Every degree of freedom you give AI is a source of variance you will eventually blame on the model.

When it comes to automation, everyone makes the same mistake. Large document in. Big expectation out. When the output is wrong, inconsistent, or half-baked — the model gets the blame.

But is the model the problem? Shouldn't we be looking closer home?

The model is capable. Genuinely capable. But asking it to find the signal in the noise, structure it, reason over it, and hand you output that would make an analyst proud — all in one pass — is asking for too much. When it falls short, we blame the model. Not the thought. Not the process. Not the prompt.


The fix isn't a better prompt. It's a better pipeline. A three-course meal, one course at a time — not a buffet plate piled high.

One pass
Document
↓ Prompt
↓ Output
Staged pipeline
Extract
↓ Determine
↓ Analyse
↓ Output

What delivers the better result? Has the model changed? Is the prompt suddenly smarter? Or is something else at work?

Extract
Retrieval only. Pull the relevant fields from the raw source. Not reasoning. Not interpretation. Retrieval.
Determine
Structure what came out. Resolve it into clean, bounded inputs. This is the step almost everyone skips. It's also where almost all inconsistency originates.
Analyse
Clean data. Defined schema. One job. The output quality difference isn't marginal — it's categorical.
Output
Neat tables, structured data, formatted exactly as needed.

Why does the staged approach work? Errors localise. Tokens go to reasoning, not retrieval. Same input, same output — auditable, repeatable, fixable.


Tell people this and you'll hear the same objections every time:

"You're describing RPA. What happened to generative AI?"
"How is this different from UiPath, Power Automate, Selenium scripts?"
"You're limiting the model's potential. This is going backwards."

My answer: if I get better, consistent output at scale — call it RPA for all I care.

Traditional automation needs perfect inputs. This pipeline is built for imperfect ones — unstructured documents, variable formats, inconsistent data. The steps are deterministic. The inputs don't have to be.

Asking the model to retrieve, structure, and reason simultaneously degrades its performance. Focus it on one job per step and it executes at its ceiling. That's not constraining the model. That's deploying it at full potential.


I know this because I spent time blaming the model — until the output kept falling apart every time the document structure changed. Blaming the model is the easy thing to do. Looking inward isn't.

This isn't an AI insight. It's a manufacturing insight applied to AI.

You wouldn't hand a senior analyst 300 pages of raw data and ask for a conclusion in one pass. AI is no different. We just forget that sometimes.

The pipeline is the process. The model is incidental.

← The Pipeline