Enterprise AI teams are discovering that their most powerful agents are only as effective as the workflows they follow. When processes are vague, rely on human intuition, or contain hidden workarounds, agents fail — tasks stall, handoffs break, and costs soar. Salesforce has introduced Agentforce Operations, a new workflow platform that imposes deterministic structure on back-office processes, breaking them into explicit tasks for specialized agents. Below, we explore how this approach works, why it matters, and what challenges remain.
Why do enterprise AI agents often fail in back-office workflows?
Most enterprise workflows were never designed for machine execution. They evolved through years of human workarounds: loosely defined steps, implicit decision points, and coordination that relies on individuals knowing what to do next. When AI agents try to follow these processes literally, they hit a wall. Tasks break because the logic is fuzzy, handoffs fail because roles are ambiguous, and the problem compounds as more agents are deployed. Sanjna Parulekar, Salesforce SVP of Product, told VentureBeat that the brokenness often originates in the product requirements document itself. Upload that into a workflow system, and it doesn't work. The result is an agent that increases cost instead of solving problems.
What is Agentforce Operations and how does it work?
Agentforce Operations is a workflow execution control plane that turns back-office processes into a clear set of tasks for specialized agents. Users can upload their existing process documentation or choose from predefined Blueprints provided by Salesforce. The platform then breaks the workflow down into explicit, deterministic steps. Unlike traditional automation that relies on agents to decide what to do next, this system enforces a structured sequence. Parulekar explained that focusing on what makes the process tick and breaking it into explicit steps makes the system more deterministic. Agents then know their specific tasks ahead of time, reducing failure points and enabling smooth handoffs.
How does this approach differ from traditional workflow automation?
Traditional workflow automation tools often route tasks based on probabilistic decision-making — agents guess the next best step. This works well for simple, repetitive processes but fails when complexity grows. Agentforce Operations takes a different approach: it uses a deterministic structure where the system, not the agent, decides the next action. Every step is predefined and enforced through a control plane. This eliminates ambiguity and ensures that agents execute tasks in the correct order. It also introduces observability through session tracing, so teams can monitor exactly how a process flows. Human checkpoints can be embedded, making the entire workflow transparent and auditable.
What role does session tracing play in agent workflows?
Session tracing is a key feature of Agentforce Operations that records every step an agent takes within a workflow. Parulekar highlighted that this forces companies to rethink their processes and brings much-needed observability to AI-driven automation. With session traces, teams can see where tasks succeed, where they stall, and where human intervention is needed. It also helps identify hidden bottlenecks or flawed steps that were previously invisible. By tracing each agent's path, organizations can continuously refine their workflows. The data from tracing feeds back into process redesign, making the system smarter over time while maintaining deterministic control.
What risk does codifying workflows introduce, and how can it be mitigated?
A major risk is that encoding a broken workflow locks in problems at scale. If a process has flawed steps or illogical sequences, making it deterministic only amplifies those issues across all agents. Parulekar cautioned that simply uploading a process doesn't fix it — the platform can optimize and cut out broken parts, but human oversight remains essential. To mitigate this, Agentforce Operations allows human checks to be built into the sequence. Teams can review and adjust the workflow before full deployment. The session tracing also helps catch errors early. The lesson: codify only after validating the process logic, and treat workflow automation as an iterative, not one-time, exercise.
Who is Sanjna Parulekar and what insights did she share?
Sanjna Parulekar is Salesforce's Senior Vice President of Product. In an interview with VentureBeat, she provided deep insights into why enterprise workflows fail with AI agents. She explained that the root cause often lies in the product requirements document itself — a document written for humans that assumes implicit knowledge. Her team found that by breaking processes into explicit, deterministic steps, they could dramatically improve agent success rates. She also emphasized the importance of observability through session tracing and the need for human-in-the-loop checkpoints. Parulekar's comments underscore a shift from probabilistic automation to structured, controllable agent execution.
What does the future hold for workflow-driven AI agents?
The emergence of platforms like Agentforce Operations signals a new architectural layer for enterprise AI: the workflow execution control plane. As organizations push agents deeper into back-office systems, the need for deterministic structure will only grow. Future developments may include more sophisticated blueprints, adaptive workflows that learn from tracing data, and tighter integration with human decision points. The key challenge remains balancing structure with flexibility — workflows must be explicit enough for agents to follow but not so rigid that they break when exceptions occur. Companies that master this balance will unlock the full potential of AI automation without the cost of broken processes.