The BeAIReady Brief | Week 29
July 13–19, 2026 | Microsoft is working to replace the AI partner it paid $5 billion for, seven in ten heavy AI users are submitting work they haven't reviewed, and CEOs are measuring the wrong thing
Big Tech earnings started this week, hot on the heels of a deceptively cooling CPI — prices are still up except for energy, which could rise again as the U.S. continues to push the conflict with Iran — and a labor market that added just 57,000 jobs in June. I’ll be watching those earnings reports for what’s underneath: the structural displacement of knowledge work, and whether it’s showing up in the numbers yet. The stories I was reading last week — from the Fed Chair before Congress to Microsoft reckoning with its own AI supply chain — kept returning to the same friction: the narrative around AI and the execution of AI are not moving in the same direction, and the gap between them is getting harder to ignore.
This week’s coverage:
Microsoft Is Done Renting Its Intelligence
Nadella warned that AI is leaking your IP through every prompt and correction. His sales team was coaching against Anthropic and OpenAI. And he called Fable “editorially controlled” to the engineers building Copilot on it — a notable set of moves inside a $5 billion partnership.
Who’s Accountable When Nobody’s Reading the Output
Seven in ten heavy AI users are shipping work they haven’t reviewed. And a quarter of organizations can’t detect the AI agents running inside them. Two failure modes, one accountability gap.
Underinvested and Measuring the Wrong Thing
Two-thirds of CEOs say they’re behind on AI investment. A separate argument says they’re also measuring it wrong. The problem may be both at once.
The Workforce Is Being Rewritten
AI fluency became a baseline job expectation, $162 million in federal training funds moved, 20,000 engineers were standardized on a single model, and Walmart ran its entire supply chain through AI agents and digital twins — all in one week’s reading.
On the Bigger Picture
OpenAI’s pattern of high-profile partnership collapses, Anthropic’s model-war promotional tactics, and Canva’s push to bring vibe coding to a quarter-billion users.
Here’s what I was reading.
Microsoft Is Done Renting Its Intelligence
Satya Nadella has been making a consistent argument this year: that companies should own their AI stack, not rent it. Last week that argument sharpened considerably across three separate disclosures. The first was a piece from The Next Web naming what Nadella calls the Reverse Information Paradox. The classic information paradox in economics is about disclosure destroying value. What Nadella is describing is the reverse: every time your team uses an AI tool, it gives something back — the prompts you write, the corrections you make, the patterns of how your business actually works. That “exhaust” flows upstream to the vendor. And in most enterprise setups, your competitive intelligence is part of what you’re paying with. (Nadella says you pay for AI twice, and Microsoft helped build the trap)
The same week, a second Next Web piece reported that Microsoft’s internal FY27 sales playbook is coaching salespeople to position in-house AI as cheaper and better integrated — and to steer customers away from Anthropic and OpenAI models, even the ones still embedded in Microsoft products. The goal, per the reporting, is for Microsoft to eliminate what it pays Anthropic — a stark internal framing when the partnership involves $5 billion going to Anthropic and $30 billion in Azure commitments flowing back. (Microsoft is coaching its salespeople to talk down the models it still runs on)
Then came Thursday. CNBC obtained remarks Nadella made to Copilot engineers in which he described Anthropic’s Fable model as “editorially controlled” after it refused certain requests — adding, simply, that “it doesn’t make sense.” The context is worth noting: Fable was suspended briefly in early July to comply with U.S. export control directives, and when Anthropic restored it, the updated safeguards flagged a higher share of benign requests than before. Nadella’s critique is technically grounded — but making it to your own engineering team, in an internal session, about a model your company is both invested in and ships to customers, is something other than a product complaint. It’s a repositioning. (Microsoft CEO Nadella criticizes Anthropic’s Fable AI over refusals in internal Copilot meeting)
Three moves in a single week: a public IP warning, an internal sales motion against external models, and a CEO calling out his own AI vendor to the engineers building on it. That’s not partnership drift. It’s a deliberate reorientation.
Who’s Accountable When Nobody’s Reading the Output
A Glean report published last week gave a name to something I suspect is far more common than organizations want to acknowledge. Nearly seven in ten heavy AI users admit to “botshitting” — submitting AI-generated output without reviewing it. The term is blunt, but the finding is precise: it describes the moment when AI’s speed advantage inverts, because the person signing off can no longer explain or defend the work. The same report found that heavy users also spend significant time “botsitting” — monitoring AI output closely enough to feel some ownership — and that this generates its own problems: disengagement, reduced agency, higher turnover risk. These aren’t risks that surface in most AI governance conversations. Botshitting doesn’t make it onto the compliance checklist. (Heavy AI users submit work they don’t understand, report finds)
At the organizational level, a SecureWorld survey surfaced a structural version of the same failure. Twenty-one percent of organizations cannot detect unsanctioned AI agents operating in their environments — a visibility gap that has more than tripled in a year. More than 88% reported AI-related security incidents in the same period. The prescription is clear enough: continuous automated discovery, dynamic least-privilege permissions, zero-trust controls applied to non-human identities the same way they’re applied to human ones. What the survey doesn’t address is the organizational will to implement before something goes wrong. (Your Organization’s AI Trust Infrastructure Is Failing, Survey Says)
These two pieces sit at different levels of the organization, but they describe the same failure: adoption outrunning the accountability infrastructure built to support it. One is HR’s problem. One is IT security’s. Neither is being solved.
Underinvested and Measuring the Wrong Thing
A survey covered by HR Dive found that roughly two-thirds of CEOs believe they’re underinvesting in AI — with 40% identifying infrastructure modernization as their top priority for 2026. The finding reframes the barrier: it isn’t ambition or budget conviction, it’s the foundational data, security, and network infrastructure that determines whether investment can land at all. Licensing isn’t the bottleneck. The plumbing is. (CEOs fear they’re underinvesting in AI)
A CEOWORLD piece offered the conceptual counterpart. The author’s argument is that the ROI framing most organizations apply to AI — efficiency, cost savings, task speed — captures what AI does to individual work, not what it does to organizations. The proposed shift to “Return on Intelligence” tries to measure something harder to quantify: learning velocity, adaptability, the compounding effect of building an organization that gets systematically smarter over time. The piece also noted that 66% of board members admit limited AI knowledge — which means the governance layer often encodes the wrong incentive structures into the evaluation criteria before anyone realizes it. (The New ROI of AI: CEOs Must Now Measure Return on Intelligence)
The two problems compound each other: an infrastructure gap that limits what investment can accomplish, and a measurement framework that may be evaluating the wrong outcomes even when it does. Getting one right without the other probably doesn’t get you very far.
The Workforce Is Being Rewritten
AI-related job titles have more than tripled in the U.S. since 2022, and 63% of them now sit outside traditional technology companies. That number is less about titling convention and more about expectation: AI fluency is becoming a baseline job requirement across healthcare, finance, operations, and logistics — not just in roles where anyone would have anticipated it. The skill level often implied is basic; the signal being sent is not. (More job titles include AI across every sector)
The Department of Labor moved last week to accelerate the pipeline. Nearly $162 million in cooperative agreements went toward expanding Registered Apprenticeships in AI, semiconductors, shipbuilding, nuclear energy, and defense manufacturing. The investment is significant not just for its scale but for its framing: AI workforce development is now a national infrastructure priority, treated in federal budget terms alongside energy and defense. Apprenticeship — credentials through work, not through degree programs — is the vehicle of choice. (Latest workforce development funds target specific industries — like nuclear energy and AI)
At the enterprise level, what this shift looks like in practice showed up in UST’s announcement with Anthropic. The IT services firm is standardizing 20,000 developers across healthcare, banking, telecom, and hardware practices on Claude — building reusable governance frameworks, training programs, and AI workflows around a single model. The implication I keep returning to is the one about decision authority: AI model selection is moving from individual developer preference to organizational platform policy, with all the change management complexity that entails. (Anthropic’s newest enterprise partner is training 20,000 people on Claude — here’s the shift it signals)
Walmart filled in the operational picture. Its supply chain team is deploying AI agents, LLMs, and digital twins to optimize across assortment, speed, and cost — across a network of more than two million employees. What the Walmart piece made clear isn’t the technology; it’s the decision architecture: AI agents aren’t replacing logistics judgment, they’re compressing the time between signal and response enough that human decision-makers can act before disruptions cascade. That framing of human-AI collaboration is more honest than most enterprise AI marketing manages. (Walmart bets on AI and digital twins to shape its supply chain strategy)
On the Bigger Picture
A Business Insider piece last week traced the pattern of OpenAI’s partnership collapses — first the rupture with Microsoft, now Apple, which is reportedly suing over employee poaching and trade-secret theft. Two major distribution partners, two legal disputes, within roughly a year: the pattern suggests that OpenAI’s most important relationships carry structural friction that the promotional framing around them has consistently understated. (Why does OpenAI keep breaking up with Big Tech?)
Anthropic extended free access to Fable 5 through July 19 — the second extension in a week — as a direct counter to OpenAI’s GPT-5.6 launch. The move is less interesting for what it costs Anthropic than for what it reveals: promotional access extensions have become a standard competitive response at the frontier, which says something about how difficult meaningful differentiation has become. (Here’s Why Anthropic Extended Access To Claude Fable 5 Extended—Again)
Canva launched Code 2.0 to all 265 million of its monthly users last week, its most aggressive push into vibe coding. The market is estimated at $4.7 billion in 2026 and projected toward $12.3 billion next year; well-capitalized competitors in Lovable, Replit, and Bolt.new already own significant ground. Canva’s differentiated bet isn’t that it can generate working code — it’s that the real bottleneck for non-technical users has never been generation, it’s been making the output look good enough to actually deploy. Integrating generated code directly into the Canva design environment is either a smart insight about where vibe coding breaks down for most users, or an overextension into a space with very different expectations. I’m curious which it turns out to be. (Canva launches Code 2.0, offering AI website building to every user — including free accounts)
Depth of AI commitment, this week made clear, is not the same as readiness for what commitment reveals. Microsoft built its AI strategy around Anthropic — and is now coaching against it. Enterprises moved fast on licenses — and discovered they can’t see the agents they’re running. Workers adopted AI first — and are now shipping work they can’t explain. The organizations furthest into AI aren’t further along because they’ve solved these problems...they’re further along because they’ve had them longer. That’s not a failure of the technology. It’s what happens when adoption outpaces the organizational capacity to absorb it — and it’s the problem that doesn’t show up in the deployment metrics.
That’s it for this week’s BeAIReady brief!
If you appreciate the depth of reporting and how I connect the dots, please like, share this post, and subscribe (or share the Brief with a friend!). Thanks!
~erick