May you live in interesting times. That old line was meant as a curse — the wish that you’d live through chaos instead of calm. Welcome to software in 2026.
Over the last month, investors woke up to the potential of AI for traditional software businesses, and the reaction was brutal. More than $1T in market cap for legacy SaaS evaporated. The S&P 500 Software & Services Index — 140 companies — fell 20% year to date. Some are calling it Softwaremageddon. The message investors sent to legacy SaaS was blunt: we think your future cash flows are at significant risk.
So what actually changed? For the first three years of AI, most people’s experience was chatting with an ever-smarter person. In 2025, the models got “reasoning” and “deep research” and dug deeper for better answers. But 2026 is when models stopped just answering and started doing — completing full tasks and complex, multi-step jobs across tools and APIs. New frontier models that make intelligent decisions, plus open-source agent frameworks that people are using to replace entire functions, flipped the switch. The stock pickers went “Oh shit.” SaaS, suddenly, looked vulnerable.
I think the panic is overblown. Here’s why — and how I’m investing through it.
We’re barely out of the starting gate
2023 was the big bang moment for AI — the first time the masses got this technology at scale. Researchers who’ve studied previous big bangs (the PC, mobile, cloud) find that technology adoption runs through four phases that take 30 to 50 years to complete. We are three years into Phase One. There is a very long way to go.
Societal change lags technological change — by a lot
Scaling laws describe how fast capability accelerates. But adoption has no such law. It’s governed by human and institutional friction, and friction is stubborn. Remember:
- A quarter-century after Amazon’s IPO, only 20% of shopping happens online.
- Only 60% of corporate data lives in the cloud.
- We’ve talked about cord-cutting for years, yet 50% of households still have cable.
- Mobile payments are at a 15–20% share, and in 15% of transactions we still use cash.
- EVs have been “the rage” for years and are still only 10% of car sales.
I understand the knee-jerk that discounted SaaS multiples. But disrupting incumbents will take much, much longer than people think. The capability sits with early adopters and won’t be evenly distributed for a long time. There’s a lot of money to be made helping companies make the transition over the next couple of decades.
The pace of change is faster than ever — and accelerating
For decades the industry rode Moore’s Law: transistors doubling roughly every two years. For the last ten years or so, hardware outran software complexity, and most innovation went into shipping compute to the cloud (be honest — is the new iPhone really that much better?).
AI breaks that pattern. You now have massively compute-intensive applications that aren’t bound by Moore’s Law and are advancing far faster than it. Some estimate AI intelligence is doubling roughly every six months. We went from a pretty smart chatbot to software making intelligent decisions on complex multi-step tasks in three years. And it’s accelerating. Buckle your seatbelts.
SaaS vs. Agents — know the difference
This is the heart of the fear, so let’s be precise about it.
Traditional SaaS — Salesforce, Slack, Notion — gives you structured interfaces, dashboards, and predefined workflows. You log in, click buttons, fill out forms. It charges per seat, per month, and promises a modest 10–20% productivity bump per employee.
AI agents are different animals. They’re autonomous systems that act on your behalf — they reason, decide, adapt in real time, and execute multi-step tasks across multiple tools without you lifting a finger. They tend to be priced on outcomes (tasks completed), and they promise very high productivity gains, up to full role replacement. SaaS waits for instructions; agents anticipate needs and deliver outcomes.
The tricky twist: much of the context agents need to do the job lives inside SaaS systems of record — and across them (email, Slack, CRM). That’s not a footnote. It’s why SaaS is more durable than the panic suggests.
Value is moving to the intelligence layer
Software has always been departmentalized by function — CRM for sales, ERP for the warehouse, accounting for finance. Those silos gave us dashboards. But people don’t want dashboards. They want intelligence: natural-language answers about their business without having to know where the data lives.
So the real question is whether buyers will pay for AI bolted into each legacy system, or pay for a new intelligence layer that runs across the entire enterprise. Part of why investors knocked Salesforce from 30x revenue to 15x is that they doubt Salesforce will own that layer — it only has Salesforce data to run AI on. I tend to agree. Why pay a premium for AI inside each sub-app when a meta intelligence layer can answer “who did I talk to about Project X last week?” across email, texts, and Slack at once?
Possible vs. Probable vs. Actual
Here’s where the fear gets ahead of the facts. Agentic AI ratcheted up expectations on the Possible. But the Probable and the Actual still lag — badly.
We finally have data. Anthropic’s labor-market research mapped real model usage against theoretical capability, and the findings are sobering for the doomers:
- No systematic rise in unemployment for highly exposed workers since late 2022 (though there’s suggestive evidence that hiring of younger workers in exposed jobs has slowed).
- AI is nowhere near its theoretical ceiling — actual coverage is a fraction of what’s feasible.
- Even in the most exposed jobs, real usage is less than half of theoretical. Some supposedly high-exposure fields — architecture, life sciences — show almost no usage today.
Societal change lags technological change. Again. The places with the most adoption are exactly where the LLM companies are aiming directly (coding, with Claude Code and Codex). That tells you where not to invest early — and where the long runway is.
So how do you invest in SaaS through Softwaremageddon?
Interesting times spell opportunity for new entrants. A few rules I’m playing by:
Avoid the blast radius. Applications too close to the core models — the wrappers — are the most at risk as the models improve. Generalized agents open a second front of competition. Don’t invest where the LLM labs are pointed.
Sustainable value lives in the niches. It always has. The durable companies understand and integrate a specific industry’s workflow in ways general tools can’t or won’t — fragmented POS systems, industry-specific integrations and partnerships, the unglamorous plumbing. It’ll be a long time before the LLMs reach most of these verticals.
Bet on the gap between demand and supply. In 2025, over 95% of enterprise AI applications failed to reach production. Demand for AI that solves specific business problems still far outstrips supply. Freelancers are adopting autonomous agents with a vengeance, but very few enterprises will hand them their account credentials and keys.
Back fast iterators early. Early-stage companies that find the right customer base and iterate fast enough will crush it — precisely because they’re still early in product and GTM when the ground shifts.
One more thing worth holding onto. The best vertical software doesn’t replace people — it eats the low-value work so people can do more of what humans are great at. The receptionist’s job was never the point; serving the client was. Done right, AI enables more human-to-human connection, not less.
Softwaremageddon is real on the screens. But $1T of fear is not the same as $1T of obsolescence. We’re three years into a thirty-year movie. The opportunity isn’t in betting against software — it’s in building the next layer of it.
