An AI reply assistant a viewer-services team chose to use

Complex letters in minutes, and a team that chose the tool

The setting: Viewer services at Bibel TV, a German broadcaster moving from linear TV to apps and streaming. Letters, postcards, emails and social messages, many of them personal.
My part: I built the assistant, a Microsoft Copilot agent, in September and October 2025, and ran the whole introduction myself, including the one-to-one sessions.
What changed: Complex letters with many points take minutes instead of hours. New and part-time colleagues answer in the house voice from the start. Both early holdouts now use it regularly. These are the team's observations, not a per-letter measurement.
Three takeaways: Give the assistant the first draft and keep the decision. Start from the standard you already wrote down. Invite, don't mandate, and prove it on the hardest case.

The mail still arrives on paper

The post comes as letters and postcards, some handwritten, some typed. Viewers write to thank the channel for a program and to say what it meant to them. They ask about programs, and about books and materials they saw in one. They criticize content too, some with an opinion, some with a theological argument.

Then there is the complaint that makes many points at once, complex and closely argued. Someone in viewer services reads it. The reply has to answer every point, in the house voice, with the wording of the service handbook. Nothing about that is quick.

The hard part was getting a team to want the tool

Today an assistant writes the first draft of those replies. That alone would not be worth your time.

What is worth it is how the assistant got into the team: nobody was required to use it, it is connected to no inbox, and its standard is one the team had already written down. That transfers to almost any team that writes many similar replies, none of them quite the same.

Personal mail does not forgive an autopilot

Internal guidelines say what viewer services responds to, what it doesn't, and how far it engages with contentious topics. Many on the team work part time, and finding the Bibel TV voice used to take new colleagues weeks, sometimes months.

So I agreed the goal with the head of the department first. The assistant would help with writing. Deciding what a reply says would stay a person's job.

Not everyone wanted AI at all. Two people on the team held out. As I read it: the unknown, preconceived ideas, probably some worry about jobs, and one objection I took seriously, that a reply to a viewer needs a human touch. The phone hotline is never to be handed to AI in any case.

The assistant is connected to nothing, on purpose

AI in customer service usually means automation: connect the inbox, let the model read along, send the replies. That would have missed how this team works. Viewer services works largely from paper post.

So the assistant is embedded in nothing. A person opens the letter and reads it. They decide whether the assistant is worth reaching for, and type or summarize the content in. They edit the draft, copy it into Word, print it and put it in the outbox. Email is more direct.

It sounds like a detour. It is the safeguard. A person has read every message before the model sees it, and a person decides whether and how to answer. Not answering at all stays an option.

The handbook already existed. It became the standard.

I built the assistant in September and October 2025. I didn't have to invent its standard. Viewer services had already written down what a good reply looks like: the service handbook holds the wording, the formatting and the standard phrases. I wrote a web page that condenses the handbook, its wording and its standard phrases, and the assistant consults that page for every draft. When the summary changes, the drafts change.

No model was trained. It is a Microsoft Copilot agent with a system prompt I wrote: a fixed answer structure, a house-voice specification and the channel links to close with. For emotional letters it calls for high empathy and pastoral language, with scripture or a promise of prayer where it fits.

Most teams have their standard written down somewhere. Just not for a machine.

The limits live in the instructions, and a person does the checking

The assistant has a list of things it stays out of: no medical, legal or financial advice, no political or denominational judgments, no missionary language. The broadcaster wants to stay objective and ecumenical. If the input isn't a viewer inquiry at all, an internal team note for instance, it writes no draft and asks for clarification instead.

To be plain about it: that list is an instruction, not a technical lock. Nothing checks the finished draft automatically. That is why a person reads every draft before it becomes a reply.

Optional, a week to try it, and won over on the hardest letter

I showed the assistant to the team and ran the introduction myself. Whoever wanted to tried it on their own for a week. One-to-one sessions followed. The team lead encouraged all of it and required none of it.

The turn came from a letter like the one at the top. One of the two holdouts tried the assistant on exactly that kind of complaint: complex, closely argued, many points at once. The draft answered every point, in the handbook's tone and the house voice, and they were amazed. By hand, they said, it would have taken hours at least. They have used it regularly since.

The second holdout was skeptical of AI in general. Later they told me they use it regularly too, and looking back, both saw it as a tool they stay in control of. This is where rollouts stall: a tool demoed on an easy example, then left unused.

Faster replies, and the house voice from the first day

Complex letters with many points at once now take minutes instead of hours, because the assistant gives an immediate starting point. None of this was timed letter by letter; the minutes, hours and weeks are the team's own experience.

New and part-time colleagues answer in the Bibel TV voice from the start, where it used to take weeks or months.

By 2026 part of the team used it regularly. Today, I'm told, it is part of daily work across the whole team, though not every letter is written with it. Every reply is still read, adapted and sent by a person.

The letter still comes by post

Someone opens it and reads it, as before. What changed is the sheet after that. It is no longer blank.

On it is a draft that answers every point, in the handbook's words. The person adjusts it, prints it and puts it in the outbox.

Before you hand your drafting to an assistant

Give the assistant the first draft and keep the decision.

Name who reads before the model and who sends after it. If you can't name both, the boundary isn't ready yet.
Start from the standard you already wrote down.

Handbooks, phrase banks and model replies are the best context an assistant can get, and most teams have one. Find yours before you write a single prompt.
Invite, don't mandate, and prove it on the hardest case.

A week to try it alone, then one-to-one time on each person's own work. A hard letter answered well is what won over a holdout.

Skip this setup if nobody reads the message first

If you want to answer high volumes fully automatically, with no person reading first, this is the wrong setup. The same goes if certain content must provably never appear in a reply. A list in the instructions can't promise that; you would need a separate check on the output.

Want to hear more?

Does your team write many similar replies, none of them quite the same? Look first at where your first drafts get written today. That is where this assistant earned its place: it reads no inbox, sends nothing, and one hard complaint letter won over a holdout.

Get in touch