The State of SaaS in 2026: AI Is Breaking Apart the Application Layer

For much of 2026, software investors have been asking whether Claude, ChatGPT and other general-purpose AI systems will start replacing SaaS products rather than merely adding features to them. Claude already writes code, analyses spreadsheets, builds internal tools, creates websites and produces presentations. A task that once required several applications now often starts inside one model.

A satirical website called Death by Clawd has turned that anxiety into a scoring system. Users enter the name of a SaaS company and receive a mock “death score” estimating how easily Anthropic’s Claude might replace it. Adobe and ServiceNow fare relatively well, while Workday looks more exposed. The joke works because the underlying test is serious: what does the software own that the model does not? OMR Reviews, a German B2B software platform, showed year-on-year traffic down 83 percent for AI image-generation sites and 72 percent for AI text-generation tools, while AI-agent traffic rose 108 percent and AI sales assistants gained 162 percent. General models absorbed tasks that once justified a dedicated interface, while products connected to a larger workflow gained ground.

Software spending is still rising. Gartner expects worldwide software expenditure to reach about $1.43 trillion in 2026, 14.7 percent more than in 2025, while global IT spending should reach about $6.37 trillion. SaaS is not facing a collapse in demand. Vendors are competing for the same spending through a different architecture.

The interface no longer guarantees the business

For most of the SaaS era, companies bought a separate application for each category of work. Salesforce handled CRM, Jira project management, Adobe creative work, Tableau analytics and SAP core enterprise processes. Hundreds of smaller products filled narrower gaps between them. Foundation models now enter many of those tasks before the specialist application does.

A marketer asks Claude to create the first presentation rather than opening PowerPoint. An analyst uploads a spreadsheet to an AI model rather than starting in a visualisation tool. A small company describes an internal workflow and asks an AI coding system to build it. Designers use generative systems to produce first concepts before opening specialist editing software. Claude Design shows how far the interface has moved. Anthropic, the US AI company behind Claude, introduced the tool so users can describe a presentation, webpage or marketing asset, provide brand material and receive a first working version inside Claude. More than 1 million people used Claude Design during its first week.

PowerPoint does not disappear because Claude creates slides, but it no longer owns the beginning of the task. The same pressure applies to SaaS products whose value sits mostly in the interface. If a €50-a-month application accepts information, applies a relatively simple process and returns a document, a general model now competes with it directly. The vendor needs something else to defend the subscription.

SAP sits at the opposite end of the spectrum. Companies use it to store records, run processes, enforce permissions and connect finance, procurement, tax and reporting systems. Reproducing a SAP screen with AI does not reproduce the company infrastructure behind it. SaaS products therefore face very different AI exposure even when analysts place them in the same software market.

Klarna did not replace SaaS with AI

Klarna became the most cited case for the claim that companies would use AI to build their own technology and cut software vendors out. The Swedish buy-now-pay-later company abandoned Salesforce as its CRM and moved away from Workday. Its AI customer-service assistant took over work Klarna initially compared with hundreds of full-time agents, while the company cut external supplier spending and promoted AI adoption across its operations.

The simplified version of the story said Klarna had replaced enterprise SaaS with internally built AI. In practice, Klarna replaced Workday with Deel, another SaaS provider focused on payroll and workforce management. It combined internal technology with external products for CRM and continued using Slack, which Salesforce owns. CEO Sebastian Siemiatkowski later warned against assuming that every company should copy Klarna’s decision to leave Salesforce.

Its supplier cuts show where AI produced the larger financial effect. Klarna reduced or cancelled contracts with more than 1’700 suppliers as it standardised operations and adopted AI. Spending on external marketing suppliers including translation, production, CRM and social agencies fell by $7 million in the first quarter of 2024 compared with the same period a year earlier. Klarna attributed 37 percent of the reduction to AI and estimated another $10 million in savings from AI-powered marketing during 2024. By 2025, its AI assistant handled 80 percent of customer-service chats, while annual revenue per employee had risen from about $344’000 in 2022 to roughly $1.24 million.

Klarna did not build every function itself. It removed suppliers where AI changed the economics, switched SaaS providers where another vendor offered a better product and kept outside software where buying remained more sensible than building. That purchasing behaviour now appears across the wider market.

German companies are buying AI, not building everything themselves

The Munich-based ifo Institute, one of Germany’s main economic research institutes, found in May 2026 that 54.5 percent of German companies already used AI in their business processes, up from 40.9 percent a year earlier. Adoption reached 67.2 percent among large companies and 51.2 percent among small businesses. Only 18.7 percent of AI-using companies developed their own systems.

Almost three quarters relied on paid external AI applications, while 48.4 percent used free external tools. Among self-employed people and microbusinesses, only 7 percent used systems they had developed themselves. Most German companies are not replacing vendors with internal AI engineering teams. They are changing what they buy.

A company might keep SAP for finance and procurement, use an AI-native CRM, give employees Claude for analysis, build one narrow internal application and retain specialist software where regulation, proprietary data or integration makes an outside product more valuable. The stack becomes more selective rather than fully internal.

Lower development costs still put pressure on vendors. An internal tool that once required an engineering project now costs less to prototype, and AI-native competitors reach the market faster. OMR Reviews points to Attio and Folk, newer CRM platforms, Linear in project management and n8n in workflow automation as challengers in categories previously dominated by larger incumbents. Customers now have more alternatives for each layer of software: keep the incumbent, buy a newer SaaS product, use a general model directly or build a narrow application themselves.

Systems of record remain harder to displace

AI threatens software unevenly because companies use different applications for different purposes. Systems of record hold information the organisation cannot afford to lose or corrupt. Payroll systems, accounting platforms, ERP software and core CRM databases contain permissions, histories, integrations and business rules accumulated over years. Replacing one means migrating more than an interface.

Systems of work face more pressure where employees use them mainly to create, analyse, summarise or transform information. Foundation models now perform many of those activities directly and often pull information from the systems underneath. A company still needs a reliable customer record, but it does not necessarily need the same application every time an employee wants to summarise the account, prepare a sales briefing or draft a follow-up.

OMR Reviews, which tracks software products and buying behaviour across the German-speaking market, has not seen a broad DACH exodus from systems such as SAP or Salesforce. It has seen lower barriers for competitors attacking narrower workflows around those systems. The SaaS vendor most exposed to AI in 2026 is therefore not automatically the smallest one. Exposure depends on how much of the product’s value remains after the user removes the interface.

Software is starting to sell work rather than tools

Traditional SaaS sold employees software and left the work to them. A salesperson used a CRM to organise prospects, a designer used software to produce an asset, a developer used an IDE to write code and an analyst used a dashboard tool to investigate data.

AI products now execute part of the task. An AI sales agent researches prospects, ranks them and drafts outreach. A design model produces a first campaign visual. An agentic coding system receives a requirement and writes parts of the application. Less human work sits between instruction and output, which pushes software vendors into budgets beyond the IT department.

Klarna’s reduction in agency and supplier spending shows the effect. AI did not merely move money from one software subscription to another. It removed expenditure previously paid to translators, production companies, agencies and service providers.

Matthew Gallagher, the founder of US telehealth company Medvi, offers an extreme example of AI-enabled operating leverage. Medvi sells access to weight-loss treatments and reportedly used more than a dozen AI tools across software development, customer service, website copy, advertising and business analysis. Gallagher outsourced functions that still required external providers rather than staffing each capability internally.

Medvi generated $401 million in sales during 2025. By 2026, Gallagher and his brother were reportedly its only two employees and the company was on track for around $1.8 billion in annual sales. Medvi remains an exceptional case and its marketing practices have attracted scrutiny, but its staffing structure explains why investors pay so much attention to agentic software. AI gives small companies access to functions that once required employees, agencies or outsourced teams, and the market opportunity expands when software replaces paid work rather than another piece of software.

Thin AI products are already losing ground

A dedicated text-generation application competes with Claude, ChatGPT and Gemini, which already generate text. A stand-alone image generator faces the same pressure when general platforms include high-quality image generation inside a broader subscription. Adding another interface around the same underlying capability becomes harder to price.

AI agents occupy a different position when they connect several steps of a business process. A sales agent linked to CRM data, account history and prospecting systems does more than generate an email. It researches the account, decides what information matters, prepares the message and records the activity. Proprietary data, integrations, permissions, workflow depth, regulatory infrastructure and distribution become the parts worth paying for. An AI button inside an existing SaaS product offers little protection on its own.

Incumbents want to own the AI layer themselves

Salesforce is responding by bringing agents closer to the data already stored inside Salesforce. Its Agentforce product is an AI-agent platform designed to automate tasks using customer data, workflows and permissions already held inside the system. By its second fiscal quarter of 2027, reported in August 2026, Agentforce annual recurring revenue had passed $1.5 billion, up more than 240 percent year on year. Agentforce and Data 360 together approached $3.9 billion in ARR.

Salesforce owns customer histories, permissions, workflows and integrations inside thousands of organisations. Its defence against general-purpose AI rests on making those assets the environment where agents act. Anthropic, OpenAI and Google are coming from the other direction, connecting their models to external systems so users stay inside the AI interface while working with company data elsewhere. AI-native SaaS companies attack individual workflows between them.

Vendors now compete not only over who stores the data, but over where the employee gives the instruction and where the work gets completed.

AI has also entered software purchasing

US software marketplace and review platform G2 found in its 2026 research that 51 percent of B2B software buyers start research with an AI chatbot more often than with Google. Seventy-one percent use AI chatbots somewhere during software research. Among respondents, 69 percent eventually selected a different vendor from the one they initially expected, while 33 percent bought from a company they had not known before AI surfaced it.

DACH buyers remain more conservative. OMR Reviews found that roughly one in four software buyers in Germany, Austria and Switzerland currently uses AI chatbots for software research. Integration remains a major buying criterion and 81 percent expect GDPR compliance.

Google gives buyers links. AI systems often return a shortlist with explanations and comparisons, which means a vendor that never appears in the answer loses the opportunity before anyone visits its website. OMR Reviews reports that 80 to 90 percent of brand mentions influencing AI-search visibility come from domains outside the vendor’s own website. Reviews, industry publications, editorial coverage, comparison sites, videos and customer discussions supply much of the material models use to understand which companies belong in a category.

A SaaS website still needs clear product information. Models need readable text explaining what the product does, who uses it, where it operates, what it integrates with and how it differs from alternatives. Important information buried inside graphics or interfaces remains harder to extract. Third-party evidence also carries more weight because the vendor cannot manufacture all of its credibility on its own domain.

OMR Reviews highlighted one experiment in which a fictitious agency reached a ChatGPT recommendation within 14 days after appearing in engineered list articles. The experiment shows how heavily current AI-search systems still rely on external mentions and how quickly marketers will try to manipulate them. Longer-term visibility depends on evidence the company does not fully control, including customer reviews, specialist media references, consistent product descriptions and credible third-party discussions.

SaaS in 2026 has lost the right to charge for an interface alone

Companies are still spending more on software, while AI gives them more ways to avoid individual software products. European companies already show what that looks like in practice. Most use outside AI services rather than build their own systems. Klarna kept external vendors while cutting others. Salesforce is putting agents on top of its own data. New challengers attack individual workflows, and foundation models absorb tasks that previously justified separate applications.

A SaaS company now needs to own more than the screen through which the user performs the task. Proprietary data, deep integrations, regulated processes, network effects, critical records and workflows that reach far into an organisation remain difficult to reproduce. Products that mainly package a task behind a convenient interface face a weaker position once a general model performs the same work inside a conversation.

SaaS is not dying in 2026. The application layer is being repriced.