Why product leaders who refuse to eat their own business model are choosing slow extinction over fast reinvention.
The Thesis
SaaS, as we built it for two decades, is over. Not declining – over. The per-seat pricing model that funded the entire industry is breaking in real time, and the companies that survive the next ten years will not be the ones with the best product or the deepest moat. They will be the ones willing to cannibalize their own business model before the market does it for them.
I call this the Radical Cannibal Test. Most product and executive teams will fail it. Not because they lack the talent, but because self-disruption requires killing the thing that currently pays your salary, and almost nobody does that voluntarily until the stock price forces their hand.
The Evidence: This Isn’t a Thesis, It’s Already Happening
In a six-week span between January and February 2026, the software sector lost approximately $2 trillion in market capitalization. This wasn’t a recession-driven correction or a rate-hike shock. It was a structural reassessment: investors stopped believing that per-seat SaaS pricing deserves a premium when AI agents can do the work that used to require five licensed humans.
Microsoft alone shed roughly $360 billion in a single trading session – the worst day for software since the COVID crash. If a vertically integrated hyperscaler isn’t immune, no point solution is.
The mechanism is simple and brutal: when one person equipped with AI agents can do the work of five, the revenue-per-seat math that powered SaaS economics for twenty years collapses. Gartner projects that 35% of point-product SaaS tools will be replaced by AI agents by 2030. That also means 65% survive – but only the ones that change shape.
This is no longer a projection. Workday announced 8.5% layoffs explicitly attributed to AI efficiency gains – fewer people doing the same work, which means fewer seats needed, which means less revenue for every SaaS vendor whose pricing model follows headcount. Publicis Sapient is actively cutting traditional SaaS licenses by roughly 50%, including enterprise standbys like Adobe, replacing them with generative AI tools the company describes as an order of magnitude faster than the junior staff those tools displaced. The SaaSpocalypse is not a forecast. It is a current event.
The Salesforce Case Study: Self-Disruption Is Necessary But Not Sufficient
Salesforce is the cleanest example of why “doing something” isn’t the same as “doing enough.”
Benioff didn’t wait for the SaaSpocalypse to hit before disrupting his own pricing. About a year before the 2026 selloff, Salesforce moved away from a pure per-seat model toward a hybrid: customers keep their licenses, but now also pay per completed Agentforce action – a new metric called Agentic Work Units. It’s the equivalent of switching from billing per call-center employee to billing per resolved ticket. That’s genuine cannibalization of the model that built a $30 billion business.
And the stock still fell – at one point down 43% year-to-date, with the longest losing streak in company history. Why? Because investors weren’t asking “did Salesforce adapt?” They were asking “did Salesforce adapt enough, fast enough, and credibly enough that customers won’t just build their own version with a coding agent?” Benioff’s counterargument – that nobody will “vibe-code” a regulation-compliant CRM from scratch because of the compliance and data-security layer built over decades – is a real moat. It’s also exactly the kind of argument every incumbent makes right before the moat gets bridged by someone with less to protect.
The lesson for product leaders: one act of self-disruption, however genuine, is not a strategy. It’s a single data point. The market is pricing in continuous, structural willingness to keep cannibalizing – not a one-time pricing pivot you can point to in a board deck.
The Product Exec’s Dilemma: Three Numbers That Should Be on Your Roadmap Slide
Strip away the market drama and there are three numbers every product leader should be running through their own P&L right now – because they describe the exact mechanism that will hit your roadmap, not just the stock ticker.
Number one: your gross margin is about to compress, structurally. Traditional SaaS ran on 75–85% gross margins because marginal cost per customer was near zero. AI changes that math permanently – inference, model hosting, and orchestration scale with usage. ICONIQ’s 2026 State of AI survey puts average gross margins for AI-native products at roughly 52%. If your roadmap still assumes SaaS-era margin structure for the AI features you’re shipping, your pricing model and your margin model are already misaligned – and finance will find that gap before your customers do.
Number two: the revenue metric you report on is the wrong one. Salesforce’s pivot is instructive precisely because it’s a metrics change, not just a pricing change. They introduced Agentic Work Units – completed actions, not logged-in seats – as the headline number, and Agentforce ARR grew 169% year-over-year to $800 million on that basis. As a CPO or VP Product, the strategic question isn’t “should we add AI features” – it’s “what is the unit of value we charge for once usage decouples from headcount?” If your board deck still leads with seats, MAUs, or logins as the primary growth metric, you’re reporting on a variable that’s about to become structurally disconnected from the value you deliver.
Number three: 65% is the real planning number, not 100% or 0%. Gartner’s projection that 35% of point-product SaaS tools get replaced by AI agents by 2030 is usually read as a doom number. Read it as a CPO or VP instead: it means two out of three product categories survive – but only the ones that have already moved off pure point-solution positioning. The strategic work is diagnosing, today, whether your product is sitting in the 35% (a workflow an agent can fully absorb – task tracking, data entry, status logging) or the 65% (something with a data moat, regulatory surface, or network effect an agent can’t trivially replicate). That diagnosis should be driving your next two roadmap cycles, not your next earnings call response.
The Executive Framing: This Is a Capital Allocation Decision, Not a Feature Decision
Here’s the strategic trap most product organizations fall into: they treat AI disruption as a roadmap problem – “which features do we ship” – when it’s actually a capital allocation and organizational design problem that happens to surface through the roadmap.
Three decisions that belong on a Product Exec’s desk, not buried in a sprint backlog:
Where do you put R&D dollars when most of the org still runs on the old revenue model? Avenir’s 2026 survey of enterprise buyers found 63% expect their existing vendors to benefit from generative AI, while only 8% expect them to lose out. That’s a real asset – customer goodwill toward incumbents making the transition – but it has a shelf life. Every quarter you spend protecting legacy ARR instead of funding the cannibalizing product line burns down that goodwill. How do you structure the org chart for a product that needs fewer humans to deliver the same outcome? This is the part most CPOs and VPs avoid saying out loud: if your own product makes your customers’ teams smaller, the honest internal question is whether your product org should be getting smaller and more concentrated too – fewer, more senior product/eng pairs working with AI leverage, rather than scaling headcount the old way. What’s your actual time horizon for margin recovery? IDC’s framing is useful here: the negative impact on SaaS applications is front-loaded in the next 1–2 years, with recovery expected by 2030 for vendors who’ve refactored their pricing around consumption, outcomes, or capability rather than seats. That’s a five-year capital allocation horizon, not a one-year product roadmap. Boards and CFOs need to hear that timeline explicitly, or they’ll keep evaluating AI investment against quarterly SaaS-era growth expectations it was never designed to meet.
This is the executive distinction that separates a CPO/VPfrom a head of product: a head of product asks “what should we build.” A CPO/VP asks “what is the cost of capital tied up in the model we’re refusing to cannibalize, and how long can we afford to carry it.” Salesforce can afford to carry that cost while it works through Agentforce, because it has $41.5 billion in revenue and three years of internal Agent Albert development behind it. Most product organizations don’t have that runway – which is exactly why the cannibalization decision needs to happen earlier and faster the smaller your balance sheet is.
My Own Take: Shrink the Suite, Build the Specialist
Here’s where I’d actually place my chips as a Product Exec today, not just diagnose the market.
I would pull a dedicated slice of my product team out of the regular roadmap and put them on a single R&D mandate: translate the product we sell today into the product an AI-native customer will actually use tomorrow. Not “add a copilot to the existing UI” – that’s the chatbot-overlay trap, and the SaaSpocalypse data is explicit that overlays don’t protect valuation, genuine architectural integration does. This has to be a parallel team with its own budget line, insulated from the pressure to ship this quarter’s feature requests for the legacy product, because that pressure will always win if the same team owns both mandates.
The deeper strategic question this R&D team needs to answer is the one that actually decides your TAM: does tomorrow’s user sit inside one integrated suite, or does the user increasingly interact through an agent that orchestrates the best specialized component from across the market? I think the honest answer, looking at where the agent-as-interface-layer thesis is heading, is the second one. If an agent is the orchestration layer between the user and a federation of specialized services, the competitive prize stops being “who owns the most modules in one suite” and starts being “who is the best-in-class component an agent reliably calls for a specific job.” That’s a fundamentally different competitive game than the one most SaaS roadmaps are still designed to win.
Which leads directly to the scope question: how far can I narrow my own product’s footprint and still maximize customer value? My answer is: much further than incumbents are comfortable with. The “eierlegende Wollmilchsau” – the all-in-one suite trying to cover every adjacent use case – is exactly the asset class carrying the most technical debt and becoming less maintainable every month, because every adjacent module was built to defend against losing a deal, not because it’s where the product is genuinely best. A narrower, specialized portfolio is faster to harden, faster to re-architect around an agent interface, and faster to react to the next shift in the market – three advantages the bloated suite structurally cannot match, no matter how much R&D budget you throw at it.
That’s exactly where strategic partnerships come back into play, and why I think they’re underused right now. If I deliberately cut scope to the two or three things my product does better than anyone, I need partners covering the adjacent jobs my agent-orchestrated customer will still need done. That’s not a weaker competitive position – it’s the condition for being the specialist an agent picks reliably, instead of the generalist an agent quietly routes around. The companies treating “shrink the suite, partner the rest” as a retreat are missing that it’s the more aggressive bet: it trades the illusion of broad coverage for genuine, defensible depth in the few things that actually justify a premium once an agent – not a human clicking through a dashboard – is doing the evaluating.
This is also where I’d push back on Benioff’s compliance-and-trust argument directly, rather than just noting it as a counterpoint. His logic assumes the moat that matters is decades of accumulated brand safety and regulatory infrastructure. I think that’s the wrong moat for a narrow AI product to compete on in the first place. Most SaaS contracts today aren’t exclusive or one-shot decisions – a customer can run a focused AI challenger in parallel, alongside the incumbent, without ever formally switching. That collapses the real entry barrier: the decision isn’t “replace or stay,” it’s “test for free, with almost no downside.” A site called canivibecodeit.com now catalogues over 976 SaaS applications – exactly the question the CEO conversation raised – and gives a direct verdict on each: YES (rebuildable in a single AI coding session), KINDA (a weekend), or NOT REALLY (a genuine moat). For every YES, it provides the exact prompt to paste into Claude Code, Cursor, or Codex. That’s the Challenger 4.0 dynamic made operational: not a theoretical threat, but a publicly available checklist of which subscriptions your customers are one prompt away from cancelling. Trust doesn’t need to be proven upfront through certificates and procurement reviews – it gets earned afterward, simply by the challenger working. Effort and risk are minimal; upside is enormous. That’s challenger-versus-incumbent dynamics, version 4.0: the same pattern that’s disrupted every mature market for a century, except this time the incumbent doesn’t have a decade to watch the challenger mature before responding. The cycle time has collapsed from years to quarters. (This logic is strongest for workflows that can be tested in isolation – a research agent, a reporting tool, a single process step. It’s weaker for products deeply embedded in data custody or compliance, like a system of record holding customer data, where the test itself carries some of the risk. That’s roughly the same line Gartner’s 35%/65% split is drawing – point-solution workflows fall first, data-moat systems hold longer.)
That parallel-running pattern has a dangerous second-order effect most incumbents aren’t watching for: it produces no clean signal until it’s too late. Customers don’t cancel when they start testing a challenger – they just quietly route more of their real work through it while the legacy contract auto-renews out of inertia. Usage erodes, logos stay flat, NPS looks fine, and the business believes it’s healthy because the metrics it’s used to watching haven’t moved. Then, once the challenger crosses a threshold of trust and functionality, the customer doesn’t downgrade gradually – they cancel outright, in one step, because there’s no longer a reason to pay for two solutions. Most companies are dead before they know it: either they’re silent-quitting their way toward a contract that won’t renew, or they’re one quarter away from a cancellation wave with no warning in the dashboards that currently define “healthy.”
The diagnostic rule that follows from this: if usage on your own product declines and nothing in the market or your customer’s business explains it – no seasonality, no known competitor win, no headcount change on their side – that gap is itself the trigger for a direct customer conversation, not a footnote to investigate later. And that conversation can’t be a one-off response to a bad quarter. It has to run continuously, as a standing practice, because the qualitative signal – a customer quietly trialing something else, a power user who’s lost enthusiasm – almost always arrives before the usage data confirms it. Data is lagging by definition. A live, ongoing customer dialogue is the only channel that gives you the lead time the challenger’s compressed cycle time has taken away everywhere else.
Which is precisely why the incumbent’s best move isn’t defending the trust moat – it’s building the challenger itself, internally, before someone with no legacy revenue to protect does it faster. That’s the cannibal logic applied one level deeper: it’s not enough to adapt your pricing, you have to be willing to compete with your own product using rules a challenger would use, on a timeline a challenger would use.
What “Radical Cannibal” Actually Means for Product Leaders
This is not a call to “embrace AI” in the vague, conference-keynote sense. It’s a specific, uncomfortable discipline:
1. Price against your own incumbency, before someone else does it for you. If your revenue model depends on seat count and your product makes humans more efficient, you are structurally incentivized to under-sell your own AI features. Build the consumption- or outcome-based model now, while you still control the timeline.
2. Treat your core workflow as the thing most likely to be replaced – not protected. The SaaS categories hit hardest in 2026 (project management, CRM, customer logging) were exactly the ones whose core workflow – task tracking, data entry – is what AI agents automate best. If your product’s value proposition is “we make a manual process less manual,” that’s the first thing an agent eats.
3. Stop optimizing the org chart you have and start designing the org chart AI makes possible. IDC predicts that by 2028, pure seat-based pricing will be obsolete industry-wide, with most vendors refactoring around consumption, outcomes, or organizational capability. Waiting for that prediction to become a press release is not a strategy.
4. Accept that “good enough” self-disruption will still get punished by the market – and do it anyway. Salesforce’s experience is the uncomfortable proof point. You cannibalize not because it guarantees survival, but because refusing to guarantees extinction.
The Counterargument – And Why It Doesn’t Change the Thesis
Not everyone agrees the apocalypse is real. Jensen Huang called the “software is dead” narrative the most illogical thing he’d heard. A meaningful school of thought argues the real 2026 risk isn’t AI capability – it’s “business model debt”: years of pricing commitments and billing architecture built for a seat-based world that simply hasn’t caught up to what the AI era demands. On that view, companies don’t need to destroy themselves; they need to pay down structural debt faster than the market loses patience.
I think both things are true at once, and that’s the point. Whether you frame it as “self-disruption” or “paying down business model debt,” the action required is identical: you have to be willing to dismantle the parts of your own model that no longer match how value gets created. The framing is a matter of taste. The radicalism required is not optional.
The Ten-Year Bet
In ten years, the SaaS companies still standing will not be recognizable as SaaS companies in the 2015 sense. They will price differently, organize differently, and very likely employ far fewer people per dollar of revenue than they do today. The ones that get there will be the ones whose leadership treated cannibalization not as a crisis response, but as a standing operating principle – something you do every planning cycle, not once when the stock crashes.
Only the most radical cannibals will survive. Everyone else is running a very well-designed SaaS business for a market that is already gone.
Now it’s time for you to reflect. Are you radical and cannibal enough to survive the SaaSpocalypse?
