The software you stopped paying for.

The most expensive software a company owns in 2026 is increasingly the software it stopped paying for. The build-versus-buy pendulum — which swung decisively toward “buy SaaS for everything” for fifteen years — is swinging back, and the market is pricing it in real time. Per-seat software stocks are down 30–45% on the year on fears that AI agents do the work seats used to. Internal teams are spinning up custom tools with AI coding assistants in days instead of buying a subscription. And the survey data backs the anecdotes: a third of enterprises have already replaced at least one SaaS tool with something they built. But the smartest version of this story is not “SaaS is dead.” It is a precise rule about what to build and what to keep buying — and the data cuts harder in both directions than the headlines admit.

The evidence the dam is breaking

Start with the cleanest number. Retool's 2026 Build-vs-Buy report — 817 respondents — found 35% of teams have already replaced at least one SaaS tool with a custom build, and 78% expect to build more internal software this year. The categories getting replaced first are exactly the ones you'd predict: workflow automation, internal admin, then the edges of CRM, BI, and support. The enabling technology is AI coding assistants collapsing build time. Anthropic's own published examples are concrete: Stripe deployed Claude Code to 1,370 engineers and did a 10,000-line code migration in four days that would have taken an estimated ten engineer-weeks; Rakuten cut feature delivery from 24 working days to 5. When a custom internal tool takes days instead of a quarter to stand up, the math on a $50,000-a-year SaaS contract for a workflow you could own changes.

The carnage in the public market

Investors are not waiting to find out. The seat-based SaaS complex has been repriced on AI-disruption fear, and the drawdowns are brutal.

The single most telling line in the data is the bottom one: Twilio, which prices by consumption rather than by seat, is up ~39% on the year while its seat-based peers fell by a third or more. That is the whole thesis in one comparison — the disruption is not really “SaaS is dying,” it is “the per-seat pricing model is dying.” When an AI agent does the work of a human seat, a vendor that bills per seat watches its own revenue model erode, while a vendor that bills per unit of work done gets more usage as agents run. Software's forward earnings multiple collapsing from 84x to below the S&P 500 for the first time ever is the market making exactly this distinction.

The counter-stat that keeps you honest

Here is where the “rip out all your SaaS” crowd runs into a wall. MIT's 2025 study of enterprise AI found that 95% of GenAI pilots delivered no measurable P&L impact — and, critically, that buying from specialized vendors or partnering succeeded about 67% of the time, while internal builds succeeded only about a third as often. The root cause was not model quality; it was the “learning gap” — integrating a tool into real workflows and culture. Klarna is the cautionary tale made flesh. Its CEO declared in 2024 that the company had “shut down Salesforce” and would kill Workday too; months later he admitted Klarna had mostly swapped vendors and built an internal data layer, not replaced SaaS with an LLM, and on customer service he reversed course entirely — “we went too far… the result was lower quality” — and began rehiring humans. Building is cheaper than ever and still fails more often than buying. Both things are true.

AI didn't make “build” always right. It made build cheap — which means the discipline is no longer cost, it's judgment about which problems are actually yours to own.

The seat-pricing scramble

The incumbents are not standing still; they are dismantling their own pricing models before the agents do it for them. Salesforce launched Agentforce at $2 per conversation, then introduced Flex Credits at $0.10 per action and pay-per-resolution tiers — three pricing models where there used to be one seat. HubSpot moved its Breeze agents to outcome pricing in 2026: $0.50 per resolved conversation, not per seat, not even per attempt. Gartner expects 40% of enterprise apps to embed task-specific agents this year and roughly 35% of point-product SaaS tools to be replaced or absorbed into agent ecosystems by 2030. The vendors that survive will be the ones whose pricing tracks work done rather than chairs filled — and whose moats (proprietary data, compliance, integration gravity) can't be rebuilt by an internal team in a weekend.

The operator rule

For a lean consumer holdco like ours, the synthesis is a simple decision rule, and it is the one we actually run. Build the thin custom layer — your differentiated workflow, your proprietary-data logic, your agent orchestration, the middleware stitching commodity tools to your process. This is where AI coding assistants now make in-house builds genuinely cheap, and where the work is unique enough that no vendor will ever serve it as well as you can serve yourself. Buy the commodity infrastructure — databases, identity, payments, comms, HRIS, anything compliance-heavy — because the maintenance and compliance tax on building those yourself is exactly the tax MIT measured, and buying succeeds twice as often. And when you buy, prefer consumption or outcome pricing, because it tracks AI-driven usage and you never pay for idle seats. It is the same logic as our operating standard: own what compounds, rent what doesn't.

Sources: Retool 2026 Build-vs-Buy report; Bain; SaaStr: valuation reset; MIT via Fortune; diginomica: Klarna reversal; Anthropic: build velocity; Gartner; TIKR: Twilio.