Two opposite stories are being sold to us. First: "SaaS is dead, agents will sweep it all away." Second: "nothing changes, we just bolt an assistant onto the corner of the screen."
Both miss the point. If you want to know what the software you'll buy in three years looks like, don't watch the demos - look at the invoice. That's where the shift is already visible, including at the incumbent vendors.
Pricing Gives Away the Architecture
Classic SaaS is priced per seat: one subscription per person who logs in. That model assumes something very specific - that the value of software is letting a human do the work faster.
Picture an office building rented by the number of occupied desks. The rent follows the chairs. The day the work is done by machines that sit on nothing, the meter stops measuring anything at all.
That's exactly the problem vendors are running into. a16z puts it coldly: when a company pays ~$115 per month per support agent, and an AI starts handling the tickets, the vendor that stays on seats is engineering its own revenue decline. So the metric moves to the outcome. Zendesk prices its AI agents around $1.50 per automated resolution rather than per human seat. And per PYMNTS, ServiceNow, SAP and Workday are shipping agents billed on consumption or outcome: "we're moving from counting people to counting activity."
Salesforce is feeling its way in public - and that's instructive. The vendor launched a support agent billed per resolution, and according to The Register it's hunting for pricing aligned with the benefits actually delivered (customer service agents paid per case resolved)… but it also re-packaged its agents into per-seat licenses - AELA, launched in October, with reusable credits - because customers moving to production want budget predictability. Three models coexist. This isn't indecision: it's the sound of an industry whose unit of account is changing.
The Front Door Changes Shape
Second signal, more structural than price: who the software is built for.
Classic SaaS is a building with a reception desk, counters and forms - everything designed for a human to click. The software that's coming adds something else: a loading dock, designed so an agent can pick up and drop off work directly.
The name of that dock, to be precise, is MCP (Model Context Protocol): an open convention through which an application declares, in a standard format, the data and actions it exposes - "here are my tools, here are their parameters, here's what they return" - so an agent can call them without a hand-stitched integration per piece of software. It isn't an AI model, it's a power socket. And it's spreading: MCP server downloads went from roughly 100,000 in November 2024 to over 8 million in April 2025, and per a Stacklok report cited in 2026, 41% of surveyed software organizations have some in production, even if limited.
What that means for the buyer: the question "is the interface pleasant?" now comes with a new one - "can my agents get in?" Software that only exposes screens becomes a dead end for your company brain.
And Sometimes, You Just Don't Buy
Third signal, the most uncomfortable one for vendors. A McKinsey survey relayed by Forbes finds that nearly a third of organizations walked away from buying at least one software product or feature because they can now build it in-house with agentic coding tools. The same work reports that 40% of enterprise applications will embed task-specific agents by the end of 2026, versus under 5% in 2025 (Gartner forecast).
In other words: the scope of "buy off the shelf" is shrinking from the bottom. What used to be a €40,000-a-year module becomes, sometimes, two weeks of assisted work.
What Doesn't Change (and the Mirage to Avoid)
Careful: "the next software won't look like Salesforce" doesn't mean the game is over. The numbers cool things down. Again per McKinsey, 88% of organizations use AI somewhere, but only 23% are scaling an agentic system, and fewer than 10% of functions have actually deployed agents. Gartner forecasts that over 40% of agentic AI projects will be scrapped by the end of 2027 - runaway costs, fuzzy business value, inadequate risk controls.
And straight replacement has a price. Klarna has become the canonical case: per press summaries, its agent absorbed 2.3 million conversations in one month - the equivalent of several hundred people - before customer satisfaction slipped on complex cases and the company reopened hiring. Volume, yes. Complexity, no.
Marc Benioff, who rejects the SaaS-apocalypse narrative, sums up the real shift better than his critics: "this isn't the end of software, it's the end of software that makes humans do all the work" - human being, agent doing. The human stays in command; they stop being the executor.
What We Do
At Elephant in the Room, we don't bet on a vendor. We build the layer that keeps those choices reversible: a company brain that holds your context, business agents plugged into it with tight permissions, standard connections to your existing software, and multi-model orchestration - because the model is a commodity and context is not. That's the concrete definition of an AI-native company: the software does the heavy lifting, you keep your hands on what matters.
Tomorrow's software won't be sold to those who click, but to those who get results.
