There’s a big gap between:
“AI can generate code”
and
“AI can help you ship something real”
I’ve been working on tools that bring AI directly into production workflows, specifically inside ArcGIS Pro.
The goal isn’t novelty.
It’s usability.
The dream
You type:
“Find all parcels within 500 feet of a fault line and summarize by county”
And you get:
- correct code
- correct output
- something you can trust
The reality
What you actually get is:
- mostly correct code
- missing assumptions
- subtle errors
The output might run, but:
- it might use the wrong projection
- it might ignore null values
- it might silently fail
So how do you make it usable?
You don’t just generate code.
You build a system around it.
Pattern 1: Constrained generation
Don’t let AI generate anything.
Give it structure:
- known layers
- known fields
- known operations
This reduces hallucination dramatically.
Pattern 2: Execution guardrails
Before running anything:
- validate inputs
- check schema
- simulate execution
I’ve been using:
- dry-run modes
- logging outputs
- explicit validation steps
This turns “risky automation” into something usable.
Pattern 3: Human-in-the-loop
The best workflows aren’t:
- fully manual
- fully automated
They’re hybrid.
AI accelerates:
- setup
- boilerplate
- exploration
Humans handle:
- correctness
- judgment
- edge cases
The mental model shift
Stop thinking of AI as:
“a coder”
Think of it as:
“a fast but unreliable junior developer”
You wouldn’t:
- blindly run their code
- skip review
Same rules apply.
Where this gets interesting
Once you have guardrails, you can start doing more:
- chaining multi-step workflows
- evaluating outputs automatically
- iterating on results
This is where “agentic workflows” actually start to mean something.
What I’m working on now
- CLI tools that make workflows reproducible
- evaluation systems for AI-generated outputs
- better ways to inject context into prompts
Final thought
The question isn’t:
“Can AI write code?”
It can.
The real question is:
“Can we build systems where that code is safe, correct, and useful?”
That’s the problem worth solving.