Enterprise AI teams are hitting a wall – not because their models can’t reason, but because the workflow beneath them was never built for agents. Tasks fail, handoffs break down, and the problem grows as organizations push agents deeper into back-office systems. A new architectural layer is emerging to address this: workflow execution control planes that impose deterministic structure on the process agents they are expected to run.
One of the companies bringing this to the forefront is Salesforce, with a new workflow platform that turns back-office workflows into a set of tasks for specialized agents. Users can upload their own processes or use one of the set blueprints provided by Salesforce, and AgentForce Operations will break it down for agents.
The problem is that many enterprise workflows aren’t built for agents, Sanjana Parulekar, senior vice president of product at Salesforce, told VentureBeat in an interview. “What we’ve seen with customers is that oftentimes the fault in a process is probably in your product requirements document,” Parulekar said. “So when it’s uploaded to a product, it doesn’t quite work. We can customize it and cut out some things and replace it with an agent.”
Without this control panel layer, enterprises can risk deploying agents that increase costs rather than fixing their workflow problems.
Making workflows work not just for humans, but also for agents
Enterprises deploying agents are learning an expensive lesson: Their workflows were designed around human decision intervals rather than machine execution. Processes that evolved through years of workarounds – loosely defined steps, implicit decisions, coordination that depends on individuals knowing what to do next – break down when agents are asked to follow them verbatim.
Even with all the context of an enterprise at its fingertips, an AI system will have difficulty completing tasks if it is not clear what it is supposed to do.
Parulekar said his team found that focusing on what drives the process and breaking it down into more clear steps and workflows makes the system more deterministic. Then, when platforms like Agentforce Operations introduce agents, those agents already know their specific tasks.
“It forces companies to rethink their processes and brings observability into the mix because of the session tracing model in the system,” she said.
Parulekar said human checks can be built into the system, so the process is more transparent.
What makes this approach different from other workflow automation offerings is that it doesn’t rely on agents to decide what to do next; The system does. Unlike more traditional automation tools, which route tasks and agents to probabilistic decision making, it enforces execution on a more pre-defined, deterministic structure.
The problem it presents
Codifying a workflow doesn’t fix a broken workflow. If a process has erroneous steps, encoding it for agents largely fixes the problem. And once the workflow is distributed among agents, the challenge shifts from execution to governance: who owns the process, who validates it, and how it evolves as business conditions change.
This puts the onus on teams to keep a close eye on what works for them and what doesn’t.
Organizations need to consider that, with the execution control plane offered by platforms like AgentForce Operations, someone must be made accountable for task completion and success.
Brandon Metcalfe, founder and CEO of workforce orchestration company Assembl, told VentureBeat in a separate interview that the key is a shared goal for both the humans and agents following the workflow.
“You have to understand the goal otherwise the agent or human will not complete the task successfully,” Metcalf said. “Someone has to manage the result that has to be delivered. This could be a person or an agent.”
The obstacle has been removed. As Metcalf formulated it, the question is no longer whether agents can reason through a task, but whether the workflow beneath them is coherent enough to execute. For enterprises that have built their processes around human judgment and institutional memory, this is a harder solution than swapping in a smarter model.
<a href