Most things sold to PMs as AI agents are not agents. They're automations or workflows in a trench coat. That's not a naming nitpick. If you're building product around AI, the distinction changes your architecture, your pricing, your reliability story, and what you can honestly promise a customer.
Start with the three, because from the outside they look identical.
Automations follow rules. If X, do Y. No judgment, no variation. An email drip that fires on day 3 after signup is an automation. Fast, reliable, and it'll do exactly the wrong thing the moment conditions change and nobody updates the rule.
AI workflows add a language model to a pipeline that's still deterministic. A support ticket comes in, an LLM classifies it, routes it, maybe drafts a reply. There's intelligence at specific steps, but the flow is fixed. Step 1 always leads to step 2. The AI makes the steps smarter. It does not decide which steps to take.
Agents decide what to do next based on context. They change approach when something surprises them, compose different skills for different situations, and produce variable outputs. That non-determinism is the whole point, and the whole risk.
So how do you tell what you're actually looking at? Ask four questions about any "agent" someone hands you. Does it decide what to do, or only how to do a predefined thing? Can it handle a situation it wasn't designed for? Does it learn from feedback? Is the output predictable across runs? That last one has a cheap test. Run it ten times on the same input. Same result every time, and you've got an automation or a tightly constrained workflow. Not an agent.
I run a fleet of agents in my own PM practice, and honestly, not all of them are agents. My daily focus agent is closer to a workflow. It pulls from fixed sources, runs a prompt, hands me a summary, and the sequence never changes. My competitive intel agent is the real thing. It decides which competitors to dig into based on what changed, picks its own sources, adjusts depth as it goes. It surprises me sometimes. That's the tell.
Why does any of this matter for the work? Reliability expectations differ. Automations should work 100% of the time. Workflows should be highly reliable with the occasional AI misstep. Agents are less predictable and need guardrails and human checkpoints. Pricing follows the same split: automations fit per-action, workflows fit per-task or per-seat, and truly autonomous agents should be priced on outcomes.
Here's the part people skip. Most of the real value in enterprise AI right now lives in the workflow tier. Agents get the headlines. Well-built workflows get the work done. Pick one "agent" on your roadmap this week and run it through the four questions. If it fails one, call it what it is, and build it that way.