Salesforce admits it: your workflow was built for humans, not agents

Salesforce shipped a product this spring whose entire pitch is a confession. It's called Agentforce Operations, and it exists to fix the workflows that keep breaking enterprise AI. The framing in their own materials is unusually blunt:

Enterprise AI teams are hitting a wall — not because their models can't reason, but because the workflows underneath them were never built for agents.

Read that twice, because a vendor selling agents just told you the agent is not the problem. The model is fine. The thing breaking is the process you handed it. That's the most useful sentence anyone has said about enterprise AI this year, and it's worth understanding exactly why it's true.

Humans run on workflows full of holes

Here's the part we never notice. Almost every business process you've ever followed is quietly broken, and it works anyway — because a human is patching the holes in real time without telling anyone.

The steps are loosely defined. Half the decisions are implicit. The handoff between two teams "just works" because Maria in accounts receivable knows to check the second spreadsheet on Fridays, and has known for nine years. As one write-up put it, these are processes that evolved through years of workarounds — loosely defined steps, implicit decisions, coordination that depends on individuals knowing what to do next. They were designed around human judgment gaps, not machine execution.

A person treats the written process as a rough suggestion and fills the rest from experience. That improvisation is invisible, unpaid, and absolutely load-bearing.

An agent does exactly what you wrote down

Now drop an agent into that same process. The agent does the one thing the human never does: it follows the instructions literally. It doesn't know about Maria's Friday spreadsheet. It doesn't sense that "get approval" means "Slack Dana, but only if the amount is over five grand." It hits the first undefined step and either stops or confidently does the wrong thing.

This is why so many pilots stall, and it has nothing to do with intelligence. A smarter model follows your broken process more precisely off the cliff. The agent isn't failing the workflow; the workflow is failing the agent, by assuming a human will be there to improvise — and now there isn't one.

The fix is the work nobody wants

The unglamorous truth is that making a process agent-ready means actually defining it — writing down the implicit decisions, encoding the rules Maria keeps in her head, drawing the line between what's fixed and what's judgment. The emerging consensus on how is the same thing I keep arguing for: you blend deterministic steps — rules, APIs, system checks — with agent reasoning only where it adds value. The approval thresholds, the escalation triggers, the compliance gates become hard, deterministic rules. The agent reasons in the gaps between them, not over the top of them.

That's the same point as asking the right first question: not "which part do we agentize," but "what's the deterministic source of truth here, and where exactly does judgment belong." You're not teaching the agent your process. You're finally writing your process down well enough that anything — human or machine — could follow it. The agent just won't forgive you for skipping it the way a person would.

Be skeptical of the miracle numbers

One caution, since this is a sales pitch wearing a diagnosis. The same launch waves around figures like 70% faster cycle times and 80% of manual chores eliminated, and surveys claim 79% of enterprises already "run agents." Treat those the way you'd treat any number a vendor needs to be true. "Running agents" in a pilot is not the same as agents doing real work unsupervised, and the 80% that gets automated is always the clean, well-defined 80% — the part that was already close to agent-ready. The messy 20%, the part where Maria's judgment lived, is exactly the part that doesn't automate, and it's where the hard redesign work actually is.

The takeaway

You don't need to buy Agentforce to use the lesson, and it scales all the way down to a solo project. Before you blame a model for failing at your task, go read the process you gave it and ask: would a brand-new employee, with zero context and no one to ask, get this right by following these instructions exactly? If the answer is no, the agent was never your problem.

The companies that win with agents over the next year won't be the ones with the smartest model. Everyone has that. They'll be the ones who did the boring work of turning their human-shaped, gap-filled processes into something precise enough for a machine to run — and discovered, along the way, that they understood their own operations a lot less well than they thought.