Which SaaS ideas will work after the SaaSpocalypse?

What is the SaaSpocalypse actually killing?

The SaaSpocalypse is hitting generic SaaS much harder than software as a whole.

The clearest change is that software itself has become much easier to create. Features that once required a funded startup, several engineers and months of work can increasingly be built with coding agents, APIs and existing infrastructure. That puts pressure on SaaS products whose main advantage was simply that customers did not want to build the software themselves.

We can already see this affecting real buying decisions. McKinsey’s latest State of AI survey found that 32% of organizations had decided against buying at least one software product or feature because agentic coding tools let them build it internally. Among large companies, 31% were already scaling coding agents.

Retool found something similar in a survey of 817 builders. Some 35% said their teams had already replaced at least one SaaS tool with custom software, while 78% expected to build more internal tools. Workflow automation, internal admin software and BI dashboards were the three categories being replaced most often.

But companies have not suddenly stopped paying for SaaS. Zylo’s latest dataset shows average annual SaaS spending reaching $55.7 million per organization, up 8% year over year. BetterCloud also found that the average SaaS portfolio grew 11% after two years of contraction, helped largely by AI-powered applications.

So the market is splitting. Software that mostly packages code behind a convenient interface has become easier to challenge. Software tied to proprietary data, difficult operations, regulation, transactions, security or a system customers cannot afford to break looks much healthier.

Getting easier to replace	Still difficult to replace
Basic dashboards	Authoritative business records
Simple internal tools	Regulated workflows
Generic AI wrappers	Security and governance
CRUD applications	Payments and transactions
Basic reporting	Physical operations
Feature bundles	Deep vertical workflows

Are companies really replacing SaaS with software they build themselves?

Yes. Companies are already replacing some SaaS with custom software, although the replacement is concentrated in relatively simple applications.

Retool’s data shows where the pressure is strongest. Workflow automation was cited by 35% of respondents who were replacing software, followed by internal admin and operations tools at 33%, BI and reporting at 29%, CRM and form-builder use cases at 25%, project management at 23% and customer support at 21%.

Those categories share a useful characteristic: the company usually owns the underlying data and understands the workflow already. Once generating an interface, database queries and business logic becomes much easier, paying another vendor can look unnecessary.

One Retool customer, Harmonic, rebuilt a tool that had cost roughly $20,000 a year because waiting for the vendor’s support had become slower than recreating the product internally. Harmonic eventually operated 33 internal applications connected to systems including Salesforce, Gong and Slack.

Still, generating version one is the easy part. Production software needs permissions, authentication, integrations, audit logs, backups, maintenance and someone responsible when an API changes at the worst possible moment.

That keeps the threat uneven. A company may happily rebuild a simple employee dashboard. Rebuilding payroll, security infrastructure or a system processing millions of dollars carries a very different risk.

Founders should assume that a product a competent customer can reproduce with an AI coding agent over a weekend will face serious pricing pressure.

Which SaaS products are easiest for AI to replace now?

Simple workflow software, generic dashboards and thin AI wrappers look most exposed right now.

Internal admin panels are an obvious case. So are lightweight approval tools, uncomplicated CRMs, basic reporting dashboards, form builders, simple project trackers and small workflow automations.

These products benefited for years from the high cost of custom development. A company might tolerate paying $20,000 or $50,000 annually for software because reproducing it internally would require months of engineering work. AI coding changes that calculation.

Generic AI applications have another problem: their underlying capability keeps moving into broader platforms. Basic summarization, rewriting, document chat, research assistance, meeting notes and content generation are increasingly available inside ChatGPT, Claude, Gemini, Microsoft 365 and other large products.

Retention data gives us another reason to be cautious. RevenueCat’s latest subscription-app analysis found median annual retention of 21.1% for AI-powered subscription apps versus 30.7% for non-AI apps. That dataset includes consumer software as well as business tools, so we would not use those percentages as an enterprise SaaS benchmark. The gap is still hard to ignore.

An AI feature can attract users very quickly. Keeping them once the same capability appears somewhere else is much harder.

The SaaS ideas in the most danger are the ones where the customer essentially pays for access to a feature rather than for an ongoing operational dependency.

Will systems of record and workflow SaaS survive AI?

Yes. Systems of record and software that actually runs important workflows are among the better places to be after the SaaSpocalypse.

Salesforce, Workday and ServiceNow remain useful tests because AI should theoretically threaten enormous parts of what their users once did manually.

Instead, the companies are putting agents on top of the records and workflows they already control.

Workday recently said AI generated more than 25% of new annual contract value in its latest quarter, while more than 5,500 customers were already using at least one of its own agents. Salesforce subsequently reported Agentforce ARR above $1.5 billion and said Agentforce plus Data 360 had reached nearly $3.9 billion in ARR. ServiceNow has also built a billion-dollar-scale AI business on top of its workflow platform.

The user interface may matter less over time. An employee could ask an agent to update an opportunity, approve an expense or open an IT ticket without spending much time inside the original application.

Yet Salesforce still needs to know which opportunity exists. Workday still needs the employee record, payroll rules and permissions. ServiceNow still needs the workflow state and audit history.

This changes what good workflow SaaS should aim to do.

A tool that merely tells an employee that invoice 481 is overdue has a weak position. A product that checks the invoice, reads the contract, contacts the customer, handles the response, escalates an exception and records everything inside the company’s financial workflow is much harder to remove.

The useful test for a SaaS founder now is simple: if an AI agent stopped opening our interface tomorrow, would the agent still need our product to finish the job?

For the strongest SaaS products, the answer is yes.

Is vertical SaaS safer than horizontal SaaS now?

Yes. Deep vertical SaaS looks considerably safer than another generic productivity application.

Veeva gives us one of the cleanest examples. The company builds software specifically for life sciences, including clinical, regulatory and commercial workflows. Its latest quarterly revenue reached $928 million, up 18%, while subscription revenue grew 16% to almost $767 million.

Veeva is also adding AI without abandoning its vertical model. The company is combining specialized applications, industry data and agents inside workflows that pharmaceutical companies already depend on.

Toast follows a different version of the same playbook. Restaurants use Toast for point-of-sale software, ordering, payments and day-to-day operations. Samsara connects software with vehicles, cameras, sensors and industrial equipment. Samsara has now passed $2.1 billion in ARR, with large enterprises increasingly standardizing on its platform.

These businesses have very different customers, but the reason they hold up is similar. Their software sits inside the actual operation.

That is a much stronger position than building “ChatGPT for restaurants” or “an AI assistant for construction.”

The best vertical SaaS ideas now tend to combine several things that would be irritating for a customer to recreate: specialized records, industry-specific integrations, unusual edge cases, changing rules and actions that have to happen reliably every day.

Is cybersecurity one of the best SaaS markets after the SaaSpocalypse?

Yes. Cybersecurity is one of the strongest SaaS markets because AI is creating new security problems almost as quickly as it solves old ones.

CrowdStrike’s latest quarter gives us a useful reality check. ARR reached $5.84 billion, up 25% year over year, while the company added a record $333 million of net new ARR during the quarter.

The bigger opportunity goes beyond traditional endpoint security. Companies are introducing agents that can access internal databases, call APIs, open SaaS applications and act under delegated credentials. That creates new questions around machine identity, permissions, data leakage, malicious instructions and what an agent should be allowed to do without human approval.

BetterCloud’s latest State of SaaS report shows how messy the environment already is. Companies in its survey were using an average of 27 AI-powered SaaS applications, yet only 56% of all applications in use had IT approval. One in five organizations had discovered sensitive corporate data being shared publicly during the previous year.

That makes agent identity, AI access control, connector security, AI data-loss prevention, model governance and autonomous security remediation especially interesting SaaS categories.

Customers can tolerate an improvised content tool failing. They are far less relaxed about an improvised security layer failing.

That difference gives security SaaS pricing power even while many ordinary software categories are being squeezed.

Will AI governance become a real SaaS market?

Yes, but the useful AI governance products will control what agents can actually do rather than produce another compliance dashboard.

The governance problem is expanding because companies can deploy AI much faster than they can supervise it. BetterCloud found an average of 27 AI-enabled applications per organization, representing roughly 22% of the SaaS portfolio. Only 56% of applications had IT approval.

Retool found another version of the same problem: 60% of respondents had built software outside official IT oversight during the previous year. Among those surveyed, 51% had already shipped production software using AI.

That means companies increasingly need to answer very practical questions. Which agent accessed Salesforce? Whose credentials did it use? What customer data did it retrieve? Which external tool did it send information to? Who approved the action? Can the company reconstruct what happened three months later?

A governance product that merely tells employees how they should use AI will be easy to copy.

A product sitting between agents and production systems can become much harder to remove. Agent authorization, machine identity, policy enforcement, model routing, audit trails, AI asset discovery and approval gates all fit that pattern.

The easier it becomes for employees to create software and agents, the more software someone has to govern.

Does data infrastructure get stronger when software gets easier to build?

Yes. Cheap application development can actually increase demand for the infrastructure underneath those applications.

Every new internal app or AI agent creates more API calls, data movement, logs, permissions, production failures and security issues. Someone still has to keep that environment working.

Snowflake, Datadog and Cloudflare have all continued growing quickly while fears about a SaaS collapse have intensified. Their businesses sit underneath applications rather than depending mainly on employees opening another interface.

The logic also works at a smaller scale.

A company may no longer pay much for a dashboard that shows warehouse data. It can probably generate that dashboard itself.

But the company still needs clean warehouse data. It still needs to know whether pipelines broke, whether the AI agent received stale information, who changed the underlying table and whether sensitive records were exposed.

This creates opportunities in data quality, lineage, synchronization, AI observability, agent evaluation, infrastructure monitoring and integration maintenance.

Large cloud vendors and open-source software still compete aggressively here, so infrastructure is hardly automatic safety. But infrastructure tied directly to growing machine activity has a much better setup than another human-facing productivity dashboard.

Are compliance-heavy SaaS ideas still attractive?

Yes. Compliance SaaS remains attractive because customers are paying somebody to make a regulated process work correctly, not merely to give them software.

Payroll makes the distinction easy to see. Producing a payroll interface is increasingly trivial. Correctly applying tax rules, employee classifications, deductions, filing deadlines and jurisdiction-specific requirements is not.

The same applies to pharmaceutical compliance, healthcare documentation, financial reporting, privacy, customs, insurance, environmental reporting and workplace safety.

AI should help these companies rather than automatically destroy them. Models can extract evidence, read documents, classify records, detect missing information and prepare filings. The SaaS vendor can therefore automate a larger percentage of the work.

The best opportunities go further than giving the customer regulatory information. An LLM can already explain many regulations reasonably well.

What companies will keep paying for is software that monitors the operation, notices when something falls out of compliance, gathers evidence, fixes what can be fixed and preserves an audit trail.

The closer the product gets to proving that the company actually followed the rule, the stronger the SaaS becomes.

Is seat-based SaaS pricing dying?

Seat-based SaaS pricing is weakening, especially when AI agents do work that used to require several human users.

The old model was straightforward: more employees used the application, so the customer bought more seats.

That relationship breaks when one employee supervises several agents doing the work of a larger team.

Pricing data is already moving in that direction. Vertice recently found consumption-based pricing had become the largest individual pricing model in its dataset at 36.5%, ahead of per-user pricing at 32.4%. Another 31.1% of vendors used hybrid approaches.

Salesforce offers an interesting example of what could come next. It measures Agentic Work Units, meaning tasks performed by agents rather than humans occupying conventional software seats.

A collections product could charge for invoices processed. A support agent could charge for resolutions. A security product could charge for workloads or identities. A reconciliation system could price against transaction volume.

Consumption pricing also creates a problem that founders should not ignore. Zylo found that 78% of IT leaders had experienced unexpected charges from AI functionality or consumption-based pricing. Some 61% said unexpected SaaS cost increases had forced them to cut projects.

Customers want pricing tied to value, but they also want to know roughly what the bill will be.

The strongest post-SaaSpocalypse pricing models will probably combine both ideas: a predictable platform fee plus usage or outcomes that scale with the work performed.

Does proprietary data still protect an AI SaaS company?

Yes, but useful proprietary data comes from owning a workflow, not simply collecting a large pile of information.

Most AI SaaS companies can access similar frontier models. If two competitors are both calling the same underlying model, the model itself gives neither company much protection.

The difference becomes what that model knows about the customer.

A generic sales agent might understand public information about a prospect. An agent sitting inside years of opportunities, emails, support history, renewals and pricing exceptions can make much better decisions.

The strongest datasets also keep changing because customers create them while using the product.

Samsara receives operational telemetry from vehicles and equipment. Toast participates in restaurant transactions. CrowdStrike sees security activity. Datadog monitors production systems. Workday holds years of HR and financial records.

That information gives the software context that a newly generated application does not automatically have.

Static data is less impressive. Public records, common datasets and information that can easily be licensed or exported will rarely create a durable moat on their own.

A more useful question for a founder is whether each month of customer usage leaves the product with context that makes replacing it more painful.

If six months of usage changes nothing, the customer can often switch easily.

If six months creates valuable history, rules and operational context, AI can make the product stronger instead of making it disposable.

Can a small SaaS business still work after the SaaSpocalypse?

Yes. Small SaaS can still work, but narrow operational products look much better than narrow feature products.

A small SaaS company does not need Salesforce-level lock-in. It needs one recurring problem that customers really do not want to maintain themselves.

A contractor might build a simple project calculator with AI and never pay for one again. The same contractor may happily subscribe to software that monitors permits across dozens of municipalities, notices changes and updates active projects.

An accountant can ask Claude to interpret a document. A product that continuously imports client records, checks transactions, finds missing documents and prepares exceptions for review takes on much more responsibility.

A logistics company could build a dashboard internally. Maintaining connections to dozens of carrier portals, normalizing inconsistent data and handling failures every day is less appealing.

That is the opportunity for micro-SaaS and bootstrapped SaaS now.

The customer does not need the problem to be technically impossible. The problem needs to be annoying enough, repetitive enough and consequential enough that outsourcing it remains worthwhile.

Which SaaS ideas look strongest after the SaaSpocalypse?

The strongest SaaS ideas today control important workflows, risks, data or transactions that customers cannot casually rebuild and forget about.

AI security is one obvious area. Agent identities, permissions, MCP and connector security, AI data-loss prevention and agent audit trails all become more important as autonomous software spreads.

Vertical operations is another. Instead of building a generic AI agent, a startup can own a specific process in healthcare, logistics, insurance, construction, finance or another industry where dozens of integrations and edge cases accumulate over time.

Compliance automation has similar characteristics. Data quality and agent observability benefit from the rapid creation of AI systems. Transaction software gains an advantage because it participates directly in money movement. Integration maintenance becomes more useful as companies connect an expanding number of custom applications.

Physical-world SaaS is particularly interesting because AI cannot generate away the need to collect information from trucks, factories, buildings, equipment or stores.

The common thread is recurring responsibility. Customers keep paying when removing the vendor means taking that responsibility back in-house.

SaaS idea	Why customers may keep paying	Natural pricing
AI agent identity and permissions	Agents create new access risks	Agent + platform
MCP and connector security	AI needs controlled access to company systems	Connection + usage
AI audit infrastructure	Companies need records of autonomous actions	Agent run + platform
Vertical workflow agents	Industry rules and edge cases accumulate	Workflow + outcome
Compliance automation	Rules change and mistakes have consequences	Entity + platform
AI data-quality monitoring	Agents depend on reliable company data	Data volume
Agent observability	Autonomous workflows need debugging	Traces + usage
Physical-operations SaaS	Software connects to real-world assets	Asset + location
Transaction software	The product sits directly inside money movement	Transaction
Vertical systems of record	History and workflow make switching painful	Platform + usage
Integration maintenance	APIs continually change and fail	Connection + execution
Exception automation	Difficult cases still need reliable handling	Case + outcome

Which SaaS ideas should founders avoid now?

We would be very cautious about SaaS products whose entire pitch is a nicer interface around something a general AI platform can already do.

Generic writing software, uncomplicated summarizers, basic PDF chat, simple meeting notes, general research assistants and thin customer-support bots are obvious examples.

Basic analytics dashboards also look vulnerable when the customer owns the underlying data. If an AI agent can query that data directly and explain what changed, simply displaying it in charts becomes less valuable.

Simple internal CRUD applications face the same problem. Retool’s latest survey already shows internal admin, workflow automation and BI among the categories companies are replacing most often.

Broad “AI employee” products deserve caution too. A startup promising to handle everything from research to email to project management ends up competing directly with the largest model providers. Owning one painful workflow is usually a better position than claiming to replace an entire employee.

We would also avoid assuming an incumbent’s missing AI feature creates a permanent opening. Workday says AI already generates more than 25% of new ACV. Veeva is adding specialized agents to life-sciences workflows. Salesforce, ServiceNow and other major vendors are doing the same inside products customers already trust.

The dangerous place to build is one step above the model with little else underneath.

The interesting place is deep inside the customer’s operation.

So which SaaS ideas will actually work after the SaaSpocalypse?

The SaaSpocalypse is real for generic software, but the evidence does not support the idea that SaaS itself is dying.

McKinsey now finds that 32% of organizations have skipped at least one software purchase because coding agents let them build internally. Retool sees 35% of surveyed teams already replacing at least one SaaS product with custom software.

At the same time, Zylo finds SaaS spending up 8%, BetterCloud sees application portfolios growing again, and companies such as CrowdStrike, Veeva and Samsara are still producing strong double-digit growth.

Those facts fit together.

Writing software has become dramatically cheaper. Owning a difficult business process has not.

The SaaS ideas we would back today are therefore concentrated around vertical workflows, cybersecurity and AI governance, data infrastructure, compliance, transactions, physical operations and the tooling needed to control agents and internally generated software.

Generic dashboards, thin wrappers and simple productivity features face a much harder future because customers have more ways to reproduce them.

The best post-SaaSpocalypse founders will stop asking whether they can build a useful app. That bar has become too low.

The better question is whether a customer will still want someone else to own this problem after creating the software itself becomes almost free.

That is where SaaS still has room to build very large businesses.