sites · Free tier (paid from ~$10/mo)
When I have twenty minutes and need something board-ready, this is the tab I open instead of fighting PowerPoint. You type what you want — "turn these five bullets into a six-slide update" — and then edit by sentence rather than by dragging text boxes around. The free tier is plenty to judge whether it fits how you work; a small watermark on free exports is the only real nag. Not magic, but it's the fastest "rough draft of a deck" tool I've tested this year, and the one I keep coming back to.
|
02 |
Fathom
AI meeting notetaker · Free (unusually generous)
It joins your Zoom, Meet, or Teams call, transcribes it, and hands back a clean summary with action items — and the free plan is genuinely generous, not a teaser. I keep it on for internal calls and off for anything sensitive, which is the whole point of this issue: a notetaker is itself an AI sitting in your meeting, reading everything said. Useful — and exactly the kind of tool that belongs on your company's approved list, not quietly running on a personal login. Vet it first, then enjoy it.
|
03 |
Helicone
AI usage & cost dashboard · Free / open-source
This one's for the builders among you, but it's in the stack because of the Copilot story. Helicone sits in front of your AI API calls and shows you exactly what they cost — per request, per model, per project — before the bill surprises you. Even if you never wire it up, knowing this category exists is the point: as every tool flips to metered pricing, the people who stay calm are the ones who can see their AI spend instead of guessing. The free tier covers a solo project, and it's open-source if you'd rather self-host the whole thing. ---
|
|
07 |
Inside FindSkill
What's new for members this week
|
| New |
Lead an AI Rollout for Your Team
Eight lessons on doing what KPMG just did, sized for a 50–2,000-person team and a budget that isn't Big Four: the shadow-AI audit, right-sized governance (no Center of Excellence required), picking one use case, building a champion network, and the adoption math that separates the 5% from the 95%. The practitioner companion to this issue's Deep Insight. Start the course →
|
|
| New |
Two Claude legal-plugin courses
Claude Law Student and Claude Legal Clinic. The role-specific builds for anyone doing legal work with AI: case briefs, IRAC writing, and bar prep for students; cold-start setup, client intake, and deadline supervision for clinics. Governed prompts, not "paste the case and hope." See all courses →
|
|
| New |
The Opus 4.8 field guides
Five plain-English explainers shipped the day after launch: which effort level to actually pick, the dynamic-workflows and ultracode walkthrough, whether Fast Mode is worth it, the Opus 4.8 vs GPT-5.5 vs Gemini comparison, and the KPMG rollout playbook — plus the June-1 Copilot billing explainer from Story 2. If you read one, the effort guide: it'll save you the "why is this suddenly so slow" confusion this week. Browse blog →
|
|
|
|
What's the one AI tool you use for work that your IT team doesn't officially know about?
Reply with it — no judgment, I keep a list of my own — and I'll tell you whether it'd survive a basic data-safety check. Worst case, your answer becomes an anonymized example in the AI Rollout Toolkit.
↵ Hit reply
|
|
The Skill · by FindSkill.ai
|
|