A few years ago, AI stopped being something we experimented with on the side and started showing up everywhere at once. Development. Support. Documentation. Internal tools. Reporting. Ideas for customer-facing features. All in the same stretch of time, across teams that hadn’t necessarily talked to each other about it.

That was a good problem to have. It also meant we didn’t have a consistent way to think about approved tools, data protection, ownership, or where AI should and shouldn’t be used yet. So we formed an AI Governance Committee, not to slow things down, but to give people enough guardrails to keep experimenting safely once customer data and production systems were in the mix.

Owning the Process, Not Touching Every Step

“Human in the loop” gets used a lot right now, often loosely enough that it stops meaning much. Austin Shaw, MACC’s AI Champion, put it simply when we were talking through this internally:

“To me, it does not mean a human has to manually touch every single step forever. It means a human owns the process, understands where AI is being used, and reviews the parts that carry real risk.”

That distinction matters more than it sounds like it should. AI-generated testing guidance can save QA a lot of time, but QA still owns quality. AI can summarize a support ticket in seconds, but the rep still owns what they tell the customer. The AI does work. A person is still accountable for the outcome.

What That Looks Like Before Anything Ships

Internally, that principle turns into a short list of questions we ask before any AI use case moves forward, not as a checkbox exercise, but because skipping them is how teams end up with a workflow nobody fully understands:

  • What problem are we actually solving?
  • Is AI the right tool for it?
  • What data would it need access to?
  • Who reviews the output?
  • What happens if it’s wrong?
  • Can we measure whether it helped?

None of that is complicated on paper. It’s harder in practice, because the honest answers sometimes slow a launch down or rule a use case out entirely. We think that tradeoff is worth it. As Austin put it, “the model is only part of the solution. The harder part is everything around it: the workflow, the data boundaries, the review process, the permissions, the documentation, and the measurement.”

The Line We’re Trying Not to Cross

It’s easy for AI oversight to become its own burden, another system your team has to babysit instead of the thing that was supposed to save them time. That’s not the goal. Good AI should reduce the manual effort it takes to run the business, not hand your team a new job managing the AI.

So the standard we hold ourselves to is pretty simple: AI should be useful, governed, measurable, and accountable. Not flashy. Not hands-off. Just built so that a person is still the one responsible for what happens next.