AI agents
An AI agent with no workflow behind it is a demo
The demo works because a human is standing next to it. Production is the condition where nobody is.
3 min readNazmul Hasan
I have watched a lot of AI pilots die at the same moment, and it is never the moment people expect. It is not when the model gets something wrong. It is the first time it gets something wrong and nobody notices for four days.
That gap is not a model problem. It is a missing process.
Workflow, agents, tools
At PodFleet we break every operational problem into three parts, in a fixed order, and the order is the whole point.
- Workflow: the repeatable, documented process. Who does what, in what order, and what counts as done.
- Agents: who executes it. Some are people, some are AI, and the distinction matters far less than most vendors want it to.
- Tools: the systems and integrations underneath.
Almost everyone starts at tools, picks an agent second, and never writes the workflow at all. Then they wonder why the pilot did not survive contact with a normal week.
What a workflow gives you that a prompt does not
A documented workflow answers the questions an agent cannot answer about itself. What is the acceptable failure rate. Who reviews the edge cases. What happens on the day the API is down. Who gets told when the volume triples.
None of that is AI work. All of it determines whether the AI work survives.
It also makes the agent swappable. If the workflow is written down, replacing a model, or replacing the model with a person during a bad week, is a configuration change. If it is not, the process only exists inside whatever you built it on, and you have made a vendor into a dependency.
Where agents genuinely win
I am not talking anyone out of AI. The wins are real and they cluster in a specific shape: high volume, narrow scope, tolerant of a review step.
- First-touch response, where speed matters more than nuance and a human still closes.
- Classification and routing, where being right 95 per cent of the time beats a queue nobody triages.
- Drafting, where the agent produces and a person approves.
- Extraction from messy inputs, which is genuinely tedious and genuinely mechanical.
What they lose at is anything where being wrong is expensive and being wrong is quiet. Those need a person in the loop, and the workflow is how you decide which is which.
The question to ask a vendor
Not "what model is it". Ask what happens when it fails, who finds out, and how fast. If there is no answer, you are being sold a demo, and the demo works because someone is standing next to it.
Production is the condition where nobody is standing next to it.