Rent, Build — or Wait

Klarna just terminated its Salesforce and Workday contracts.

Not because it found better CRM software. Not because its procurement team negotiated a competitor deal. Because its AI agents don’t need a user interface to get the work done — and paying per seat for software that a human has to log into is a model that no longer fits how the work actually gets done.

Marc Benioff’s public response was three words: “How is he doing this?” — while simultaneously pivoting Salesforce’s own pricing from per-seat licenses to $0.10 per agent interaction. He saw the threat clearly enough to start cannibalizing his own business model before someone else did.

This is the board-level software decision question of 2026. And most governance frameworks are one option short of an answer.

The Framework That’s Missing a Column

Every board uses some version of build vs. buy — or in the more precise formulation, build vs. rent. It’s a clean two-column decision: do we develop proprietary capability, or do we license someone else’s? The criteria are familiar: cost, time to value, strategic differentiation, vendor risk.

That framework was built for a stable market. In AI, the market is not stable.

The hyperscalers — Microsoft, Google, Anthropic, OpenAI, AWS — are closing capability gaps in weeks. A startup category that takes 18 months to mature can be absorbed into an enterprise platform contract in a single product release. We are watching it happen in agent observability right now: a dozen startups spent 2025 building what Microsoft shipped as Agent 365 in early 2026.

The framework needs a third column: wait.

Rent is the right call when the gap is urgent, the regulatory or competitive cost of delay is real, and the vendor has a defensible moat — proprietary data, deep domain workflow, a community the platforms won’t replicate.

Build is right when the capability is core to your competitive differentiation, you have proprietary data no vendor can replicate, and you can maintain it as a strategic asset.

Wait is right — and it is chronically underused — when the gap is real but not critical, the hyperscaler roadmap is publicly pointing at it, and adopting a point solution now means integration debt you’ll unwind in twelve months when the capability ships natively in your existing enterprise contracts.

Patience is a procurement strategy. In a market moving this fast, “wait” is not indecision. It’s capital preservation.

The Question Under the Framework

But there is a harder question sitting underneath rent/build/wait — one that didn’t exist two years ago.

Does this capability even need a SaaS wrapper?

When I sat with Bas Brekelmans and the Copilot Cowork team at Microsoft in March, I watched agents operate software through browser automation — clicking buttons, navigating workflows, completing multi-step tasks across applications — without a human in the loop and without requiring the target application to have an API. Microsoft has extended the same capability into Copilot Studio for any website or desktop application.

Anthropic’s Claude achieves 72.5% accuracy on real-world computer use tasks across Google Drive, Excel, and enterprise apps. OpenAI’s Operator, now folded into ChatGPT’s agent mode, navigates websites and fills forms autonomously.

The implication: a meaningful category of SaaS spend is not being evaluated against better SaaS alternatives. It is being evaluated against having no SaaS at all — replaced by an agent that operates the underlying workflow directly.

Then there is the Model Context Protocol.

Anthropic open-sourced MCP in November 2024. By April 2025, it had 8 million downloads. By late 2025, 97 million monthly SDK downloads, over 5,800 published servers, and formal adoption by OpenAI, Google, Microsoft, AWS, and Salesforce. The Linux Foundation took over governance in December 2025, cementing it as a neutral enterprise standard.

MCP solves what engineers call the M×N problem: before it, connecting M AI systems to N data sources required M×N custom integrations. MCP collapses that to M+N — one server per tool, one client per AI system. Salesforce’s MuleSoft arm is now converting any existing API into an MCP-accessible agent asset. GitHub, Slack, Notion, Jira, Postgres, Stripe — all have published MCP servers.

The practical consequence: workflows that previously required a purpose-built SaaS product with a UI, a per-seat license, an implementation project, and an ongoing admin overhead can now be assembled from MCP connections to underlying data sources. No UI required. No seats to count.

What the Shift Actually Looks Like

The market data confirms the structural change is underway, not merely predicted.

Seat-based pricing has dropped from 21% to 15% of SaaS companies in twelve months. Hybrid and usage-based models surged from 27% to 41% in the same period. Salesforce’s own Agentforce introduced “Flex Credits” — $0.10 per agent interaction — a direct outcome pricing model layered on top of its legacy seat licensing, acknowledging the threat to its own core business.

Bain’s conclusion in their 2025 enterprise AI report: “Stop charging for access and start charging for work done.” Gartner predicts 35% of point-product SaaS tools will be replaced by AI agents or absorbed within larger agent ecosystems by 2030.

What is being paid for is shifting. The new primitives are data access, API connectivity, and enrichment — not the interface built on top of them. The companies with proprietary data moats are in the strongest position; the companies whose value proposition was primarily the UI are the most exposed.

The Counterpoint Worth Acknowledging

This disruption is real and it is structural — but the timeline is longer than the headlines suggest. Deloitte puts full transition at five or more years. The 2025 “SaaSpocalypse” that many predicted did not arrive on schedule.

There are also new forms of lock-in emerging to replace the old ones. Salesforce is raising API prices. SAP is channeling enterprise AI through its proprietary Joule interface. Celonis has already filed litigation against SAP over data access. The shift from per-seat pricing to data-access pricing may simply replace one vendor constraint with another.

And reliability remains a genuine limit. Claude’s 72.5% accuracy on computer use tasks is impressive — but a 27% failure rate on routine workflows is not yet enterprise-grade for regulated or mission-critical processes. Human oversight is not optional in those contexts.

The board question is not “will SaaS die?” It is: “which of our current SaaS contracts would not survive a serious review of whether the work could be done by an agent against the underlying data?”

The Question to Ask Management

“For each significant SaaS contract we carry — is the value in the data, the workflow logic, or the user interface? And if it’s primarily the interface, what is our exposure if agents can replicate that workflow without one?”

This is the audit most software portfolios have not had. It surfaces the contracts most at risk — and the ones defensible enough to keep.

The Bottom Line

The rent-vs-build framework was built for a world where software required human operators and purpose-built integrations. That world is eroding.

The board’s job is not to stop buying software. It is to apply the right framework to the decision — one that includes “wait” as a legitimate option, and “does this need a UI at all?” as a preliminary question.

The companies getting this right are not moving faster or slower than everyone else. They are asking better questions before they sign.

---

Next issue: TBD

— Ian

AI CoE Lead, Altria | Board Chair, rvatech | iantyndall.com