Guardrails that come off as it learns

What I've noticed with AI is that people expect AI to be perfect out of the box, and they over engineer it, or they just accept whatever it puts out, without any learning.

I like to take more of a hybrid approach, where you treat AI as if teaching a human to do a job. They aren't going to get it right off the bat. You can give them guardrails so they don't make big mistakes or create big problems from the beginning, but they're going to have to learn. And that's going to take input, direction from somebody else who already knows the business.

I think that's a happy middle ground between the two. And it's okay to make a small mistake in email to a client, as long as you come back and correct it and learn from it going forward.

That's the approach that I've taken. You have this core system that has guardrails, and then those guardrails start to come off as it's learning more and more. It's learning from a human that knows how to run the business.

What waits for a person, and what stops waiting

One step that I wouldn't let an agent do alone is push, write something to production, or send a price to a client. Critical things like that. It just sends whatever needs a human back for approval.

From our records:

• Trust is earned one kind of work at a time. Each kind starts with a person approving every item, and one complaint puts it back there. A person decides money, scope, dates and production changes.

Over time, as it learns the business and as the system itself gets better, it becomes more and more automated and requires less and less human.

It has to learn and remember

I don't think the systems need to be perfect. They just need to be able to learn. And they need to remember what they've learned. They do need to have review in place and governance, oversight in place. They need systems, checks and balances.

It's very similar to if I have somebody on my team. They can make some small mistakes too, as long as they learn from it and do it better going forward. And over time they get better and better and better.