The Next Vendor Lock-in May Be Agentic

There is an obvious reason organizations are moving toward agentic AI. It promises to make work move faster across environments that were never designed to move as one.

Most companies are not short of software. They have systems for customers, contracts, documents, finance, planning, support, reporting, compliance and operations. Each system does something useful. Each also carries its own logic: a way of structuring data, permissions, decisions, workflows and user behavior.

For years, the user has been the bridge between those systems. People moved from screen to screen, copied context, chased approvals, interpreted exceptions, translated information and made the practical connections the stack itself did not always make.

In this context, agentic AI refers to AI systems that can interpret context, use tools, retrieve information, trigger actions and support the movement of work across systems. Agentic AI changes that pattern. It does not simply add another interface. It introduces a new layer through which work can be interpreted, routed and executed across the stack.

When does speed really become an advantage?

Speed is one of the main reasons agentic AI is attracting so much attention. Faster response can mean more leads. Faster analysis can mean better decisions. Faster routing can mean lower operational drag. Faster execution can mean more relevance, more attention and more revenue. In a market where delay is increasingly visible, it is easy to understand why organizations want work to move faster.

Speed only becomes an advantage when the organization can absorb it.

When workflows, ownership, governance and decision paths are clear, speed can shorten commercial cycles and reduce operational drag. When speed enters an environment that is already fragmented, it can create more exceptions, unclear ownership and movement that is difficult to trace afterwards.

Many organizations know this pattern. A solution is bought to create speed. It delivers movement in one area, then creates new dependencies around implementation, integration, ownership or control. Another solution is added to manage the complexity. Over time, the stack becomes faster in parts and slower as a whole.

The cost of uncontrolled speed is rarely just the tool. It is the rework, fragmentation and loss of confidence that follow when speed is added without operating control.

That loss of confidence matters. Once an organization has gone through enough half-absorbed implementations, every new change proposal becomes harder to trust. Change slows down not because the need disappears, but because confidence in the change process has been damaged.

When AI becomes the work layer

Anthropic's move into legal is a useful signal. Reuters reported that Anthropic expanded Claude's tools for lawyers with secure connections to third-party platforms including Thomson Reuters for legal research, plus integrations with Box, Everlaw and DocuSign, and 12 legal practice plug-ins. [1]

This is not only relevant for the legal sector. It shows a broader movement: AI is beginning to sit above existing systems as the place where work is interpreted, routed and executed. The workflow no longer has to start inside the SaaS interface. It can start inside an AI-mediated work layer.

That shift changes what becomes valuable in software. When work starts in an AI layer, the SaaS interface is no longer always the primary surface of work. The API, data model, permissions structure and connector layer become more important, because they determine whether the product can be used by agents at all.

For SaaS companies, this creates pressure to become more headless and agent-readable. For enterprise buyers, it raises a different question. APIs create access to system capability, and access is not the same as control. The more work moves through APIs and connectors, the more important it becomes to govern the logic that decides what should happen, when, by whom and under whose authority.

The API is not the operating model. It is an access point into the system.

Agentic AI enters organizations shaped by years of systems, process decisions, workarounds, integrations, reporting gaps and ownership boundaries. These organizations are often highly capable. Teams are delivering. Systems are working. The more subtle issue is that the operating environment can be functional enough to continue, while no longer being coherent enough to absorb more complexity cleanly. That is also the core of Ongo's existing fragmentation thesis. [2]

Agentic AI can help this environment move faster. But it can also make the environment harder to understand. A user may experience less friction. The organization may inherit more hidden movement between systems, permissions, models, connectors, context windows, approval paths and decision records.

The new dependency is less visible

A system that assists a user is one thing. A system that interprets work, selects context, chooses routes, retrieves data, triggers next steps and coordinates across applications begins to sit closer to the operating model.

That is why the larger advisory market is starting to treat agentic AI as more than a tooling issue. McKinsey describes the agentic organization as one in which humans work with virtual and physical AI agents to create value, with implications for operating models, business models, governance and workforce design. [3] Deloitte makes a related point from the workforce side: agentic AI is beginning to reshape workforce operating models before organizational design, controls and talent models have fully shifted, and leaders have a critical window to establish guardrails before multi-agent systems scale. [4]

The old vendor-dependency conversation was easier to recognize. It often lived inside an application: the CRM, the ERP, the contract platform, the document environment, the workflow system. The organization knew which system held which part of the operation, even when the total environment was fragmented.

Agentic AI makes that picture less tidy. The new layer may not own the record directly. It may own the route. It may decide which systems are consulted, which context is used, which action is taken next, which exception is escalated and which decision path becomes normal.

The next dependency may not come from the system where work is stored. It may come from the layer where work is interpreted, routed and executed.

This goes deeper than AI vendor lock-in. AI vendor lock-in is about dependency on a model or platform. Agentic lock-in is about dependency on the operating layer that determines how intelligence moves work.

That is a different kind of dependency. Less visible, because it does not always sit in the database. More consequential, because it begins to shape how work moves. Once a layer shapes how work moves, it also begins to shape how the organization learns, improves and makes decisions.

Workflow intelligence will accumulate somewhere

Without governed orchestration, companies may reduce interface friction while creating a new layer of operational dependency. They may move faster and understand less about how the work actually moved. They may gain automation and lose clarity around who owns the logic of that automation.

This is where agentic lock-in becomes a serious business issue. Not as a dramatic warning. As an operating consequence.

Gartner has pointed to the same tension from a different angle, predicting that more than 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. Gartner also notes that integrating agents into legacy systems can be technically complex and may disrupt workflows or require costly modifications. [5]

Agentic AI clearly has value. Speed alone, however, does not create an operating advantage. The deeper issue is whether organizations can use agentic AI without allowing workflow intelligence to drift into places they do not fully govern.

Workflow intelligence is the practical operating knowledge of how work actually moves: which context is needed, which systems are consulted, which decisions require human judgment, which actions can be automated, how exceptions are handled and how outcomes are recorded.

Once AI begins to support, route or execute work, that intelligence will accumulate somewhere. It may sit in the SaaS vendor's agent layer, an AI platform, connector logic, local team configurations, prompts, permissions or undocumented operational habits. Or it may be governed deliberately inside a layer the organization controls.

Ongo's earlier operating logic thesis asked whether existing systems only support operations, or whether they also begin to define the boundaries of what an organization believes is possible. [6] Agentic AI brings that question into a new phase. If AI becomes the layer through which work is performed, the company needs to understand whether it still owns the logic by which work moves.

Useful layers become structural layers. Structural layers deserve governance.

The layer between systems matters

A customer-controlled orchestration layer is useful only if it increases clarity. Its purpose is not to replace organizational judgment with automation. Its purpose is to create a governed place where AI, systems, workflows and human authority can work together without losing visibility or control.

That means the organization can shape where agents belong, which systems they may access, which actions require human approval, how context is passed, how decisions are logged, how exceptions are escalated and how workflow logic remains portable as vendors and models change.

For leadership teams, this is not a technical detail. It affects how the company preserves accountability, controls risk, absorbs speed, protects future optionality and avoids building another layer of dependency into the operating model. Once AI begins to interpret, route and execute work across systems, the question becomes less about which tool is used and more about who governs the way work moves.

The companies that benefit most from agentic AI will not simply be the ones that deploy it fastest. They will be the ones that understand where speed needs structure, where human authority should remain visible, and where workflow intelligence should live as work begins to move across the stack.

Ongo sees this as the emerging role of a customer-controlled orchestration layer: not a replacement for existing systems or organizational judgment, but a governed structure above and between systems where AI-enabled workflows can be designed, observed, adjusted and improved without surrendering operating logic to a single vendor, model or local configuration.

Enterprise work is moving from SaaS interfaces to AI-mediated execution. That creates a new strategic control problem: agentic lock-in.

The next phase of enterprise software will not only be shaped by the best model or the largest SaaS suite. It will also depend on the layer that governs how AI, systems, workflows and human authority move together.

For organizations, that layer is too important to leave undefined.

Sources

• [1] Reuters, 'Anthropic expands Claude's AI tools for law firms, lawyers', May 2026.
• [2] Ongo, 'Why Fragmentation Makes Change Hard'.
• [3] McKinsey, 'The agentic organization: Contours of the next paradigm for the AI era'.
• [4] Deloitte, 'Operating models for humans and AI agents'.
• [5] Gartner, 'Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027', June 2025.
• [6] Ongo, 'Who Owns the Operating Logic?'.

Share