Photo by Taylor Vick on Unsplash
What's on the Table
Two of the most powerful people in enterprise software looked at the exact same technology and drew opposite conclusions. Satya Nadella, speaking on a podcast in December 2024, argued that traditional SaaS business applications would collapse — that business logic would migrate out of individual apps into an agent orchestration layer, leaving the apps themselves as little more than CRUD databases (create, read, update, delete — the basic storage operations under any application). Marc Benioff, whose company had shipped Agentforce two months earlier in October 2024, pushed back: agents need trusted data, metadata, and workflow foundations, and those live inside established SaaS platforms.
Our read, as of August 26, 2026: both men are describing the same architectural shift and disagreeing only about where the margin lands — and the pricing data suggests the answer is neither purely one nor the other.
According to AI Fallback, whose reporting frames this debate, the disruption-versus-incumbency split has defined enterprise software strategy since late 2024. What follows is an attempt to do something the two camps' press cycles have not: put their numbers next to each other and compute what the difference actually costs a buyer.
The Pattern: Agents Are Tool-Callers, and SaaS Apps Just Became Tools
Strip away the vendor framing and the underlying pattern is unremarkable to anyone who has wired an LLM to an API. An agent is a loop: reason about a goal, select a tool, call it, read the result, decide whether it is done. The ReAct pattern, essentially. What changed between 2023 and 2026 is not the loop — it is that the tools became entire enterprise applications rather than a weather API.
That reframing is the whole argument. If Salesforce is a tool call, then the CRM's user interface — the part customers pay per seat for — stops being the product. The schema and the permission model become the product. Nadella's "CRUD database" line is not an insult so much as an architectural prediction: value migrates to whoever owns the orchestration.
The non-obvious part is that this migration does not require any SaaS vendor to lose a customer. It requires only that the number of humans clicking through the interface falls. Per-seat licensing is a tax on human attention. Agents do not have attention, and they do not buy seats.
Gartner's forecast puts a timeline on it: as of the firm's published projection, fewer than 1% of enterprise applications embedded agentic AI in 2024, and roughly 33% will by 2028, with about 15% of day-to-day work decisions made autonomously by AI agents by that same year. That is a jump from under 1% to 33% across four years.
Chart: Gartner's projection for agentic AI penetration in enterprise applications, 2024 versus 2028. Source: Gartner newsroom forecast.
Per-Seat vs. Per-Conversation: Who Wins Under Which Condition
Here is the comparison no single source article lays out, because each source only has half the numbers.
Salesforce priced Agentforce at launch around $2 per conversation — a consumption model, not a subscription. Klarna, separately, reported that its OpenAI-powered assistant handled roughly 2.3 million conversations in its first month, work equivalent to about 700 full-time customer-service agents, with a projected profit improvement of about $40 million.
Put those two figures in the same sentence and the arithmetic gets uncomfortable. At roughly $2 per conversation, 2.3 million conversations would run to approximately $4.6 million for a single month. Annualized at that volume, that is a line item in the neighborhood of $55 million per year against a reported $40 million profit improvement. The two numbers come from different companies with different cost structures, and Klarna built in-house rather than buying Agentforce — that is precisely the point. The build-versus-buy decision at high conversation volume is not close.
Run the same math the other direction and the incumbency case looks much stronger. A company handling 5,000 support conversations a month pays roughly $10,000 at $2 each — about $120,000 a year. Hiring and managing an equivalent internal AI platform team, plus evals, plus on-call, plus the model bill, does not come in under that. For mid-market volumes, consumption pricing is a bargain.
So the honest breakdown: consumption-priced agent platforms win when conversation volume is low-to-moderate and engineering headcount is scarce. In-house agent stacks win when volume is high enough that per-unit pricing compounds faster than salaries. Klarna sits firmly in the second bucket. Most companies do not.
And the per-seat model? It loses in both scenarios — not because it is expensive, but because it prices the wrong unit. This is the same cost-structure logic that drives the model-tier decisions AI Trends examined in enterprises picking cheaper Claude tiers: once the workload is machine-generated rather than human-generated, unit economics stop resembling a software subscription and start resembling a utility bill.
A careful skeptic pushes back here, and the pushback is fair: launch pricing is marketing, and vendors routinely discount volume. True. But the direction of travel is what matters — Microsoft's Copilot Studio agents, Google's Vertex AI Agents, ServiceNow, Workday, and OpenAI all shipped agent-building frameworks across 2024 and 2025. Every one of those frameworks meters something. None of them meters seats.
Where This Breaks in Production
The demos hide the retry logic. They always do.
The failure modes in an agent-layer architecture are not the ones the strategy decks discuss. Three worth naming:
Tool-call loops against billed APIs. When an agent's tool call returns an ambiguous error — a 429, a partial write, a timeout that may or may not have committed — the reasoning loop retries. If every retry is a metered conversation or a billed API call, a single malformed prompt template can generate a five-figure invoice overnight. Per-seat pricing had a natural ceiling; consumption pricing does not. Budget alerts are not optional infrastructure here, they are the first thing to wire.
Context window blowups on real enterprise schemas. Nadella's "agents orchestrate across data sources" framing assumes the agent can hold enough schema in context to reason about it. A production CRM with hundreds of custom fields, a Workday instance with a decade of org history, and a data warehouse joined across both will exceed any practical context budget. This is where Benioff's argument has real teeth: the metadata layer that makes an enterprise dataset legible to an agent is genuinely hard, and incumbents already built it. Retrieval helps, but retrieval over an unmodeled schema returns plausible-looking wrong joins.
The quality floor nobody benchmarks. Klarna's story is the cautionary one, and the divergence in how it has been reported is instructive. Initial coverage framed the AI assistant as replacing roughly 700 agents and cutting SaaS spend at named vendors including Salesforce and Workday. Later reporting noted Klarna rehiring human agents for quality reasons. Both are true. The lesson is not that automation failed — 2.3 million conversations is not a failure — but that the last few percent of cases are where customer trust is decided, and that tail is expensive to automate and cheap to staff.
Eval-driven development is the only defensible answer to all three. Not vibes, not a demo that worked on Tuesday: a fixed test set of real tickets, scored on every prompt change, with cost-per-resolved-case tracked alongside accuracy. Teams that skip this find out about regressions from customers.
Bottom Line
Our analysis: the "SaaS is dead" framing is wrong, but so is "nothing changes." What is dying is a pricing model, not a category. The per-seat subscription assumed a human at a keyboard, and Gartner's projection that about 15% of daily work decisions will be autonomous by 2028 is, read correctly, a forecast that roughly 15% of seat-hours stop existing. Vendors that convert to consumption or outcome pricing before that happens will look prescient. Vendors that defend seat counts will report flat revenue on growing usage and struggle to explain it.
On balance, the more likely outcome is a split: incumbents keep the system-of-record business because schema and permissions are genuinely defensible, while the orchestration layer — and the margin attached to it — becomes contested ground among Microsoft, Google, OpenAI, and whoever else can make cross-system tool calls reliable. Benioff is right about the data. Nadella is right about the logic. They are describing different halves of the same stack.
For anyone building rather than watching: pick one workflow with a measurable outcome, wire an agent against it with hard spend caps and a real eval set, and instrument cost-per-resolved-case from day one. That single number will tell you more about whether the agent layer is worth it for your organization than any vendor forecast.
Frequently Asked Questions
Will AI agents replace SaaS entirely by 2028?
The available forecasts do not support total replacement. Gartner projects roughly 33% of enterprise applications will include agentic AI by 2028, up from under 1% in 2024 — that describes SaaS absorbing agents, not disappearing. The more supported claim is that per-seat pricing gets displaced, since Salesforce's Agentforce launched with consumption pricing around $2 per conversation rather than per-user fees.
What is the actual difference between an AI agent and a SaaS application?
A SaaS application is a hosted system of record with a user interface — a human logs in and performs actions. An AI agent is a reasoning loop that selects and calls tools autonomously to complete a goal, potentially across several SaaS systems at once. In Nadella's framing, the agent holds the business logic while the SaaS app holds the data. In practice most 2026 deployments run agents inside or alongside existing SaaS platforms rather than replacing them.
Is it cheaper to build AI agents in-house or buy an agent platform?
It depends almost entirely on volume. At roughly $2 per conversation, a company handling 5,000 monthly conversations pays about $120,000 annually — cheaper than staffing an internal AI platform team. At Klarna's reported scale of about 2.3 million conversations in a single month, the same per-unit rate would imply roughly $4.6 million monthly, which is why Klarna built internally and reported approximately $40 million in projected profit improvement.
What is agentic AI in enterprise software, in plain terms?
Agentic AI means software that decides what to do next rather than waiting for instructions. Instead of a chatbot answering a question, an agent reads a support ticket, queries the order system, issues a refund, updates the CRM record, and reports back — each of those being a tool call the model chose. Microsoft's Copilot Studio, Google's Vertex AI Agents, ServiceNow, Workday, and OpenAI all shipped frameworks for building these between 2024 and 2025.
Disclaimer: This article is editorial commentary for informational purposes only and does not constitute financial, investment, or procurement advice. It reflects analysis of publicly reported information and does not involve independent product testing or benchmarking by this publication. Vendor pricing and product capabilities change frequently; verify current terms directly with vendors before making purchasing decisions. Research based on publicly available sources current as of August 26, 2026.