The Pitch That Should Never Leave Your Mouth
If your answer to every workflow problem is an AI agent, you are not an engineer, you are a salesperson. I have built and run production agents across finance back-office, plant floor operations, distribution routing and field service dispatch. The honest number is that roughly a third of the workflows I get called in to automate should not use an agent at all. A deterministic script, a structured form or a hard-coded rule will run faster, fail more predictably and cost a fraction of the inference bill.
Saying that out loud is the fastest way to earn trust with an operations or technology leader who has been burned before. It is also just true.
Stable Inputs and Stable Logic Belong in a Script
A regional distributor asked me to build an agent to generate daily pick-list summaries from their warehouse management system. The data was structured, the format never changed, the business rules were written down and had not changed in four years. An agent would have added latency, a monthly inference cost and a new failure mode where the model occasionally reorders list items or adds a sentence nobody asked for.
We wrote a script. It runs in under two seconds, costs nothing per execution and has no opinion about the output. The operations manager can read it, modify it and own it without my involvement. That is the outcome.
The rule is simple. If you can write the logic down completely, and the inputs arrive in a consistent shape, a script is the right tool. An agent earns its place when the logic cannot be fully written down or the inputs arrive in forms a script cannot parse.
Compliance Steps Are Not a Place for Probabilistic Reasoning
A mid-market lender wanted an agent to handle loan covenant monitoring. The covenants were defined in legal documents, the thresholds were exact numbers and the required action on breach was specified to the letter. There was no ambiguity to resolve, no judgment call to make.
Putting an agent in that loop would have introduced exactly the wrong property. Covenant monitoring needs determinism. A breach at 3.2x leverage is a breach. A model that reasons about it and occasionally hedges or misreads a decimal is a liability, not an asset. We built a rule engine with a structured exception queue. Auditors can trace every decision to a line of code. That matters enormously in regulated finance.
Anywhere a regulator, an auditor or a safety officer needs to reconstruct exactly why a decision was made, you want deterministic logic. An agent's reasoning trace is not the same thing as an audit trail.
Forms Beat Agents When the Goal Is Structured Capture
Field service teams in manufacturing and utilities spend real time on inspection reports. The instinct is to give technicians a voice interface or a chat agent so they can describe what they found in natural language and have the system extract structured data. I have built that. It works in some contexts.
But I have also watched it fail when a technician is standing next to a loud compressor, when connectivity is marginal, or when the extracted field values need to feed directly into a maintenance management system with strict schema requirements. In those cases a well-designed mobile form with conditional logic, mandatory fields and offline sync is faster to complete, cheaper to maintain and produces cleaner data.
The agent version adds a transcription step, an extraction step and a validation step, each of which can introduce error. The form removes those steps entirely. Fewer steps means fewer failure points. That is not a limitation of AI, it is just engineering.
High Volume Low Variance Tasks Will Drain Your Budget
A logistics operator processing several thousand shipment status updates per day wanted to use an agent to parse carrier emails and update their tracking system. The emails followed about six templates. The fields they needed were always in the same positions. Variance was low.
At that volume, running every message through an inference call is expensive and slow relative to a pattern-matched parser. We built the parser. It handles the six templates at near-zero marginal cost. We added a small classifier that routes genuinely ambiguous messages, maybe two or three percent of volume, to a lightweight agent that can handle edge cases.
That hybrid is the right architecture for high-volume low-variance work. The agent is not the default. It is the exception handler. Designing it that way cut the inference cost by roughly ninety percent and made the system faster overall.
Where Agents Actually Earn Their Place
To be clear, I am not arguing against agents. I run them in production every day. They earn their place when inputs are unstructured and variable, when the logic genuinely cannot be fully specified in advance, when the task requires synthesizing information from multiple sources in ways that change with context, or when the volume of edge cases is high enough that maintaining explicit rules becomes its own engineering burden.
A good example is exception handling in accounts payable. Invoice matching against purchase orders is straightforward rules work. But when an invoice does not match and a human would normally read the vendor's attached PDF, cross-reference a contract amendment and make a judgment call, that is where an agent adds real value. The agent is not replacing the rule. It is handling what the rule cannot reach.
The framing that works for me is this. Rules handle what you can predict. Agents handle what you cannot fully predict but can still evaluate. If you cannot evaluate the output reliably, you are not ready to automate it at all.
The Practical Takeaway
Before you scope any automation, ask three questions in order.
- Can I write down the complete logic and will the inputs arrive in a consistent structure? If yes, use a script or a rule engine.
- Is this a compliance or safety-critical step where every decision must be fully traceable? If yes, use deterministic logic and keep humans in the loop for exceptions.
- Is the variance high enough and the judgment requirement real enough that a rule engine would require constant maintenance to keep up? If yes, now you have a case for an agent.
Telling a client that their problem does not need an agent is not a concession. It is the work. The teams I have seen get the most value from AI are the ones who are ruthless about where it belongs and where it does not. That discipline is what makes the places where you do deploy an agent actually work.