Software’s Rebound Reflects a Longer Timeline for AI Disruption

Alex Kantrowitz frames the current AI debate as a timing problem. The models may improve rapidly, and the eventual disruption of enterprise software may still be substantial. But neither organizational deployment nor financial markets move at the same rate as model capability.

That distinction explains the apparent reversal in software. Investors had treated increasingly capable AI systems as an immediate threat to the subscription economics of companies such as Salesforce, Adobe, and Microsoft. The concern was straightforward: customers could use models to accomplish more work with fewer software seats, build tools they had previously bought, or eventually replace applications altogether.

Salesforce’s results challenged the implied timetable rather than disproving the threat. A CNBC article shown during the discussion described investors’ earlier worry that AI would weaken traditional software subscriptions and enable companies to make their own tools. Salesforce had fallen 22% through the Wednesday before its earnings report, then rose more than 12% after reporting a strong quarter and better-than-expected guidance.

Kantrowitz cited a much broader market rebound. By August 28, he said, Salesforce had risen 41% over the preceding month and 72% since June 22, though it remained down 1.94% for the year. He also cited one-month gains of 19.68% for the iShares Expanded Tech-Software Sector ETF, 16.49% for Adobe, and 29% for Microsoft.

The figures shown on screen make the same point visually: a Google Finance chart for IGV showed a 19.68% one-month increase, while corresponding charts showed Salesforce up 41.87% and Adobe up 16.49% over the month. The reversal is not evidence that software is permanently insulated. It is evidence, in Kantrowitz’s reading, that markets had priced a faster transition than enterprises can execute.

A company that has run its customer operations through Salesforce for years does not discard its accumulated workflows merely because an AI assistant can answer some questions about accounts. It must decide what to migrate, which records and permissions need to remain governed, where new tools fit in established routines, and how to manage the operational risk of a change. Even if an AI-native alternative proves better, Kantrowitz argues, adoption could take three, four, or five years rather than a quarter.

That lag creates a period that is consequential for both incumbents and AI companies. Salesforce has time to revise its product and business model. But it still has to determine whether it will remain the place where employees work or become an underlying database accessed through somebody else’s assistant. For investors, the relevant question is less whether AI will matter than whether a current valuation assumes the eventual answer has already arrived.

The Salesforce–Anthropic partnership branded “Claudeforce” became the clearest symbol of software’s temporary reprieve. A CNBC frame shown in the discussion placed Marc Benioff beside Anthropic CEO Dario Amodei beneath a banner describing an integration of Claude into Salesforce sales tools. Amodei described a use case in which Anthropic’s chief commercial officer could ask about the company’s largest accounts and the risks involved in closing them, with the relevant information managed in Salesforce.

Ranjan Roy does not regard that basic product concept as novel. At Writer, where he works, people already connect Salesforce data to an AI interface and query it. He said similar workflows are available through ChatGPT. Asking which accounts need attention, what risks are attached to them, or what sales work should come next can be useful; the question is why that familiar capability required such an elaborate public presentation.

For Roy, the answer was communications rather than technical novelty. Benioff’s ability to turn the announcement into a market event was itself notable: Roy pointed to Salesforce’s 22% stock jump as evidence of the value of the appearance. More striking was seeing Amodei publicly discuss the mundane mechanics of enterprise selling alongside a major SaaS executive. Roy said he had not heard Amodei speak publicly in that register before.

Roy’s more speculative interpretation is that Anthropic was presenting a friendlier face ahead of an IPO. A company portrayed as the “death star” that will absorb every enterprise software category may have an expansive long-term valuation story, he argues, but it also makes prospective public-market investors and established technology partners uneasy. Sitting beside Benioff allowed Amodei to show that Anthropic could work with incumbent software companies instead of merely threatening to replace them.

Kantrowitz initially offered a more ordinary explanation: it was the final quiet week of summer, and the companies had an opportunity to put the two executives together. But he came to agree that the pairing carried clear branding value. It made Anthropic look less like a hostile successor to enterprise software and more like a collaborator.

The underlying incentives nevertheless point in different directions. Kantrowitz’s concern is that, if advanced assistants become capable of handling sales work end to end—listening to customer conversations, conducting interactions, organizing information, and deciding what should happen next—a conventional CRM interface could become unnecessary. His hypothetical name for the more disruptive endpoint was “Sales Claude”: not Claude querying Salesforce, but an assistant that directly owns the sales workflow.

Roy believes that outcome is central to Anthropic’s longer-term growth narrative. In his view, a company pursuing a two-trillion-dollar valuation cannot ultimately depend only on helping other software companies preserve their existing positions. It needs to capture work now performed within applications such as Salesforce.

For Salesforce, the risk lies in giving up the interface while retaining only the data layer. Roy sees genuine value in Salesforce as a secure, governed, structured place for customer information. But he questions whether that role can sustain the margins associated with traditional enterprise software if employees begin their day in Claude or another assistant rather than in Salesforce itself.

The relevant measures during this transition are therefore operational, not rhetorical. Does the incumbent retain the workflow through which people make decisions? How expensive and disruptive is migration for the customer? Does the vendor still control the interface at which work is initiated, reviewed, and completed? And can its data layer remain valuable if the primary user relationship shifts to an AI system owned by someone else?

Software’s rebound buys time to answer those questions. It does not answer them.

Meta’s Project OT, short for Organization Transformation, provided a second version of the same timing problem. Reuters reported that the company envisioned an “AI native” workforce in which AI would take over much of the daily work performed by thousands of employees. Smaller, “talent-dense” groups of people would oversee virtual workers. In scenario-planning exercises, Meta executives explored cutting the size of many teams by as much as 60%.

The Reuters documents shown during the discussion described a more comprehensive organizational redesign than simply giving employees coding tools. Meta planned to make traditional product-design and engineering roles disappear into a generic “builder” title, remove layers of middle management, and use “agent-assisted analysis” to set day-to-day priorities.

The initial pilot involved five small tech pods, each with two or three engineers and a designer equipped with AI tools. Instead of working through fixed six-month planning cycles, the pods would build prototypes in four-week sprints. Reuters reported that by June, at least 11 engineering and research units had implemented small pods.

Roy’s response was not that the model was inherently misguided. Separated from the planned layoffs, he said, this is exactly the kind of operating model companies should be testing: leaner teams, shorter cycles, fewer managerial layers, and deliberate use of AI. The failure, as he sees it, comes when a company makes AI usage itself the objective rather than asking whether a team is completing useful work reliably.

Reuters’ internal reporting exposed the difference. AI-assisted development was generating substantially more activity, but the relationship between that activity and user value was weak. Changes to internal software platforms and infrastructure rose 220% year over year; changes that resulted in new or upgraded features reaching Meta users rose 36%. At the same time, infrastructure teams warned that unchecked agents were undertaking disruptive actions that people would be unlikely to take.

The Reuters Project OT document shown on screen explicitly linked the experiment to smaller AI-equipped pods, eliminated management layers, and agent-assisted priority setting. The later figures give the model its practical test: the company could generate more code while simultaneously shipping only modestly more user-facing change and spending far more time dealing with incidents.

Kantrowitz treated that as a concentrated version of a broader AI implementation problem. More generated code is not automatically more productive work. If a team can create changes faster than it can test, secure, operate, and review them, the extra throughput can create a backlog of failures. Service disruptions, possible data leaks, and remediation work are not peripheral costs; they are part of the real output of the system.

Roy attributed much of the problem to incentives. Employees who believe they are being evaluated on speed, token consumption, or the amount of AI they use will optimize for those measures. The better test, he argues, is whether a lean team can complete a defined project with quality. AI should be a means of reaching that outcome, not a scorecard category that substitutes for it.

Meta eventually curtailed the broader workforce plan. Kantrowitz said the company proceeded with a 10% cut but called off planning for a larger November reduction. Zuckerberg told employees not to expect further company-wide layoffs that year and, according to remarks obtained by Reuters, described Meta’s two principal expense categories as compute and “people-related” costs. As compute spending rises, the latter, in his formulation, has to decline.

Kantrowitz rejected the idea that this makes layoffs mechanically necessary. Meta is, he argued, a highly profitable company with massive margins, and its willingness to spend billions on AI talent shows that it will pay for people it considers valuable. The real judgment embedded in Project OT is not that compute inevitably displaces labor. It is that particular workers or layers of work are not valuable enough to keep.

That makes the implementation question more demanding than “Can AI do this task?” Companies also need to ask what quality controls must remain human, how work should be evaluated, which actions agents should not be allowed to take unchecked, and whether the output reaching customers has actually improved.

The South Korean market episode shows what happens when a plausible long-term AI thesis is treated as a smooth, immediate financial trajectory. Kantrowitz cited Wall Street Journal reporting that the Kospi had more than tripled, driven by confidence in the AI boom and by the rise of Samsung Electronics and SK Hynix. It then fell about 40% over six weeks in June and July before recovering roughly 20% from its lows.

For unlevered holders, a 40% decline followed by a partial rebound is already a severe drawdown. For investors using double- and triple-leveraged single-stock ETFs, it can be ruinous. Kantrowitz’s simplified example was a 5% move in a stock: it becomes a 10% loss in a double-levered product or a 15% loss in a triple-levered one. A prolonged decline can erase a concentrated position.

The investors involved are known in South Korea as “ants”: individually small traders with collective market power. Kantrowitz said they account for 60% to 70% of the Kospi’s daily trading volume. Roy noted that the label emerged during the COVID-era trading boom and had taken on renewed relevance during the AI-driven semiconductor surge. The double- and triple-levered ETFs discussed entered the market at the end of May, shortly before the sharp swings.

The result was not merely an abstract market correction. Kantrowitz cited Yun Kyung-min, a 44-year-old sound engineer who had put half of his severance pay into semiconductor stocks after leaving his job and saw $7,200 disappear in a week. He also cited Lee Ka-young, a 25-year-old software developer who began investing after deciding her paycheck would never be enough to buy a home.

Lee put roughly $14,000 from savings and crypto gains into SK Hynix and later into single-stock leveraged ETFs. She opened an Instagram account to document what appeared to be a rise to riches. By May, Kantrowitz said, the account was up more than 58% and her posts drew an average of four million views. After the reversal, her holdings were worth 5% less than when she had begun.

That example connects the market story to the discussion about social-media design. Meta’s $18 billion settlement with nearly all U.S. states over claims that Facebook and Instagram were designed to addict children included product changes discussed by the speakers: like counts turned off by default for minors, access to a non-algorithmic chronological feed, a default two-hour daily limit for users under 18, overnight restrictions, and no notifications during school hours. The states had sought $200 billion in civil penalties.

Roy argues that these details matter because recommendation, visible popularity metrics, and notifications do not merely surround a product; they shape behavior within it. He favors going beyond the settlement by removing engagement counts for ordinary users and making chronological feeds the default. He also sees Meta’s willingness to accept limits as a competitive maneuver. Reported discussion of a one-hour daily limit if TikTok and YouTube followed suit, he argues, would put pressure on rivals that are also dependent on algorithmic engagement.

The same mechanisms can apply to investing. Roy described Lee’s Instagram account as an unsettling example of investment content becoming entertainment, public performance, and audience-building around a trade. A feed can repeatedly present the same success story, turn participation into identity, and make a crowded position feel like a collective route to security.

Roy does not argue that every AI-related investment is irrational or that South Korea necessarily forecasts a systemic crisis. His comparison with the period before the 2008 financial crisis is narrower. Before Lehman, he recalled, there were earlier signs of strain: Countrywide, Bear Stearns, and hedge-fund collapses. Such episodes did not settle the entire market narrative, but they revealed vulnerabilities in specific places.

The vulnerability here is household exposure to concentrated, levered positions. Roy contrasted it with 2008, when a typical person’s central exposure was often a home and its equity. Trading platforms now make it easier for more people to take direct positions in volatile themes. The relevant question is not only whether the AI boom eventually receives the label “bubble.” It is whether people who share the thesis can survive the air pockets before its long-term outcome is known.