This deep dive expands on the Loop Asia conversation with Uli Hitzel, AI educator and former Microsoft/Axway/Dyson engineer.
There's a pattern Uli Hitzel says he keeps hearing from companies: "we know nothing about AI, and can we deploy something in production in three weeks from now." It reminds him of early cloud computing, when boards wanted the whole stack migrated overnight and technical leaders had to explain that nobody moves a core production database to the cloud in a few months. You start smaller, build the muscle, then move the thing that actually matters.
Two rollouts that look similar and aren't
Hitzel drew a distinction that's easy to miss if you're thinking about "AI in customer service" as one initiative: a customer-facing chatbot built with no prior AI experience is a completely different risk profile from giving existing agents AI assistance so they close a call in five minutes instead of thirty. Both use the same technology. One puts an unproven system in front of your customer with no human in the loop. The other keeps a human accountable and uses the model to make that human faster.
If you're choosing where to spend your first six months, that distinction is the whole ballgame. Augmenting people who are already accountable lets your organization build judgment about where the model is reliable, without betting the customer relationship on getting it right the first time.
The reliability gap nobody puts in the pitch deck
Hitzel's sharpest comment was about debugging. With traditional software, when something breaks, you can trace it. With a language model, you get an apology instead: "I'm sorry, you're right, I shouldn't have done this, I won't do this again." That trade-off is structural — the "five-nines" reliability language your infrastructure team uses everywhere else doesn't transfer cleanly to AI, not yet.
That's not a reason to wait, it's a reason to sequence deliberately: start with something non-critical, and if it has to matter, make sure it isn't customer-facing and someone can catch a bad output before it does damage.
Treating your AI rollout like a feature launch instead of a capability build?
Sequencing from low-stakes to high-stakes, with the team's judgment improving every cycle, is the same discipline I bring to advisory and training engagements.