These days, everything with an LLM inside gets called an “agent.” A chatbot? Agent. A script that summarizes emails? Agent. A five-step process that classifies tickets and routes them to the right team? Also an agent, apparently.
Part of this is marketing. “Agent” sells, so vendors slap the label on anything that touches AI. But part of it is that most of us using AI at work haven’t stopped to draw the line. And I think we should, because the difference between an agent and an agentic workflow is exactly where most of the real business value is hiding.
A simple test: who decides the next step?
Here’s the easiest way I’ve found to tell them apart:
- Agent: the model decides what to do next. You give it a goal and some tools, and it figures out the path on its own: what to look up, what to call, when it’s done.
- Agentic workflow: the process decides what to do next. The steps are defined up front, and the model does its work inside specific steps: extracting data, classifying, drafting, deciding between a few options.
Same technology, very different behavior.
It’s the same line Anthropic draws in Building Effective Agents: workflows orchestrate LLMs through predefined code paths, while agents direct their own process and tool use.
Determinism: the part nobody talks about
A system is deterministic when the same input always produces the same output, following the same path. A calculator is deterministic. Your payroll system had better be deterministic.
Interactive agents are, by nature, non-deterministic. Give an agent the same task twice and it may take a different route each time: different tools, different order, different result. That’s not a bug; it’s what makes agents great at open-ended work. But it also means they can wander off, skip steps, or get stuck in loops, retrying the same thing over and over.
Agentic workflows strike a better balance. The overall structure of the process is fixed and predictable. The non-determinism lives only in the specific points where an agent steps in. You get the flexibility of AI where you need it, and consistency everywhere else. No loops, no surprise detours, no “why did it do that this time?”
Same task, two ways
Take something every company does: processing supplier invoices.
The agent way: you give an agent access to the inbox and the ERP and tell it, “Process incoming invoices.” Some days it checks the purchase order first; some days it doesn’t. Sometimes it asks for clarification, sometimes it guesses. When a field is missing, it might try three different approaches, or the same one ten times.
The agentic workflow way:
- A new invoice arrives (trigger).
- An AI step extracts the vendor, amounts, dates and line items.
- Regular code matches it against the purchase order.
- If everything matches, it’s posted to the ERP.
- If not, an AI step drafts a short explanation of the mismatch and routes it to a person for approval.
Only steps 2 and 5 use AI. Everything else is plain, boring, reliable software. And that’s the point.
Traceability: opening the black box
This is where the difference really hits home for enterprises.
With an agent, every run is basically a black box. You can log which tools it called, but the path changes every time, so getting consistent, comparable data out of it is hard. Why did it approve this invoice and flag that one? Which step failed? How long does each part take? Good luck building a dashboard on top of that.
With an agentic workflow, every step is a known checkpoint. You know what went in, what came out, how long it took and what it cost. You can audit it, monitor it, measure it, and plug in human approvals exactly where they’re needed. For regulated industries (finance, healthcare, government), that’s often not a nice-to-have. It’s the deciding factor.
Costs you can actually predict
There’s a money side to this too. An agent decides how many steps a task needs, so the same job might take five model calls one day and fifty the next. Cost and response time per task vary, which makes ROI hard to forecast.
In a workflow, you know how many AI steps each run has and roughly what each one costs. You can budget it, measure the savings and show your CFO a real number. And since most steps are plain code, most of the process doesn’t need a model at all.
So, where do agents fit?
Don’t get me wrong: agents are great. For one-off, exploratory or knowledge-heavy tasks (researching a topic, analyzing a document, writing code, preparing a proposal), an agent working alongside a person is a huge productivity boost.
But when it comes to the repetitive processes that run your business every day, the ones behind most of the efficiency gains and cost savings, agentic workflows do the heavy lifting.
It’s not either/or
Agents and workflows aren’t competing ideas. An agent can be one step inside a workflow: say, a step that researches a new supplier before a person approves the contract. The process stays in control, and the agent handles the part that needs open-ended reasoning.
A good rule of thumb: start with a workflow, and add autonomy only where it pays off. If a step keeps needing flexibility the process can’t give it, that’s your candidate for an agent.
The takeaway for leaders
If you’re figuring out where AI can make a real dent in your organization, start by mapping your repetitive processes. That’s where the savings are. Turn them into agentic workflows, with AI only in the steps that need judgment. Then give your teams agents for the specific, unpredictable work where flexibility beats consistency.
And next time someone pitches you an “agent,” ask one question: who decides the next step?
