Overview
Letting people hand tasks to an AI assistant, with control at every step
Role: Sole Product Designer, with 1 PM, 3 engineers & CTO.
Context
Mavue is an early-stage B2B SaaS company that helps small and mid sized manufacturers in Germany create their annual sustainability report, from collecting ESG data and spotting risks to setting goals and writing the report itself. Throughout the year, they collect data on Environment, Social and Compliance topics and use it to identify risks, challenges and opportunities. Once a year, much like a financial report, they turn these insights and their goals into one report.
Problem
The platform had no help at all. There was no chat assistant, no customer support inside the product, no way to share an issue and no place to find FAQs or recent product news. The only way to give feedback was the weekly or biweekly user interview, which I participated on an ongoing base. Anything that came up between got lost or led to frustration because of slow problem solving.
Goal
Create a chat assistant that solves all of this. It should be a simple help assistant and also an AI agent that does repetitive work, for example “Assign, schedule and update XYZ for me.”
Constraints
Small team: Support was handled by a small team, and at that point partly by me, so the assistant had to answer as much as possible on its own.
Early AI tools: The product was early stage and AI assistants were new for this audience.
A mistake is costly: One wrong change can distort a report that the company has to stand behind.
Solution & Key Decision
The assistant starts at a home screen with three ways to get help:
Home: The main page shows recent requests, the most recent articles and the most requested FAQs, so people see what matters without searching.
Self service: People find answers on their own through FAQs and recent product news, to overall reduce the human contact for simple questions.
Ask a question: The support team answers, or passes the problem on to tech if it is a bug, or to the PM if it is about the product.
Perform a task: This is the AI agent. People describe a task, such as assigning, scheduling or updating items, and the agent does the repetitive work.
Decision 1: one clear starting point: The home screen offers all three, so people don’t need to know where to look.
Decision 2: every question reaches a person: Issues go straight to the team that can solve them, and people no longer wait for the next interview to be heard.
Decision 3: the agent proposes, the person decides: An agent that acts on its own saves the most time, but one wrong change can distort a report. So I chose a middle path. The agent shows every change in a plan card, and each line can be confirmed or cancelled on its own. Nothing runs until Start automation, a progress state offers Stop, and the result offers Undo. Confirm all keeps the common case to one click.
Process
Understand the gap: I listed the status quo and what users could not do: find answers, share an issue or get help while they work.
Benchmark: I looked at Intercom and other help assistants. Most start with a home screen that gives a central view of recent requests and recent articles. From there, people reach the different kinds of help. I took this pattern because it is clear and simple.
Define the structure: I fitted our needs into this pattern and added the agent as the third way to get help.
Sketch and design: I drafted in Figma, then built high fidelity designs for different user flows, from the assistant home to a finished task.
Test and refine: I tested the assistant with different users, gathered the feedback and finalized the design.
Outcome
Help inside the product: People get answers and support without leaving the platform.
Feedback anytime: Issues no longer wait for the next interview.
Less repetitive work: Many changes can be handed over in one request.
In control: People see every change before it happens, and Stop and Undo make mistakes easy to correct.
Learnings
The biggest learning came from testing. Most users were not familiar with AI agents. When they opened Perform a task with an empty prompt box, they were lost. They did not know what to ask, and so they did not use the full power of the agent. Specific examples at the start fixed this. They showed people what the agent can do and inspired them to try it.
What I would do differently today
AI tools were still very early when I worked on this. Today I would:
Test with a real model: Use Claude Code to build a working prototype early, to see how the agent handles unclear or wrong requests instead of designing only the happy path.
Synthesize research faster: Use Claude Projects to collect interview notes and support requests and find the common themes, while I decide what matters.
Explore more drafts: Use Claude Design and Figma Make to produce more variations of the plan card and test them with users.
Design for mistakes: Plan the unhappy paths from the start, such as a wrong suggestion or a partly failed task.
Teach the agent in use: Suggest new examples based on what a person already does in the product, so the examples stay useful after the first visit.
Note: The screens shown were recreated, anonymized and adapted: data, copy and branding are invented to protect confidential product details.
