The claim has been kicking around since Satya Nadella said it in early 2025: AI agents will replace SaaS. Every analyst deck since has run with it. The problem is it's imprecise in a way that misleads decisions.
Agents are not going to delete Salesforce or ServiceNow from your stack. What they are doing - right now, in contracts being signed this quarter - is making the seat irrelevant as a unit of value. That is a different kind of disruption, and it has sharper near-term consequences.
The seat was never a natural unit of value. It was a proxy for the human who did the work.
Seat-based SaaS had clean logic for twenty years: a human uses the tool, the tool's value scales with how many humans use it, so you price the seat. When an agent handles the same task - no login required, no human triggering the workflow - that logic breaks. When an agent does the work, the seat becomes fiction.
The vendors know this. Watch what they're actually doing, not what they're saying.
Salesforce shifted its Agentforce pricing away from the per-seat model toward an outcome-based structure.
Charging $2 per conversation signals the end of the inefficiency where businesses purchased 500 seats with only 100 in active use. Meanwhile, Intercom's Fin AI agent went from $1M to $100M+ ARR on a single pricing move: $0.99 per resolved ticket - not per seat, not per month, but per outcome.
Then came the retreat. Across the market, a switch back to seat-based pricing is a sign that, in the current phase, users are not buying AI agents to replace people - they want some certainty in the pricing model before investing.
When a $300B company like Salesforce abandons seat-based pricing for AI and then partially walks it back, it signals that the SaaS pricing playbook is dead for agents - even if the transition is messier than anyone predicted.
Microsoft is collecting seat revenue from customers who don't know the transition is happening, and consumption revenue from customers who do. Microsoft 365 Copilot still runs $18-$42.50 per user per month, while Copilot Studio - where you build your own agents - runs on consumption credits. That is not a company that has figured out the right model. That is a company running two bets simultaneously and waiting to see which one survives contact with enterprise procurement.
The steelman case for SaaS staying intact
There is a real argument that systems of record don't get disrupted by agents; they get a new layer on top. When you look at how enterprises are actually deploying AI agents in 2025 and 2026, they're not replacing their systems of record - they're building orchestration layers on top of them. The data lives in Salesforce. The compliance trail lives in ServiceNow. A "reality check" is occurring among CIOs who realize LLMs lack the deterministic consistency required for critical processes - a system that provides a correct answer six out of ten times is insufficient where 100% consistency is required. So the argument goes: agents handle interpretation and routing; the SaaS system does the actual execution. Coexistence, not replacement.
That is probably right for the top tier of platforms - the ones with deep data gravity. It is much less right for point solutions that exist purely to execute a narrow, repeatable workflow. Those tools are genuinely at risk. Not because an agent replaces the concept, but because the workflow gets absorbed into a larger agent pipeline and the vendor loses the budget line.
What the inference bill changes
There is a second-order effect that most of the "agents kill SaaS" analysis ignores: inference costs reshape the ROI of every feature decision. In traditional SaaS, flat-rate pricing works because the marginal cost of one more user interaction is effectively zero - there is no free marginal unit in AI. Every interaction is a real inference call with a real cost. The business model assumption that carried over from SaaS doesn't transfer.
In 2026, inference accounts for 85% of the enterprise AI budget. One reason: autonomous agents often "reason" in loops, hitting an LLM ten or twenty times to solve one task
- versus the single API call that a traditional SaaS integration would make. The pilot economics, calculated on single-query API calls, bore no relationship to the production economics of multi-step agentic loops running thousands of times per day.
This is where the "agents replace SaaS" story quietly falls apart on a spreadsheet. If your agent costs $4 in inference to save a support rep 15 minutes, the ROI is negative. The software didn't get disrupted. The math did.
What this means for teams deciding now
If you are evaluating whether to consolidate a workflow into an agent or keep the SaaS subscription, the useful question is not "can an agent do this?" It almost certainly can. The questions are:
- What does the inference loop actually cost per resolved outcome at your volume?
- Is there a system of record underneath this that creates compliance or audit obligations?
- Is the vendor's pricing model designed to share gains with you, or to capture them?
Outcome-based pricing requires a great deal of information sharing to track gains and attribute them to the agentic offering - transparency and trust are required on both sides. Most contracts being signed today do not have that infrastructure in place.
An AI teammate like Beagle - sitting inside Slack or Teams, fielding questions that would otherwise hit a tool's UI - is already surfacing this tradeoff for knowledge work: the seat you're paying for is often idle, and the inference call that replaces it only pencils out if you size the context window correctly.
The SaaS industry is not dying. It is splitting. IDC predicts that by 2028, pure seat-based pricing will be obsolete, with 70% of software vendors refactoring their pricing strategies around consumption, outcomes, or organizational capability. The software survives. The seat - the twenty-year proxy for human labor - probably doesn't. That is a narrower claim than "agents kill SaaS," but it is the one that will actually show up on your next renewal invoice.