The Fed held rates last Wednesday on a 10–2 vote. The markets responded with the expectation of hike for December as emminent. The problem with that of course, is that debt-financed AI ends up costing a lot more to carry. And last week, signs of that problem are beginning to show up in a lot more places: Microsoft sold $25 billion in bonds, Anthropic’s S-1 showed hundreds of billions in compute commitments, and Gartner found that 97% of organizations can’t cut costs on productivity gains alone. Although the vendors can finance the build, customers can’t yet show the return, and the cost keeps falling on employees — as buyouts, reskilling, and skills that fade when AI ends up doing those tasks for them. So the big question now is — who’s actually going to pay for all that productivity?
Last week’s coverage:
Microsoft Borrowed $25 Billion. Its Customers Can’t Find the Savings. Vendors are raising capital at record scale while the buyers can’t point to what they saved.
Someone Else Pays for the Productivity Dividend Buyouts, reskilling and fading skills show where the cost of AI efficiency lands.
You Can’t Govern What You Can’t Reconstruct Copilot memory, usage scorecards and unlogged agents all raise the same question: who can see what happened?
Employees Are Already Ahead of the Mandate People are building, experimenting and running models locally while the operating conditions around them lag.
Here’s what I was reading.
Microsoft Borrowed $25 Billion. Its Customers Still Can’t Find the Savings.
A Reuters piece reported that Microsoft’s $25 billion bond sale drew more than $100 billion in orders, over four times the offering. AI financing has turned into a balance-sheet advantage: the companies with the best credit borrow their way to capacity while everyone else rents it. (Microsoft expects to raise $25 billion in US bond sale, sources say)
Meanwhile, Techcrunch reported that Microsoft booked $10 billion from OpenAI in the quarter ending September 30 — nearly 11% of its total. The filing didn’t split Azure services from revenue-share payments, which means that Microsoft can point to exactly where its AI money comes from. Most of its customers can’t. (Microsoft recorded $10 billion in revenue from OpenAI)
Anthropic’s S-1 on October 26 showed first-half revenue of $16.3 billion, up roughly 1,300%, and an $80.7 billion net loss. Most of that loss is a non-cash accounting charge tied to the company’s own rising valuation (the operating loss actually narrowed to $693 million). The risk factors still warn that costs will rise and losses will continue. The fastest-growing revenue story last week came with a written warning that spending will stay ahead of profit. (Anthropic’s IPO prospectus shows $16.3 billion in revenue, an $80.7 billion net loss, and goal of an ‘age of miracles’)
On the customer-side of the equation, a Gartner survey shows that a whopping 97% of organizations can’t cut expenses through productivity gains alone, 38% are over budget on AI — and yet, 85% still plan to spend more next year. Only 21% of the time employees save gets redirected to more innovative work. Customers are spending more on AI and getting back time they can’t convert into savings or new work. (Companies are not saving money with AI, Gartner finds)
A correction seems eminent… US tech stocks lost roughly $835 billion in market capitalization in October from AI fears, but have already mostly recovered. China, on the other hand, is experience a different story. A Reuters piece reported that the STAR50 index fell 40% in under four months, erasing about $480 billion in market value, and have yet to rebound. Several causes including the variety of technologies companies involved and a variety of domestic challenges. I’m reading it as a valuation warning, not necessarily a verdict on demand. The bottom line is that executive enthusiasm for AI spending isn’t evidence that there’s an operational payoff, and a market correction is what happens when that gap gets priced in. (In China, the AI bubble has already burst)
Someone Else Pays for the Productivity Dividend
Microsoft is offering buyouts to roughly 60,000 of its 223,000 employees. The same Fortune piece put the company’s projected fiscal 2026 spending at $145.3 billion. It didn’t claim one caused the other. But readers like me will likely make the connection anyway. (Microsoft offers buyouts to nearly 30% of workforce in ongoing AI shift)
An HR Executive piece summarized McKinsey’s estimate that 6 million to 16 million U.S. workers may need to move into new occupations. That makes reskilling a planning problem for every HR team, not a training-budget footnote. The article didn’t say how organizations should build the pathways, which is the part leaders actually need. (Report: More than 70% of workers may have to reinvent their roles)
A Fortune piece on JLL carried a survey of 1,321 organizations: 89% expect facilities budgets to shift from labor to technology, and 87% expect outcome-based pricing by 2030. When a contract pays for outcomes instead of hours, every labor hour comes straight out of the provider’s margin. JLL’s CEO says he despises the word “headcount.” The contracts his customers want don’t share the sentiment. (The humanist at the top of real estate’s automation machine)
An HR Dive piece on Dayforce research landed closest to the people doing the work. More than 6 in 10 workers said they felt they were losing the ability to do tasks themselves because of AI, and 40% expected their employers to give them more work if AI saved time. Workers expect the saved hour to come back as more work, and many already feel their own skills slipping. So who decides what that hour is for? Right now it’s whoever gets to it first. (AI at work: Employees, employers may not be on the same page)
You Can’t Govern What You Can’t Reconstruct
An Office 365 IT Pros piece covered Microsoft’s separate retention policies for Copilot interactions, rolling out from late October to mid-November, plus retention for Copilot memory. Organizations that use older Purview policies for both Teams and Copilot data need to split them. Copilot memory is a record, and organizations that treated it as a feature will find that out in a retention review. (Copilot Retention Policies Cover for Interactions and Memory)
A Mashable piece covered Claude’s new smart reports, which give employers visibility into team AI usage and spending and can classify sessions as “personal or off-topic.” Anthropic advises against using them for employment decisions. The article didn’t establish how accurate the classifications are. Reports that can flag a session as off-topic turn AI use into a scorecard before anyone has agreed what good use looks like. (Claude’s new smart reports tell bosses when workers are ‘off-topic’)
A MIT Technology Review piece looked at the four most widely used agent frameworks and found gaps in persistent agent IDs and in records of which humans approved which tasks. If you can’t say which agent acted, who approved it and what changed, you haven’t deployed it safely. Put the three together: retention covers what AI remembers, usage reports cover what employees did with it, and agent logs cover what it did on its own. The third is the least built. (We urgently need to monitor AI agents’ actions in real time)
Employees Are Already Ahead of the Mandate
A CIO Dive piece reported that Liberty Mutual employees have built more than 30,000 assistants and Scotiabank’s AI platform has reached over 70,000 employees. Both paired broad access and training with security, governance and cost visibility. That pairing is what separates a sanctioned assistant boom from shadow AI. The article offered few measured outcomes, so the results are still unproven. (How Scotiabank, Liberty Mutual scale AI while watching costs)
Another CIO Dive piece told the other side. One chatbot use case at a client exposed that seven business lines had no centralized data strategy. An AI mandate without shared data, clear authority and measurable value pushes pressure down the org chart without giving anyone the means to act on it. (How to create a scale-ready AI mandate)
An HBR piece found that 70% of executives reporting great digital transformation progress said they often or always relied on younger, digitally fluent employees, against 37% of those reporting less progress. The leaders making the most progress lean on their most digitally fluent people far more often. The article didn’t show that the reliance causes the progress. (Empower Your Most Curious Employees to Drive Your Digital Transformation)
Microsoft’s announcement of its Copilot harness pitched a managed layer that supports more than 20 frontier models, so organizations don’t build and maintain their own. The pitch is that model choice is solved if you rent the scaffolding. It offered no comparative evidence on cost, outcomes or how the harness handles poor-quality organizational knowledge. Would your team’s workflows survive a model swap tomorrow? (Introducing the Microsoft Copilot harness)
An XDA Developers writer ran a 9B model, just under 7GB quantized, on an 8GB RTX 3070. It handled explanations and logic problems and struggled with coding and design. A useful assistant now runs on one employee’s graphics card, outside every policy written for procurement. (I ran a 9B model trained on Claude’s reasoning on my 8GB GPU)
Microsoft borrowed $25 billion, Anthropic told investors its losses will continue, and Gartner found that 97% of organizations can’t cut costs on productivity alone... the money is moving faster than anyone’s ability to say where the saved hours went. Meanwhile the people absorbing the difference are taking buyouts, retraining for new occupations, and watching their own skills fade. Are leaders who hand employees an AI mandate, a usage scorecard and a smaller headcount running an effective AI strategy? In my opinion, the answer is clearly no.
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