Turning a reactive support tool proactive with AI

A quick pilot with 150 agents, built to test two things at once: the AI's performance, and how naturally agents adapted to it.

Role:

UX / Discovery

Year:

2024/25

Context

Teaching the tool to think ahead of the agent.

MagentaView is what Deutsche Telekom's frontline agents use to handle live customer calls, 10k+ of them across Germany. An earlier redesign of its core screens had already cut call handling time by 18%.

The next question wasn't how to make the tool faster. It was how to make it think ahead of the agent, instead of waiting to be searched.

MagentaView is what Deutsche Telekom's frontline agents use to handle live customer calls, 10k+ of them across Germany. An earlier redesign of its core screens had already cut call handling time by 18%.

The next question wasn't how to make the tool faster. It was how to make it think ahead of the agent, instead of waiting to be searched.

My role

Shaping the pilot from discovery to decision



Discovery & Journey Mapping

I sat with agents and observed live calls end to end, from IVR to wrap-up. I mapped how the process actually moved, not how it was documented to move, and used that to identify where AI could genuinely help.

Mandatory dashboard detours

To get anywhere, agents had to return to the dashboard first. No direct path between sections. Agents looped Dashboard → Section → Back dozens of times in a single call.

POC Design & Build

I designed and built the three AI concepts tested in the pilot, Agent Briefing, Support Tips, and Wrap-up, working within a short timeline to get something real in front of agents quickly.

Two completely different environments

Call centre agents can view customer data before authentication. Retail agents cannot, they're face to face with the customer. One platform had to handle both without building two separate products.

Pilot Observation

Once live, I observed the two-week pilot directly across 150 agents. I tracked both how well the AI performed and how naturally agents adapted to working alongside it, not just whether they liked the idea.

Two completely different environments

Call centre agents can view customer data before authentication. Retail agents cannot, they're face to face with the customer. One platform had to handle both without building two separate products.

Eight designers, no shared ground

Call centre agents can view customer data before authentication. Retail agents cannot, they're face to face with the customer. One platform had to handle both without building two separate products.

Discovery

Watching the work, call by call.

Before designing anything, I sat with agents and observed live calls end to end, from IVR to wrap-up. Not to fix anything yet, just to understand how the process actually worked, and where the real opportunities were.

Before designing anything, I sat with agents and observed live calls end to end, from IVR to wrap-up. Not to fix anything yet, just to understand how the process actually worked, and where the real opportunities were.

"I just want to hear the customer. I shouldn't be typing while they're talking."

Agent, pilot observation session

"I just want to hear the customer. I shouldn't be typing while they're talking."

How do you build a tool that 22,000 agents choose to use, across environments and ability levels, when the foundation holding it together is nobody's problem?

The repetition tax

Customers got transferred and had to explain their problem again from scratch. Every transfer reset the conversation to zero.

Mandatory dashboard detours

To get anywhere, agents had to return to the dashboard first. No direct path between sections. Agents looped Dashboard → Section → Back dozens of times in a single call.

The information hunt

Agents searched wikis, policy docs, and internal databases mid-call, while still listening to the customer. Long pauses followed.

Two completely different environments

Call centre agents can view customer data before authentication. Retail agents cannot, they're face to face with the customer. One platform had to handle both without building two separate products.

Eight designers, no shared ground

Call centre agents can view customer data before authentication. Retail agents cannot, they're face to face with the customer. One platform had to handle both without building two separate products.

Wrap-up overload

After the call ended, agents spent minutes typing summaries and filing tickets before they could take the next call.

A tool agents didn't trust yet

Adoption was the real metric. If the platform slowed agents more than the old tools, consolidation would fail — regardless of how well individual features were designed.

The journey map also showed us where AI didn't belong. The moment a customer explains their issue stayed untouched. Agents needed to hear them without anything competing for attention.

That reframed the brief. We weren't building a smarter search tool. We were building something that removed the search entirely, so agents could just listen & resolve.

The journey map also showed us where AI didn't belong. The moment a customer explains their issue stayed untouched. Agents needed to hear them without anything competing for attention.

That reframed the brief. We weren't building a smarter search tool. We were building something that removed the search entirely, so agents could just listen & resolve.

Challenge

How do we leverage AI to make agents more efficient, not by doing their job for them, but by staying quiet, fitting their flow, and leaving the decision in their hands?

Challenge

How do you build a tool that 22,000 agents choose to use, across environments and ability levels, when the foundation holding it together is nobody's problem?

Design principles

Four constraints that shaped every decision.

That reframe gave us four filters. Every direction we considered after this had to survive all four, or it didn't ship.

That reframe gave us four filters. Every direction we considered after this had to survive all four, or it didn't ship.

Quiet, not intrusive

Not an assistant to do the agent's job. Support that fits their flow without ever breaking it.

Quiet, not intrusive

Not an assistant to do the agent's job. Support that fits their flow without ever breaking it.

The decision stays theirs

Agents rejected AI acting on its own. The system prepares. The human decides, always.

Pilot before rollout

We designed for 150 agents first, watched how they used it, and fixed the AI before it reached the full floor.

Narrow before wide

Billing came first because the data was clean and it was the most raised issue. A good idea on bad data teaches you nothing.

Introducing

Agent Brief

Agent brief turns holding time into preparation time. By summarizing the customer's initial AI-IVR intake into a crisp, scannable brief, we give agents the exact context they need to resolve issues faster. Simultaneously, the feature analyzes user intent to surface smart, contextual sales opportunities, turning a standard support call into a seamless revenue driver.

Agent brief turns holding time into preparation time. By summarizing the customer's initial AI-IVR intake into a crisp, scannable brief, we give agents the exact context they need to resolve issues faster. Simultaneously, the feature analyzes user intent to surface smart, contextual sales opportunities, turning a standard support call into a seamless revenue driver.

Saves Time: Eliminates the need for customers to repeat their issues, allowing agents to jump straight into resolving the problem.

Drives Revenue: Mentally prepares agents to pitch highly relevant offers naturally, without disrupting the support experience.

Introducing

Support Cues

As the call goes on, the system listens and queues up the next best action, sending a document, pulling an invoice, whatever the moment calls for, without the agent having to go looking for it.

As the call goes on, the system listens and queues up the next best action, sending a document, pulling an invoice, whatever the moment calls for, without the agent having to go looking for it.

The agent chooses to act on it or not, it never triggers on its own.

Queues actions like sending a document or pulling up an invoice.

Measuring Success

Small pilot, close watch.

The pilot ran with 150 agents, on billing topics only. The point wasn't to prove the concept at scale. It was to observe how the feature actually performed in real usage, and understand how the AI held up outside a controlled test, in real calls, with real variation, not just whether agents liked the idea of it.

The pilot ran with 150 agents, on billing topics only. The point wasn't to prove the concept at scale. It was to observe how the feature actually performed in real usage, and understand how the AI held up outside a controlled test, in real calls, with real variation, not just whether agents liked the idea of it.

150

Agents in pilot

3

AI features tested

2

weeks the pilot ran

150

Agents in pilot

3

AI features tested

2

weeks the pilot ran

What we heard

Agents weren't against AI. They wanted output they could trust without checking it, on a live call, there's no time to verify.

What that revealed

AI quality isn't just a model problem. It's a data problem, clean backend data is what makes an answer confident instead of just plausible.

What we decided

Agent Briefing held up best and was easiest to isolate. That's what gets built out fully first, backend included, before the wider rollout.

Reflection

What two years on one project leaves behind.

Two years on one project leaves a mark. Some of what I take away is what I'd do differently. Some of it changed how I work for good. The four below are the ones I keep coming back to.

Two years on one project leaves a mark. Some of what I take away is what I'd do differently. Some of it changed how I work for good. The four below are the ones I keep coming back to.

Watching beats asking

Agents couldn't have told me directly that split attention was the core problem. I only saw it by sitting with them and mapping the journey myself. When something feels off in a workflow, go watch it happen before you ask anyone to describe it.

Small and clean beats big and messy

Starting with billing, because the data was clean, taught me more in a short POC window than starting broad ever would have. A narrow scope that actually works beats a wide one you can't trust yet.

Small and clean beats big and messy

Starting with billing, because the data was clean, taught me more in a short POC window than starting broad ever would have. A narrow scope that actually works beats a wide one you can't trust yet.

Control is the feature, not the constraint

I went in assuming automation was the goal and speed was the metric. Agents taught me that trust was the real currency, and speed only mattered once trust was there.

Control is the feature, not the constraint

I went in assuming automation was the goal and speed was the metric. Agents taught me that trust was the real currency, and speed only mattered once trust was there.

Good AI needs good data

The gap agents felt wasn't a design gap, it was a data gap. AI is only as confident as what it's reading from. If the backend isn't structured for it, no amount of interface polish fixes that. That's not a footnote, it's now part of the roadmap.

Good AI needs good data

The gap agents felt wasn't a design gap, it was a data gap. AI is only as confident as what it's reading from. If the backend isn't structured for it, no amount of interface polish fixes that. That's not a footnote, it's now part of the roadmap.

More case studies

Previous project
Imagining the future of home with AI
Previous project
Imagining the future of home with AI
Next project
A redesign that turned passive visits into active engagement

LET'S CONNECT

Let's create something amazing together

I'm always excited to discuss new opportunities and ideas. Drop me a line. :)

© 2025 Adiy Bin Yunus / Made with ❤️‍🔥 in Berlin

LET'S CONNECT

Let's create something amazing together

I'm always excited to discuss new opportunities and ideas. Drop me a line. :)

© 2025 Adiy Bin Yunus / Made with ❤️‍🔥 in Berlin

LET'S CONNECT

Let's create something amazing together

I'm always excited to discuss new opportunities and ideas. Drop me a line. :)

© 2025 Adiy Bin Yunus / Made with ❤️‍🔥 in Berlin