AI Isn't Eating Software — It's Dividing It Into Winners and Losers

Ask a lender rather than a stock picker whether AI will destroy the software industry, and you get a more useful question in return: which software businesses get paid back, and which don't. A creditor is repaid from cash flow, not from anyone's price target, so he has to be right about durability. Blair Jacobson, co-president of Ares Management, one of the largest private credit firms in the world, carries that obligation. Speaking this week at the IPEM conference in Paris, Jacobson said the software market has split into a "bifurcation", and that the earlier fear that "all software will get eaten by AI" has not come to pass.

That fear was right and wrong at the same time. Software stocks did not all collapse; they separated. In the first half of 2025, returns across the sector ran from near +80 percent — Palantir, Zscaler, CrowdStrike, Snowflake — to about -43 percent, a spread of roughly 95 percentage points inside a single industry. One sector, opposite outcomes. The question the past year has answered is not whether software survives AI, but which software does.

The fault line runs through pricing, not technology

The boundary separating winners from losers is not whether a company is "AI-native." Some of the most battered names in the sector are large, product-rich platforms. The line tracks something more mundane: how the software is priced. Most software companies sell per user seat. An AI agent can now do much of the work that justified those seats — draft the document, triage the ticket, complete the workflow — and the seat count stops being the right unit the moment that is true.

Salesforce is the clearest in-the-open signal. The company reported record revenue, but the notable move was pricing: it is shifting customer-service software toward outcome-based pricing, an acknowledgment that AI agents will "perform work that used to justify those seats". When the market's most important seat-seller starts pricing the outcome instead of the seat, it is conceding that the seat is no longer the durable measuring stick.

What makes software worth lending to

Ares' view of which software holds up maps directly onto that fault line. Jacobson says he focuses on businesses where the "cost of failure is very high" — systems the customer's whole operation depends on — because in those places customers do not trade trusted incumbents for a cheaper newcomer. Ares' own framework draws the same distinction: resilient software is mission-critical, deeply integrated into workflows, and sold into regulated end-markets like healthcare and financial services, where the penalty for a wrong answer keeps buyers slow to switch. Vulnerable software, by contrast, is narrow — single-function tools or content creation, exactly the category an agent can reproduce cheaply.

This is a lender's discipline speaking, not a growth thesis. Ares underwrites these businesses as senior secured loans at the top of the capital structure, with average loan-to-value below 40 percent — an equity cushion that buys time for uncertainty to resolve. Credit investors can afford to divide software into survivors and casualties because they are protected by the capital stack; equity holders are not, which is why the same divide looks so much sharper in stock returns.

Where equity and credit now agree

The equity market and the credit market have converged on the same line. CrowdStrike just posted its largest quarter of net new annual recurring revenue ever — $331 million, up 47 percent year over year, lifting it past $5.25 billion in ARR. "So much for AI disrupting cybersecurity" reads the headline, and the stock is up roughly 77 percent this year. CrowdStrike sells the mission-critical layer — security monitoring a company cannot operate without — and AI has strengthened rather than weakened it. Salesforce, a fine business that must manage the seat-to-outcome transition, is down about 8 percent this year despite its record revenue. The market is paying a premium for durability and discounting the pricing rework, which is another way of saying both markets now read the seat.

None of this proves the cycle has settled. The shifts Jacobson describes are expected to unfold over years, not months, and the safest software — the entrenched, regulated kind — is precisely where adoption moves slowest for the same reason the moat is wide. But the direction of the argument has changed in a year. The open question is no longer whether AI eats software; it is where each company sits relative to the seat. If agents replace what a company sells — a replaceable, per-seat function — the value flows out of it. If agents reinforce what it sells — a mission-critical system with proprietary data and high switching costs — AI hardens the pricing power. That is the distinction a credit manager is paid to be right about, and it is the same one now visible in equity returns.