In recent conversations, five separate support teams told us how they plan to bring an AI agent into their support. None of them planned to switch it on and walk away. In more than 30 earlier conversations, we heard the same thing.
Nobody goes live unsupervised. And the way they describe it, that isn't hesitation. It's the plan.
Buyers design the on-ramp themselves
Buyers arrive with the steps already worked out.
A productivity-software company scoped its pilot in three phases:
- A sandbox, running the AI on historical tickets.
- Supervised drafting, where agents see and check every answer.
- A limited go-live.
Another team described the same shape in different words: shadow mode, then supervised, then a controlled launch. A third wanted human approval in place specifically to build acceptance inside the team.
The steps have different names, but the logic is the same every time. The AI earns its way from answering nothing to answering customers, and the support team watches it happen.
Supervision is shorter than you'd expect
The supervised phase is not a long probation. The teams we talked to measure it in hours or weeks.
A customer service team lead at a cookware brand planned her go-live day like this:
"We can go live with a supervised option tomorrow at 8, for example. We will do it with the whole team. We will try to ask things for two hours or three hours. And then we can switch to go live."
Her only concern about supervision was not quality. It was timing: "If the customer writes in the afternoon or in the evening, nobody's here."
An e-commerce operations provider running support for a fashion brand put its AI in approval mode, where every answer needs a click before it goes out:
"I'm tracking it closely now, but if the AI writes a good answer, I'd approve it."
The plan is to drop the manual release after one to two weeks. Supervision is a phase to get through, not a place to stay.
The test has to answer one question
The digital lead at a book retailer described his test setup simply:
"I'll assign one or two of my people to work only through Typewise and compare the whole thing in a proper test scenario."
And he was clear about why:
"I'm steered by very numbers-driven people. Would the value pay off, yes or no?"
That is the real job of a pilot. Not to prove that AI can write an answer, but to give the person who signs off a yes or a no they can defend.
What this means
The go-live that works is the one the support team designs and watches. The first step is supervised, the supervision is short, and the test answers one question from the people who sign off.
If you're working out how your agents and the AI share the work during that phase, we've written about human override and approval patterns, and about launching when you have no historical tickets to start from.
Quotes are translated from German where needed and anonymised. Companies are described by industry and role only.
