In August 2025, after Microsoft's Charles Lamanna predicted, as reported by The New Stack, that AI business agents would make traditional SaaS applications obsolete by 2030, we published Microsoft AI Agents Will Replace SaaS A Critical Reality Check for Business Leaders and called the timeline optimistic. Our conclusion was that SaaS would die slowly, and that its replacement "will likely follow evolutionary rather than revolutionary patterns, extending well beyond 2030 for most enterprise functions."
More than a year on, we were partly right, and the half we got wrong matters more for this year's budgets. Nothing in the pricing summaries we found suggests companies are being asked to rebuild their ledgers around probabilistic models. We were wrong to assume disruption would arrive through architecture. It arrived through the price list, and seat-based SaaS is already less stable than our talk of evolution led readers to expect.
The ledger case still holds
The strongest part of the original piece was Rocky Lhotka's point about determinism. Accounting and inventory run on discrete rules, and a model that answers the same question two ways is a liability there. The pricing we can see is sold per user or per action rather than as a replacement for core systems, which suggests the system of record is not the target. Our warnings about audit trails in regulated industries still read well.
Our evidence has limits. What we have is vendor pricing, largely as summarised in third-party breakdowns rather than checked line by line against vendor price sheets, and not adoption surveys or earnings disclosures on usage, so we can say little about how many companies are running agents in production. We can say a good deal about how vendors intend to get paid, which is where our original analysis was thinnest.
Disruption arrived as an invoice
In the original we cited Mary Jo Foley of Directions on Microsoft, who expected Microsoft to roll agents out as its next wave of paid add-ons to existing applications. We treated that as the reassuring scenario: incremental, manageable, more revenue for the vendor and little upheaval for the customer. That reading was too comfortable. A paid add-on is exactly how disruption gets into a contract, because once an agent is priced on its own, the vendor has to decide what unit it is charging for.
Salesforce's answer, as Concret.io describes it, shifted from the conversation to the action. A charge per action is a charge per unit of work completed, which is the thing a seat licence was always a rough proxy for. Microsoft, at least for Copilot, kept the seat. Taken together, Concret.io's and Oliv.ai's breakdowns point to Salesforce running per-user editions alongside metered actions.
That mix is the near-term disruption we missed. A company renewing in 2026 can face two meters on one relationship: a per-user fee for the people and a consumption fee for the agents. Licence forecasting built around headcount has no model for the second meter, which behaves oddly: the better an agent works, the more actions it takes, and the larger the bill. Finance teams that budgeted for seats will meet a variable cost line that grows with success, and none of it requires replacing a single ledger. SaaS did not have to be replaced to be disrupted, only to start charging for work instead of people.
SaaS did not have to be replaced to be disrupted, only to start charging for work instead of people.
The best case that we were right all along
The strongest defence of our original caution goes like this. Systems of record are untouched. Vendor announcements have outrun measured enterprise adoption, and a $0.10 price per action is meaningless if few actions are bought. Microsoft holding to per-user pricing for Copilot suggests the seat model is more durable than this revisit admits.
On adoption, the point is fair, and we will not pretend to usage data we do not have. But pricing architecture is a forward commitment. Vendors do not redesign their meters for products they expect to stay marginal, and Salesforce changing its billing unit from conversations to actions shows the commercial terms are unsettled whether or not adoption is broad. Buyers inherit that instability at renewal either way. As for Microsoft, its seat pricing fits our revised view. Where an agent assists a named person, charging per person still makes sense. Where an agent does work a person used to do, there is no person to charge for, and metering takes over. That boundary is where the seat-based model breaks first.
The advice left out the department that pays
Our 2025 recommendations were sensible: experiment in low-risk functions, review the application portfolio, build change-management capabilities for human-AI collaboration (training, role redefinition and performance measurement), talk to compliance, monitor standards. They spoke to technology, risk and people teams, and none of them was addressed to finance or procurement. Yet the first hard decision agents force is commercial: what unit of work to buy, how to cap it, and how to forecast a cost that rises with usage. That decision belongs with finance and procurement, and it needs them in the room before a pilot scales, because terms agreed at the first renewal with metered agents will set the cost base for years.
Two claims from the original we simply cannot score yet. Lhotka's warning that agent-run businesses would ossify, and our own view that organisational change takes generations, were presented as cautions when they should have been framed as predictions someone could check, a standard we will hold ourselves to from here.
Our 2025 piece asked whether agents could replace enterprise software and answered, correctly, not soon. It should also have asked what vendors would charge for once agents arrived. The answer is already on the price list: work, by the action. The ledger may well outlive 2030. The seat as the organising unit of SaaS economics is unlikely to, and the companies still budgeting as though it will are the ones who will learn about this shift from an invoice.
