AI agents are moving quickly from experimental tools into the machinery of everyday enterprise software, but their value will depend less on autonomy than on the quality of the workflows they are asked to operate. Far from repairing weak processes, they are more likely to expose them.
That is the central argument emerging from a growing wave of vendor announcements and industry commentary. IBM says its watsonx Agents are designed to automate complex workflows across enterprise ...
Continue Reading This Article
Enjoy this article as well as all of our content, including reports, news, tips and more.
By registering or signing into your SRM Today account, you agree to SRM Today's Terms of Use and consent to the processing of your personal information as described in our Privacy Policy.
applications, while SAP is positioning Joule Agents around trusted business data and end-to-end process reasoning. Microsoft is embedding agents into Microsoft 365 Copilot for tasks ranging from scheduling to analysis, and PwC’s Agent OS is pitched as a way to orchestrate and govern agents across different systems and vendors. OpenAI, meanwhile, has introduced Workspace Agents for ChatGPT users in business and education settings, reflecting how quickly the market is shifting from simple prompting towards workflow execution.
Yet the underlying promise of these products comes with a condition: AI agents need context. They perform best when they sit inside structured business processes with clear source data, defined stages, permissions, approval steps, output destinations and an audit trail. Without those foundations, an agent may still produce an answer or complete a task, but the result can be inconsistent, incomplete or simply wrong in a business sense.
That is why broken workflows are becoming such a visible test case for agentic AI. Disconnected systems, duplicated records, unclear ownership and manual hand-offs all reduce the quality of the information an agent can use. If pricing sits in one platform, customer notes in email and approval status in another, an agent may assemble something that looks polished while missing the details that determine whether the work is actually usable. In that sense, the technology does not fix fragmentation; it reveals it.
Process intelligence specialists have made a similar point. QPR has argued that AI agents cannot navigate broken processes effectively because they lack visibility into how work actually moves across teams and systems. That diagnosis matters because many organisations are not starting from a clean slate. They are trying to automate processes that were already messy before AI entered the picture.
This is why the idea of an AI-ready workflow has become so important. Such a workflow is one in which the input data is standardised and connected, the current stage of work is obvious, permissions are clearly defined, approvals are built into the process and the final output has somewhere unambiguous to go. Human review remains part of the loop, especially where risk is involved, and every action should be traceable for governance and compliance.
The stakes are especially high in technical industries. Construction, manufacturing, field service and engineering-led businesses rely on tightly linked documents, schedules, revisions, permits, service records and approvals. An isolated prompt is rarely enough to support that level of complexity. AV firms offer a useful example: design files, bills of materials, proposals, installation details and service records all need to line up across the project lifecycle. XTEN-AV is one of the companies illustrating how connected workflow context becomes a prerequisite for reliable AI support in such environments.
Analysts expect the adoption curve to accelerate. Gartner has forecast that a large share of enterprise applications will incorporate task-specific AI agents by 2026, signalling a shift away from prompt-based experimentation and towards embedded automation within business systems. The same firm has also warned that more than 40% of agentic AI projects could be cancelled by 2027 because of rising costs, unclear business value, weak governance or inadequate risk controls. Together, those forecasts suggest that enthusiasm alone will not determine success.
The practical lesson for companies is simple: before deploying an agent, they need to examine the workflow. That means mapping the process from start to finish, checking whether the source data is reliable, identifying who owns each step, deciding where human approval is required and confirming that the necessary systems can talk to each other. It also means setting measurable goals, such as shorter turnaround times, fewer errors or less manual re-entry.
The most effective AI agents will not be the ones that pretend to operate in a vacuum. They will be the ones that work inside disciplined, well-governed processes. In the enterprise, automation is not replacing workflow design. It is making the quality of that design impossible to ignore.
Source: Noah Wire Services
