When a support ticket comes into my managed services firm, an AI agent reads it before any human does. It pulls the client’s history, checks our documentation for that environment, drafts a triage summary, and in a good chunk of cases proposes the fix. A technician reviews it and acts. That’s not a pilot or a demo I show at sales meetings. It’s been our Tuesday for a long time now.
We also run agents for research, client onboarding, security reporting, and a pile of internal automation. I built most of this myself, on our own platform, because the off-the-shelf options either didn’t exist or wanted to own my data. So when I talk to executives about AI, I’m not summarizing a webinar. I’m telling you what broke in my own shop and what didn’t.
Here’s the honest scorecard.
What works
Anything where the AI drafts and a human decides. Ticket triage. First-draft client communications. Digesting a 40-page vendor contract into the six terms that matter. Turning a pile of alerts into a readable morning summary. The pattern is consistent: AI compresses hours of reading and writing into minutes, and a person spends their time on judgment instead of assembly.
What doesn’t
Anything where you let it act unsupervised on things that matter. I learned this with guardrails on, deliberately, and it still surprised me. An agent will be brilliant for two weeks and then confidently do something absurd on a Thursday. The failure isn’t frequent, but it’s weird, and weird failures are the expensive kind. Every agent in my company that touches anything real has a human between it and the consequence.
The part nobody sells
The other thing nobody selling AI will tell you: the model is maybe a third of the work. The rest is plumbing. Getting your data somewhere the AI can reach it, cleaning up the documentation it depends on, deciding who reviews what. Companies with a messy source of truth get a very articulate machine confidently repeating their mess back at them.
Where to start
So here’s what I tell executives who ask where to start. Don’t start with a strategy deck. Pick one workflow that eats skilled hours, where a draft-then-review pattern fits, where being wrong occasionally is survivable. Run it for a quarter. You’ll learn more from that than from any vendor briefing, and you’ll be far better equipped to judge the next pitch that lands in your inbox, because you’ll have production scar tissue instead of slideware.
And if someone tells you AI will run your operations without people? Ask them if it runs theirs. Mine comes closer than most, and I still wouldn’t make that claim.