[Issue Report] Only SaaS With Workflows Will Survive

The issue report, "SaaS Survival Depends on Owning Workflows," looks at how Kim Young-wook, SAP product engineering product expert, sees changes in the enterprise software market in the first half of 2026. The subhead says that most SaaS now serves as a data repository.

In the first half of 2026, Utila, an Israeli crypto asset management software startup, moved to cut its software budget by 50% and canceled 10 SaaS products across B2B GTM, sales, and marketing to do so. The list of canceled tools included the rapidly growing Clay and Vendelux. Utila replaced them with one AI agent, which integrated all the functions previously handled by the 10 apps, including lead tracking, email sending, and marketing campaign management.

As a result, HubSpot was the only app it kept, and HubSpot’s role was that of a data repository. This case shows a shift in which many fast-growing SaaS products were eliminated and sales and marketing functions were consolidated into a single AI agent, sharply reducing costs, while HubSpot’s remaining role was also reduced from a work execution tool to a data storage function. More broadly, it clearly illustrates the essence of the 2026 structural shift in the enterprise software market.

Sanofi built its own AI agent based on Anthropic Claude Code and startup software, reduced its use of ServiceNow, and replaced purchase order audit work previously handled by an outsourcing vendor. Starbucks also reviewed AI alternatives to Microsoft and IBM products. As companies across different industries, sizes, and regions moved in the same direction, the phenomenon pointed to a structural shift rather than a temporary trend.

The market had already recognized the meaning of that shift. In February 2026, the S&P 500 software and services index lost USD 2 trillion in just two days, and the market dubbed the event "SaaSpocalypse." The trigger was seen as the overlap of Anthropic’s launch of enterprise agent plugins and the agent product launches from Salesforce, ServiceNow, and Google, but behind it all was a deeper structural recognition.

As AI agents began autonomously carrying out workflows, the SaaS pricing formula tied value to human dashboard logins is being fundamentally shaken. Gartner said the link between user growth and revenue growth is breaking down.

Against this backdrop, AI-native company spending in the first quarter of 2026 rose 94% from a year earlier, while traditional SaaS growth came in at just 8%. Gartner projected that agent AI would crowd into enterprise software spending and reach USD 234 billion by 2030. Deloitte predicted that more than 50% of enterprise digital transformation budgets in 2026 would be allocated to AI automation. As a result, the SaaS subscription model that has been a core item in enterprise IT budgets for 20 years is facing structural pressure.

The purpose of this article is to analyze that pressure structure. Its core argument is that the survival criteria for enterprise software are changing in the AI agent era. In this environment, feature richness or polished user experience are no longer the standard; the real standard is whether a product owns the workflow versus whether it merely stores data. That premise rests on agents directly executing workflows through APIs, and as a result, software that loses its execution layer becomes a data warehouse.

In this shift, HubSpot’s fate is cited as a case in point, and that outlook was foreshadowed. In this context, interpreting "SaaSpocalypse" as a simple stock correction is deemed inappropriate. Its meaning lies not in the emergence of a new threat, but in the market’s visible recognition that a structural transition already underway has crossed a critical threshold. Accordingly, investors are left asking whether the current prices of software built on the assumption of user logins are justified in a world where AI agents execute workflows on users’ behalf.

Both Salesforce and ServiceNow reported results above market expectations through the immediately preceding quarter. That makes it possible to interpret the decline as something other than earnings weakness. In the end, this move is read as a preemptive revaluation of structural vulnerability in the business model.

AI agents and legacy RPA differ fundamentally in how they interact with software. While RPA works by clicking screen buttons and following predefined procedures, AI agents access systems directly through APIs, handle unstructured inputs, and autonomously carry out multi-step decisions. For that reason, once an agent has access to software, the existence of a sophisticated user interface is no longer a criterion for value, and even dashboard design and UX optimization backed by billions of dollars in R&D lose their meaning in front of agents.

The market reaction foreshadowed what would come next, and actual enterprise budget data confirms that the forecast has become reality. The gap between 94% growth in AI-native companies and 8% growth in traditional SaaS is not simply a difference in adoption speed for new technology; it is a signal that budget items are structurally shifting within the finite total of enterprise IT budgets.

The source of corporate AI budgets is emerging from items different from traditional IT budgets. According to Deloitte report analysis, a significant portion of corporate AI automation budgets is being shifted not from software, but from OpEx, and the more specific source within OpEx was analyzed as labor costs.

Accordingly, companies are moving to replace both software and labor with AI at the same time. That said, there is also software that benefits from this shift. In Sanofi’s case, reduced use of ServiceNow was accompanied by increased use of SAP.

This shows that even as AI agent investment expands, the status of ERP systems such as SAP is strengthening. This transition does not mean the disappearance of all software; it suggests a selective reorganization within software itself.

Under the traditional Per-Seat model, revenue came from "number of users × seat price," but software vendors, recognizing the collapse of the Per-Seat model, are rolling out billing strategies that seek to capture revenue at the data access point rather than from user counts. Their defensive strategy is to charge separately when AI agents outside the platform access their data, and that strategy has been dubbed the "Agent Access Tax." HubSpot told investors about this strategy in advance and plans to charge for external AI agents calling its CRM data. HubSpot argued that it can monetize data access regardless of declining user logins. ServiceNow is also reviewing a similar direction, while SAP and Oracle have already established similar strategies through "indirect access" billing policies.

In the short term, this strategy may be somewhat effective as a means of defending revenue. Platforms that own the data can maintain bargaining power as long as agents continue to need that data. But there is also a structural paradox. Vendors that overcharge for agent access may create an opposite outcome by giving customers an economic incentive to leave the platform.

According to Utila’s leadership, even if HubSpot charges external agents, usage will continue. But the background for that is not workflow value; it is the cost burden of moving previously accumulated data, and that data migration cost functions as lock-in. If that lock-in disappears, the effect of the Agent Access Tax will shift from keeping customers to accelerating customer churn.

Companies have already identified this risk and are making agent access conditions an explicit negotiation point when renewing software contracts. They are asking about the cost structure for agents reading and writing their own data, whether there are differential charges between external agents and the vendor’s own agents, and whether surcharges kick in if API call volumes spike. These questions are becoming standard topics in 2026 enterprise software procurement negotiations. In contrast, companies that have not finalized contracts are at risk of facing unexpected cost structures when agent deployment begins in earnest.

The ultimate limit of the Agent Access Tax strategy depends on how long data ownership lasts. The effectiveness of a hostage-style defense strategy also depends on maintaining data ownership. However, because corporate migration targets are independent lakehouses and because AI agents are strong at reconstructing data from multiple sources, single-platform data monopoly power weakens once companies move to independent lakehouses or once AI agents gain the ability to reconstruct from multiple sources.

In this scenario, the Agent Access Tax initially functions as a short-term revenue defense tool. But as platform switching costs fall, the role of the Agent Access Tax shifts and becomes a trigger for customer churn. Here lies the paradox of a defense strategy becoming a signal to exit.

3. The reality of budget movement inside companies is revealed in Sanofi’s case of building an AI agent. In Sanofi’s simultaneous reallocation, the three cost items were software subscriptions, partner software, and labor outsourcing. The key point is not simple cost cutting, but the choice structure.

Even with the same decision to introduce AI agents, Sanofi illustrates an asymmetric case of the survival rules for software in the AI agent era, where ServiceNow and SAP produce opposite outcomes. Reduced use of ServiceNow and expanded use of SAP are both under way, and this ties to the difference between areas that agents can perform directly and areas on which agents must rely. The area handled by ServiceNow is IT service management workflows, which agents may be able to perform directly, while SAP’s ERP data consists of purchase orders, financial records, and supply chain information, highlighting ERP data as a necessary access target for agents to operate.

The nature of SAP data is a "single source of truth," which is why agents do not replace SAP and instead operate on top of it. Agents cannot replace SAP, but software responsible for workflow execution becomes the replacement target. The product known to have replaced ServiceNow in Sanofi’s case is elementum.ai.

In this flow, AI adoption triggers a renewed recognition of legacy technical debt, and companies are moving their strategy toward simplifying the software stack. The shift in AI adoption approach is not to add to existing software, but to replace it with AI. And the first targets for replacement are not software with many features, but software that plays the role of workflow execution.

Software that owns the data survives, and data ownership rather than features is emerging as the criterion for survival. This order is already at work in the actual decision-making of enterprise buyers.

In this flow, the notable point in Sanofi’s case is the reduced role of the Indian outsourcing vendor. The meaning of this change goes beyond a reshaping of the software market and signals the structural end of the Labor Arbitrage model. Labor arbitrage is the practice of companies in developed markets outsourcing repetitive, large-scale, labor-intensive tasks to workers in low-wage countries to reduce costs. India was the biggest beneficiary of this model, and millions of members of the middle class in India were formed on top of it. But the conditions that made labor arbitrage possible are now being dismantled by AI agents. The conditions being dismantled are repetitiveness, predictability, and economies of scale, all of which are both the basis for outsourcing’s effectiveness and the primary targets for AI agent automation.

Large-scale layoffs and a slowdown in hiring are appearing across India’s IT industry. Oracle moved to cut 12,000 jobs in India in the first half of 2026 while also expanding AI investment. TCS also moved to lay off 12,000 people. India’s major IT companies posted net hiring growth of just 17 people in the first 9 months of fiscal 2026, and the Indian IT industry, which had long continued hiring in the tens of thousands, has effectively entered a hiring freeze.

This change is linked to a broader trend in which global companies are increasingly choosing AI agents instead of outsourcing. As AI agents accelerate the replacement of repetitive work, the structure of growth through large-scale human deployment is weakening, and this has direct implications for Korea’s IT services industry. This structural shift raises the same issue for large SI project labor-based revenue models and also places structural pressure on project billing methods based on workforce size.

That said, this pressure is both a threat and an opportunity for a new role transition. The new role is presented as agent-building capability, enterprise system integration capability, and secure orchestration capability. The purpose of these capabilities is to support the internalization of functions that Sanofi had outsourced, and whether companies can preemptively structure this new opportunity will determine future competitive advantage.

Traditional enterprise software has long been divided into System of Record and System of Engagement, and the two have worked together. Systems of Record serve as the single source of data, with ERP, CRM, and HRIS as representative examples, and they functioned as platforms for creating, storing, and referencing core enterprise data. Systems of Engagement were the interfaces for carrying out actual work based on record data, with email, collaboration tools, marketing automation platforms, and project management apps among them.

Under that traditional structure, the two layers coexisted for the long term, and the division of labor was clear. Data lived in record systems, while human interaction took place in engagement systems.

But with the arrival of AI agents, that coexistence structure is being fundamentally shaken. Agents have the ability to make the engagement layer nonessential, and once they receive a goal, they connect directly to the record system API to read data, make judgments, and record results. As a result, the intermediate steps that people once handled with engagement tools are compressed into autonomous execution by agents.

As a result, the engagement layer is demoted into an optional interface that agents can bypass, while record systems are elevated as data sources that agents must access. The meaning of this structural shift is presented as a point of survival, and the headline’s claim is that only SaaS with workflows will survive.

In the era of AI agent expansion, software survival depends on whether it functions as an "executor of workflows" or an "owner of data." Software that plays the role of an "executor of workflows" is replaced by agents, while software that plays the role of an "owner of data" remains a necessary access target for agents. Tools built for process convenience lose their reason to exist the moment agents can achieve the same result directly through API calls.

The operating criteria for this turning point have nothing to do with brand, nothing to do with market share, and nothing to do with customer relationships built over the past 10 years. Even software that has long been favored in the market faces structural pressure in the AI era if it depends on workflow execution.

Sanofi’s choice to reduce ServiceNow and expand SAP can be explained by this turning point. The logic behind Sanofi’s decision can also be explained by this same dividing line, and the case-based conclusion points to SAP’s survival and HubSpot’s continued role as a "data warehouse."

SAP’s defensive strength comes from its data and record layer. SAP is the single source of enterprise purchase orders, financial data, supply chain records, and HR information. Many AI agent tasks related to enterprise operations reference SAP data or include SAP records.

For that reason, agents are not replacing SAP but operating on top of it. Even in the AI agent era, SAP’s value is not declining; it is strengthening as an agent orchestration hub. Platforms that have accumulated core enterprise transaction data for decades will see more agent access requests as AI agents spread, and these platforms are also the most realistic layer for the Agent Access Tax debate to apply.

HubSpot, by contrast, follows a different path from SAP. HubSpot grew into an "all-in-one CRM and marketing platform" that runs multi-layer workflows for marketing automation, email campaigns, sales pipeline management, and customer support ticket handling on a single platform.

The core value of HubSpot’s positioning was integrated workflow execution. But as AI agents directly carry out email sending, lead tracking, and campaign management without going through the interface, the center of differentiation is shifting to the remaining core element: data. This data category includes customer information, transaction history, and communication records.

Even amid this change, the condition that HubSpot will not be canceled as long as the reference data remains inside the agent’s data set still holds. That said, HubSpot’s role is shrinking, and the trend is moving from workflow platform to data warehouse.

5. In GitHub’s crisis, the situation is described as being surrounded on three sides. The timing is July 2026, and GitHub is facing unprecedented competitive pressure. The two key business areas under attack are code repositories and AI coding assistance. Cursor has moved to create an AI-native code editor category based on a VS Code fork, and that move is presented as eroding GitHub Copilot’s lead in the AI coding market.

At the same time, OpenAI hired a former GitHub executive and then made a strategic declaration that it would enter the code repository market. OpenAI’s assets are presented as ChatGPT, with 900 million weekly active users, and its developer ecosystem. If OpenAI enters the code repository market, a change in GitHub’s competitive landscape becomes inevitable, leaving GitHub to compete head-on with OpenAI in the AI coding assistant business and again head-on with OpenAI in the repository business.

The startup Entire, founded by a former GitHub CEO, launched its own Git service, highlighting the emergence of a symbolic competitor. Because the provider of a GitHub alternative is someone who knows GitHub well, Entire’s strategy is read as targeting GitHub’s weak points.

Behind this trend lies a changing environment driven by the spread of AI agents. The main actors in code writing and committing are shifting from people to AI agents, and AI agent activity is taking the form of 24-hour autonomous code generation, modification, and PR submissions. As a result, traffic patterns have changed in the AI agent environment, and GitHub suffered 6 service outages in June alone, with the main cause cited as a sharp increase in AI agent traffic.

GitHub has moved to migrate from its own servers to Microsoft Azure infrastructure, and GitHub’s position is that the move aims to reduce outages. But during the period before the transition is complete, competitors are moving to exploit the gap, and this is unfolding alongside cracks in Microsoft’s vertically integrated moat.

Previously, Microsoft built its presence within the enterprise IT procurement channel around a vertically integrated stack combining Azure Cloud, GitHub, M365, Teams, and an OpenAI partnership, and that created a high barrier to entry for independent AI startups. Up to now, the flagship example of that strategy had been GitHub Copilot. GitHub Copilot reached a 90% penetration rate among Fortune 100 companies and 4.7 million paid subscribers, spreading through a procurement structure bundled with M365. That was the reality up to June 2026.

But after June 2026, cracks in that moat began to appear. Repeated GitHub outages, apart from the issue of whether the strengths of vertical integration remain valid, exposed the fact that the prerequisite for vertical integration is each layer’s ability to adapt to traffic patterns in the AI agent era. If that prerequisite is not met, cracks emerge, and the new rule revealed by GitHub’s case is that failure of layer adaptation creates cracks in the vertically integrated structure.

What GitHub’s repeated outages showed is that even within a vertically integrated stack, problems arise if the infrastructure layer cannot keep up with changes in usage patterns in the AI agent era. This shows that the condition for maintaining the strengths of vertical integration is continuous response to the new environmental requirements of each layer. In the end, GitHub’s case shows that vertical integration alone does not make an advantage permanent, and that if each layer cannot respond to the traffic and usage changes of the agent era, structural strengths will still crack.

The source of Microsoft’s structural problem is its relationship with OpenAI. Microsoft invested tens of billions of dollars in OpenAI and has served as a key partner for deploying OpenAI models through Azure, making model capability acquisition a core means of its AI strategy. But as OpenAI now directly targets GitHub’s repository business and turns into a competitor, Microsoft finds itself in the paradoxical position of having its most important AI partner threaten a core asset.

That said, this crack does not mean the collapse of the entire vertical stack, and Fortune 100 companies’ dependence on Microsoft is unlikely to change quickly. Instead, GitHub’s case offers an important lesson. The moat created by vertical integration is not a static structure, and in the face of new traffic patterns, new usage patterns, and new competitive dynamics created by AI agents, each layer of that moat must keep adapting. If it fails to adapt, cracks appear in the layer, and those cracks become openings for competitors.

As a leadership implication for responding to change, the first action should begin with reclassifying the organization’s software portfolio.

The starting point for strategic decision-making is the reclassification of the organization’s software portfolio, and the categories are workflow executors and data owners.

This classification is not an IT department system audit; it is something the organization’s leaders themselves must directly engage in.

The result of that classification will determine future software contract structures, outsourcing strategy decisions, and AI agent investment priorities.

The criteria are simple, and each software product only needs to be tested with 2 questions. Question 1 asks whether the main source of value creation is people carrying out internal work or whether the main source of value creation is stored data being referenced by other systems or agents.

Question 2 asks whether AI agents can directly call that software’s API and whether the same result can be achieved through direct API calls.

Software can be divided according to the answers to those 2 questions. Survival criterion 1 is that the answer to the first question is data, and survival criterion 2 is that the answer to the second question is no. Software that meets both criteria is classified as a survivor in the agent era, while software that does not becomes a candidate for replacement or role reduction.

Utila’s decision to cancel 10 SaaS products is also explained by this standard. Utila kept only HubSpot, and its logic for canceling and retaining products followed the standard above.

After classification is complete, 3 action priorities are presented. Priority 1 is to maximize negotiating leverage when renewing contracts for workflow executor software. The basis is that alternatives already exist, and vendors also know that alternatives exist. As a result, bargaining power in the market shifts toward buyers.

Priority 2 is to immediately specify agent access conditions in contracts for data owner software. The items to spell out in the contract are the API access cost for external agents, volume-based additional charges, and differential pricing versus the vendor’s own agents. These conditions need to be finalized before renewal, and if they are not, companies will face unexpected cost structures once agent deployment begins in earnest.

The starting point for AI agent investment is presented not as simple automation, but as replacing workflow executors. Sanofi offers an example of this: it reduced its use of ServiceNow and then shifted purchase order audit work to agents. Sanofi’s sequence was reduced ServiceNow usage followed by agent-based replacement of purchase order audit work.

The trap in the transition logic here is framing it only as "cost cutting." Utila cut its software budget by 50%, but the 50% reduction at Utila was a result, not a goal. The position of cost reduction comes after answering strategic questions.

The strategic questions that must come first are about how to define the location of core organizational data in an environment where agents execute workflows. Another issue is how to ensure the quality and reliability of agent outputs on top of that data.

Amid these changes, the direction for redefining the role of Korean SI is moving toward becoming an agent orchestrator. Over the past 30 years, the core capability of large Korean SI companies has been heterogeneous system integration, with examples of integration targets including SAP, Oracle, Salesforce, and in-house legacy systems. The reason for SI’s existence was to connect multiple systems into a single operating environment and to design enterprise business processes to operate on that connected environment. The foundation of this capability has been decades of field experience, relationships with large enterprise customers, and labor-based project execution power.

The capabilities needed in the AI agent era continue from that foundation in the sense of being an extension, but the direction is fundamentally different. The central task is not system integration, but agent orchestration. This means supporting agents so they can safely read and write system data and support autonomous decision-making by agents. To do this, architecture design must determine agent access scopes and permissions by data set, design how agent judgment results are verified and recorded, and design how conflicts and errors are managed when multiple agents collaborate. This is the role of the agent orchestrator.

This shift in role is both a threat and an opportunity for Korean SI. The threat lies in the structural pressure on project billing models based on workforce size. If AI agents accelerate the replacement of repetitive work, bargaining power for service fees that scale with headcount weakens. The global move to replace Indian outsourcing vendors with agents may apply in the same way to Korea’s SI industry. At the same time, this change may create an opportunity for Korean SI to transition into a new orchestration role.

Korean SI companies have decades-long relationships with large enterprise customers, and they have accumulated deep understanding of customer systems. That relationship and understanding are unique assets in the transition to an agent orchestrator role. General-purpose agent platforms from global AI companies lack understanding of specific industry regulatory environments, lack understanding of legacy system complexity, and lack understanding of customer organizations’ decision-making structures. The core moat of Korean SI is that understanding.

Based on those strengths, the transition direction converges on 3 areas. First is internalizing agent architecture design capabilities. The scope of agent design includes deciding which tasks to delegate, setting criteria for granting data access rights, and designing how to audit and verify agent results. That design capability becomes a new core service that can be commercialized.

Second is building domain-specific agents. Target areas include manufacturing, finance, healthcare, and the public sector, where regulation and domain knowledge matter greatly. General-purpose AI agents have difficulty handling the complex workflows in these areas. Accordingly, in implementing specialized agents, SI’s domain expertise works directly as a competitive advantage.

Third, agent governance management was presented as a task. To that end, it was argued that an enterprise-wide governance layer is needed to centrally manage the systems and permissions that AI agents access. The entity responsible for building that governance layer was presented as SI, either directly or in support.

The article then posed a final question for business leaders and strategy planners. The first question is whether any of today’s subscription software products will still deliver the same role at the same cost 3 years from now. The second question is whether any of today’s outsourcing tasks will still be performed by humans 3 years from now.

It said that honest answers to those 2 questions are the first step in formulating a strategy for an era in which only workflow owners survive.

Source: IT DAILY · Kim Young-wook
Original: https://www.itdaily.kr/news/articleView.html?idxno=241280

References

• https://www.itdaily.kr/news/articleView.html?idxno=241280