There’s an uncomfortable truth many companies are discovering as they try to deploy AI agents: the problem isn’t the models. It’s what sits underneath them.
This week, Salesforce launched Agentforce Operations, a new orchestration layer that turns back-office processes into structured tasks for specialized agents to execute. The announcement has gone relatively unnoticed outside technical circles, but it points to something every business owner should understand before investing a euro in “AI for the enterprise.”
Let’s break it down.
What Salesforce has actually launched
Agentforce Operations is what the industry is starting to call a workflow execution control plane. In plain English: a layer that imposes deterministic structure on the processes AI agents are expected to run.
The idea is simple:
- You upload your process (or use one of the predefined Blueprints).
- The platform breaks it into explicit steps.
- It assigns each step to the appropriate specialized agent.
- The system — not the agents — decides what happens next.
That last part is the key. Unlike many automation tools that let agents probabilistically decide what to do, here the sequence is pre-defined. Agents execute; they don’t improvise.
The real problem: your processes weren’t designed for machines
Sanjna Parulekar, Salesforce’s VP of Product, put it in a line worth reading twice:
“What we’ve seen with customers is that often, what’s broken is in your requirements document. When it’s uploaded into a product, it just doesn’t work.”
Here’s the trap almost nobody sees coming.
Business processes — the ones that have been running for years — evolved around human judgment. They contain fuzzy steps, implicit decisions, coordination that depends on Juan knowing that after X it usually means calling María. They work because people fill in the gaps with common sense.
AI agents, no matter how smart, don’t fill in gaps. They follow instructions. If the instructions are ambiguous, the agent fails, escalates, or worse: makes a confident but wrong decision.
That’s what is breaking AI deployments in many companies. It’s not that the models can’t reason well. It’s that the workflows underneath them were never coherent in the first place.
Why this matters even if you don’t use Salesforce
Most SMBs and mid-sized companies are not going to deploy Agentforce. The price and complexity make it impractical for anyone without a sizable IT department. But the principle behind Agentforce Operations is universal and applies to any AI automation project, big or small.
The lesson, in one line: coding a broken process doesn’t fix it — it scales it.
If you automate a workflow that already had unclear steps, now you have confusion at machine speed. If you chain one agent after another without anyone knowing what outcome is being pursued, you multiply the failure points.
Brandon Metcalf, CEO of Asymbl, summed it up well in the same piece:
“You have to understand the goal, or the agent or the human won’t complete the task successfully. Someone has to manage that outcome.”
What you should review before automating anything
Before putting AI agents into any process, it’s worth asking these questions:
| Question | Why it matters |
|---|---|
| Is the process documented step by step? | If it isn’t, the agent won’t know what to do |
| Which decisions require human judgment? | They need to be explicitly isolated |
| What is the expected final outcome? | Without a clear goal, success can’t be measured |
| Who validates that the process is still working? | You need an owner, not just a system |
| What happens when something fails? | There must be a human escalation path |
If you can’t answer these clearly, the problem isn’t the tool you choose. The problem is the process.
The move many people are skipping
The dominant AI-for-business narrative says: “buy the most powerful model, give it access to your data, and let the magic happen.” The reality companies are seeing when they truly deploy agents in production looks very different.
What works looks more like this:
- First, map the real process (not the one you think you have).
- Then, identify which steps are deterministic and which require judgment.
- Next, simplify and remove the steps that only existed out of inertia.
- Only then, introduce agents for the specific tasks they can execute reliably.
- Always, keep observability and a human accountable for the outcome.
It’s less sexy than “deploy an autonomous agent.” It’s also what separates a project that saves time from one that wastes it.
How we approach it at Studio SmartWork
We’ve been building AI automations for SMBs and mid-sized companies since 2021, and we’ve seen the same pattern again and again: the client asks for “a bot that does X,” and the first thing we discover is that X, as it’s currently defined, isn’t something that can be automated reliably.
Our process starts with a quick audit of the current workflow. Not because we’re obsessed with documentation, but because we know what happens when you skip that step: you build something that works in demos and breaks in production.
We then build on proven open-source tools like n8n, which lets us keep a deterministic orchestration layer — similar in philosophy to what Agentforce Operations is proposing, but at SMB scale and without locking the client into a vendor. AI agents handle the tasks that require language or bounded judgment; the system decides the flow.
Results we’ve seen repeat themselves:
- Lead follow-up: response time from hours to under 60 seconds, +35% response rate.
- Robust workflows: 0 unrecovered failures in 6 months of operation.
- Inbox handling: from 3 hours to 15 minutes a day.
It’s not magic. It’s taking the time to understand the process before automating it.
The bottom line
The launch of Agentforce Operations is interesting less for what it offers than for what it validates: the industry is officially acknowledging that the bottleneck in enterprise AI is not intelligence, it’s process coherence.
For business leaders, that translates into a simple strategic decision. Before you invest in “AI,” spend a week looking at your processes honestly. Ask which steps exist because they make sense and which exist because “that’s how it’s always been done.” Define the outcome you’re aiming for. Assign an owner.
Then — and only then — let’s talk about agents.
The technology for automating repetitive work has been good enough for a couple of years now. What’s missing, almost always, is clarity about what exactly should be automated. And fortunately, that doesn’t require buying anything — just sitting down and thinking it through properly.