Field Notes

Operational AI isn’t chatbots. It’s queues, controls, and evidence.

If your AI doesn’t land in a queue, pass a policy gate, and leave an audit trail, it won’t survive regulated operations. Here’s the mental model we use.

• Thinking Code AI•Operational AI • Auditability • Human-in-the-loop
← All posts
Post cover

The problem with “AI demos”

Most AI demos are optimized for immediacy:

  • paste some text
  • get a fluent answer
  • feel the magic

Regulated operations are optimized for repeatability:

  • every case has an owner
  • decisions are traceable
  • exceptions have a path
  • risk is controlled

If your system doesn’t behave like an operations system, it doesn’t matter how smart the model is.

The operational mental model

We build AI like this:

  1. Work arrives (email, fax, portal, attachments)
  2. It becomes a case (an ID, an owner, a state machine)
  3. AI performs bounded work (extract, classify, summarize, suggest)
  4. Policy gates apply (required fields, thresholds, approvals)
  5. A human reviews when needed (low confidence, adverse impact, $$)
  6. The result lands in the system-of-record (structured fields)
  7. Evidence is retained (inputs, outputs, rationale, who approved)

What we log (and what we don’t)

We log what helps answer:

  • what happened?
  • why?
  • who approved?
  • what was the source?

We avoid logging more sensitive raw content than necessary.

A good default is:

  • document fingerprints + metadata
  • extracted fields + confidence
  • citations/anchors into source
  • review actions + timestamps
  • policy checks that passed/failed

Why this matters

Operational AI is not a UX pattern.

It’s a systems pattern: queues, controls, and evidence.

If you want this to work in insurance, healthcare, or PBM workflows, that’s the foundation.