Somewhere between Q4 2025 and Q1 2026, a quiet panic set in across the SaaS sector. Not the dramatic crash variety - more like the slow realization at a poker table that the rules of the game just changed. Software stocks dropped by double digits across the board, and the usual "buy the dip" chorus was conspicuously absent. The market was not punishing bad earnings. It was repricing an entire category.
The question everyone in enterprise software is asking right now: is this a correction, or is it the beginning of a structural shift? Having spent the last year building AI agents that replace chunks of SaaS workflows for clients, I have a clear (and possibly uncomfortable) answer: it is both. But the specifics of what gets repriced and what gains value are far more interesting than the headline.
The Bifurcation Is Already Here
The narrative that "AI will kill SaaS" is as lazy as the 2023 narrative that "ChatGPT will kill Google." What is actually happening is a bifurcation. Two categories of software companies are emerging, and the market is pricing them very differently.
Category 1: The Compression Layer. These are tools whose primary value was always about connecting systems, providing dashboards over APIs, or giving humans a place to do repetitive cognitive work. Think about the typical mid-market tech stack: a tool to pull data from Salesforce into a spreadsheet, another to route support tickets, another to format reports. Each one charges per seat, per month, forever.
AI agents can now do most of what these tools do. Not theoretically - right now. An agent with API access to your CRM, helpdesk, and database can route tickets, draft responses, update records, and generate reports without a $45/seat/month intermediary. Forrester's "SaaS-pocalypse" analysis nails this: the value of the integration layer itself is approaching zero.
Category 2: The Compounding Layer. These are tools where the software is not the product - the accumulated data, workflow intelligence, and domain expertise embedded in the system is the product. Epic's EHR system is not threatened by AI agents. Neither is Palantir's defense platform or Veeva's clinical trial management. The switching cost is not "find a new UI" - it is "rebuild ten years of institutional knowledge."
The market is punishing Category 1 and rewarding Category 2. This is not a blunt instrument. It is surprisingly precise.
What Actually Gets Commoditized
Let me be specific about the types of SaaS value that AI automation is compressing. This is not speculation - these are patterns we see in client engagements every month.
Data shuttle tools. Any product whose primary function is moving data from System A to System B. Integration platforms that charge per-task pricing are particularly exposed. An AI agent with access to both APIs can do the same work, and it gets smarter about edge cases over time rather than requiring you to build more Zaps.
Reporting dashboards. If the tool's core loop is "query data, format output, send to stakeholders," that is one prompt away from replacement. We built a client an agent that replaced three separate reporting tools - total monthly savings of $2,800 across 40 seats - in about two days of work.
Simple workflow automation. The irony is that first-generation automation tools (the ones that helped you avoid writing code for simple if/then logic) are now being automated themselves. BetterCloud's 2026 SaaS industry analysis found that enterprises are actively consolidating tooling, with AI-native orchestration replacing legacy automation platforms.
Template-driven content tools. Email sequence builders, social media schedulers, basic copywriting assistants. These were always thin wrappers over relatively simple logic, and the LLM price collapse of 2025-2026 made their margins untenable. Google's VP of engineering publicly warned that startups built as "thin wrappers" or "simple aggregators" face existential risk.
Here is a quick framework for identifying compression candidates in your own stack:
| Signal | Risk Level | Example |
|---|---|---|
| Per-seat pricing for repetitive tasks | High | Most CRM add-ons, basic helpdesk tools |
| Value comes from connecting two APIs | High | Integration platforms, data sync tools |
| Core logic could be a single prompt chain | High | Template builders, report generators |
| Stores proprietary data that improves over time | Low | Analytics platforms with custom models |
| Requires domain-specific compliance | Low | Healthcare, financial services tools |
| Deep workflow customization per customer | Low | Vertical SaaS with process IP |
What Compounds Instead of Compressing
The flip side of this story is more interesting. Some software categories are actually gaining value as AI proliferates, because AI makes their core asset (data, workflow depth, distribution) more valuable rather than less.
Vertical SaaS with proprietary data loops. Intellectia's 2026 analysis highlights that vertical SaaS companies embedding AI into domain-specific workflows are seeing valuation premiums. The key word is "embedding" - not "bolting on." Toast does not just add a chatbot to restaurant management. It uses years of transaction data, labor patterns, and menu analytics to power AI features that a generic agent literally cannot replicate without that dataset.
Platforms that become the orchestration layer. This is where a16z's "AI will eat application software" thesis gets interesting. The companies that position themselves as the substrate on which AI agents operate - the ones with the APIs, the data models, the execution environments - gain value as agent adoption grows. Salesforce is not threatened by AI agents. Salesforce is where AI agents go to get work done. Same for Workday, ServiceNow, and the major cloud platforms.
Compliance and trust infrastructure. Futurum Group's analysis makes a point that gets overlooked in the hype: regulated industries cannot just swap in an AI agent and call it a day. The compliance overhead - audit trails, data residency, access controls, explainability requirements - is itself a moat. Companies like Vanta, Drata, and OneTrust are more valuable in an AI-heavy world, not less, because every new agent deployment creates new compliance surface area.
Developer infrastructure. The Software Equity Group's analysis of M&A trends shows that developer tooling companies (observability, testing, deployment) are commanding premium multiples. More AI agents means more software to monitor, test, and deploy. The picks-and-shovels play is real here.
The LLM Wrapper Problem (And Why It Is Worse Than You Think)
We need to talk about a specific failure mode that is driving a chunk of the selloff: the LLM wrapper.
An LLM wrapper takes a foundation model, adds a UI, maybe a system prompt and some basic RAG, and charges a subscription. In 2023, this was a viable business. In 2026, it is a death sentence. Here is why the problem is structural, not cyclical.
Foundation model capabilities keep expanding. Every time Anthropic, OpenAI, or Google ships a model update, they absorb features that wrapper companies were charging for. Summarization got better. Code generation got better. Document analysis got better. The wrapper company's moat shrinks with every model release, and they have no control over the release schedule.
Price compression is relentless. Claude API costs dropped roughly 90% between early 2024 and early 2026. GPT-4-class intelligence went from $60/million tokens to under $3. When your cost basis drops that fast, your customers start asking why they are paying $200/month for what is essentially a prompt and a text box. Float's analysis of the SaaS-AI intersection calls this the "margin squeeze" - wrapper companies face pricing pressure from both above (smarter models) and below (cheaper inference).
Switching costs are near zero. If your product is fundamentally "our prompt + their model," a customer can switch to a competitor (or build their own) in an afternoon. Compare this to switching your ERP system, which takes 18 months and a team of consultants.
The distinction that matters is between an LLM wrapper and a vertical AI product. A vertical AI product uses the same foundation models but adds:
- Proprietary training data from the specific domain
- Workflow logic that encodes years of process expertise
- Integrations that are genuinely hard to replicate (not just API calls, but data transformations, edge case handling, compliance checks)
- Feedback loops where customer usage makes the product better
Harvey (legal AI) is a vertical AI product. Jasper (marketing copy) was, for a long time, closer to a wrapper. The market is now distinguishing between these with brutal precision.
How to Plan Your AI Automation Budget for the Next 12-24 Months
If you are an enterprise leader reading this, here is the practical framework. Skip the theory - here is what to actually do.
Step 1: Audit your SaaS stack with the 80% test.
For every tool in your stack, ask: could an AI agent with API access to our core systems handle 80% of what this tool does? If yes, it is a replacement candidate. Not today necessarily, but within your planning horizon.
Be honest about this. The answer is "yes" for more tools than most teams want to admit.
Step 2: Categorize into three buckets.
| Bucket | Action | Timeline |
|---|---|---|
| Replace - AI agent can do 80%+ of this tool's function | Build or buy an agent-based alternative | 3-6 months |
| Augment - Tool has deep value but AI can automate repetitive parts | Keep the tool, add AI agents on top | 6-12 months |
| Keep - Proprietary data, compliance, or workflow depth makes replacement impractical | Negotiate better pricing (your vendor knows they are under pressure) | Ongoing |
Step 3: Prioritize replacements by cost-per-seat times seat-count.
This sounds obvious, but most teams start with "what is cool to automate" instead of "what saves us the most money." A $15/seat tool used by 500 people ($90K/year) is a better target than a $200/seat tool used by 5 people ($12K/year), even if the expensive tool is more "interesting" to replace.
Step 4: Build for proprietary processes, buy for commodity ones.
If the process is unique to your company (your specific approval workflows, your data formats, your compliance requirements), build a custom agent. If it is a common business function (expense reporting, meeting scheduling, basic customer support), buy an AI-native tool from a vendor who specializes in it. The worst decision is building custom agents for commodity processes - you end up maintaining infrastructure that should be someone else's problem.
Step 5: Budget for the transition cost, not just the end state.
The hidden cost of SaaS-to-AI transitions is not the AI itself. It is the data migration, the workflow redesign, the change management, and the three months where you are running both systems in parallel. Budget 2-3x the "just the AI costs" number, and you will be closer to reality.
The Pricing Model Shift Nobody Is Ready For
One more structural change that deserves attention: the death of per-seat pricing.
Per-seat pricing worked when humans were the bottleneck. You paid for each person who needed access to the tool. But AI agents do not need seats. An agent that processes 10,000 support tickets per month does not care about a "per-user" license. This breaks the entire pricing model that SaaS was built on.
The companies adapting fastest are moving to consumption-based or outcome-based pricing. Charge per API call, per workflow execution, per successful resolution. This aligns incentives better (you pay for value delivered, not access granted) but it also makes revenue less predictable for the vendor, which is part of why Wall Street is nervous.
For buyers, this shift is mostly good news. Consumption-based pricing means you pay for what you use. No more "shelfware" where you are paying for 200 seats but only 40 people log in regularly. But it also means you need better visibility into your actual usage patterns, which circles back to the observability and cost management tools that are (not coincidentally) some of the best-performing software stocks right now.
The next 24 months will be messy. Some tools you rely on today will raise prices to compensate for shrinking customer bases. Others will slash prices to compete with AI alternatives. A few will simply shut down. The companies that navigate this well will be the ones that made deliberate choices about what to keep, what to replace, and what to build - instead of waiting for the market to decide for them.
Start the audit now. The repricing is not coming. It is here.