Systems of Record survive. The agentic layer moves elsewhere.

TL;DR

Surviving isn't winning. Systems of record will live on as the source of truth - but source of truth is not source of value, and the value is moving up into the agentic layer built on top of them.

Work is horizontal; records are vertical. Real workflows cross CRM, ERP, email and spreadsheets, so an agent trapped inside one system can never own the outcome. The orchestration layer that does is neutral and cross-system by design.

Incumbents are in an economic cage. Agentic AI is collapsing the cost to build and maintain software, yet incumbents are valued on seat expansion and renewals. A genuinely good agent shrinks the exact numbers Wall Street prices them on - so they can't fully commit.

They get the scraps, not the strategy. Systems of record will capture the marginal, low-stakes automations confined within their own walls. The high-value, cross-system use cases - the interesting ones - will not belong to them.

Our Argument

The fashionable take is that AI agents will kill SaaS. They won't. Systems of record - your CRM, your ERP, your HR platform - will survive, because someone still has to own the source of truth, enforce policy and absorb liability. But surviving is not the same as winning. The interesting question isn't whether systems of record live or die; it's whether they capture the value of the new agentic layer being built on top of them. They won't capture the majority of it. Source of truth is not the same as source of value.

Start with the shape of the work. Real knowledge work is horizontal: a single workflow touches CRM, ERP, email, spreadsheets, chat and half a dozen other tools. A system of record is vertical by design - single-domain, single-silo. An agent that lives inside one of them is trapped inside one of them. As Satya Nadella has put it, agents will be "multi-repo": they read and write across backends and don't much care which one they are talking to. The logic moves up into a neutral, cross-system layer - structurally the opposite of what a system of record is.

That is also where the value sits. The record stores what happened; the agentic layer gets the outcome done across systems. Accountability and execution have split apart, and the money follows execution. Worse for the incumbents, the context that makes agents actually good does not live in the database. The exceptions, the overrides, the reason a deal nearly died - that lives in email, in chat, in people's heads. Whoever captures those decision traces builds the compounding advantage, and it is not the CRM.

The incumbents also face an economic cage. Agentic AI is collapsing the cost to build and maintain software - replacing a packaged tool gets cheaper every quarter - yet incumbents are measured and valued on renewal metrics: seat expansion, net revenue retention. Those KPIs are the trap. A genuinely good agent reduces seats and cannibalises the exact numbers Wall Street prices them on. This is the innovator's dilemma in its purest form: they are punished if their own AI works. So they cannot fully commit, and the renewal contract - once the moat - becomes the thing they defend instead of the thing customers want.

And it shows. The best-resourced incumbent went all-in on agents, and of 18,500 Agentforce deals closed in 2025, only about half were paid. Building reliable agents for complex, regulated workflows is not their core competency, and the scoreboard reflects it.

None of this means zero share. Incumbents will capture the marginal use cases - the narrow, low-stakes automations that begin and end inside their own four walls: a summary here, a field update there. But the strategic use cases - the cross-system, high-value workflows where the real money is - demand a neutral orchestration layer that owns the outcome and earns trust through reliability. That layer is being built now, and it is not the system of record.