Kent Salcedo

Field Note

AI automation without the science experiment

Practical guidance for teams adopting AI workflows without turning the org into a research lab.

2 min read
aiautomationworkflows

Every company I talk to wants AI to save time. Most of them are one bad implementation away from creating more work than they eliminate.

The pattern is familiar: someone demos a chatbot, leadership gets excited, a team spins up a "AI initiative," and six months later there's a Slack channel full of prompts nobody trusts and a workflow that requires a PhD to maintain.

That's not AI adoption. That's a science experiment with a company email domain.

Start with the workflow, not the model

The best AI automations I've built didn't start with "what can GPT do?" They started with "what does someone on this team do every week that follows a pattern?"

Good candidates for AI-assisted workflows:

  • Document assembly: pulling data from multiple sources into a structured output
  • Research summarization: turning raw inputs into reviewable drafts
  • Classification and routing: sorting incoming information into the right buckets
  • Draft generation with human review: first passes that a person edits, not publishes blindly

Bad candidates:

  • Anything where the output goes directly to a customer with no review
  • Decisions that require domain expertise the model doesn't have
  • Workflows where the failure mode is worse than doing it manually

Build for maintainability, not magic

The teams that succeed with AI automation treat it like any other internal tool: understandable, documented, and owned by someone who can fix it when it breaks.

That means:

  • Clear inputs and outputs: no black boxes
  • Human checkpoints: especially early on
  • Logging and observability: you need to know when the workflow fails
  • Fallback paths: what happens when the API is down or the output is wrong?

I call this "boring AI." It's not going to win a hackathon. It will save your team ten hours a week for the next two years.

The science experiment test

Before deploying any AI workflow, ask:

  1. Can someone on the team explain how it works without saying "the AI just knows"?
  2. If it breaks tomorrow, does the team have a manual process to fall back on?
  3. Is the time saved greater than the time spent maintaining and reviewing the output?

If any answer is no, you're running a science experiment. That's fine in R&D. It's expensive in operations.

No magic. Just useful systems.