---
title: "Platform Experience & AI — Maximilian Dahl"
description: "Product design case study by Maximilian Dahl at zvoove: from a fragmented product landscape to prepared, accountable decisions. Platform, AI assistant, human-in-the-loop, live since August 2026."
url: https://max.does.systems/en/work/platform-experience-ai/
lang: en
---

Start again

zvoove · Platform Experience & AI

# From individual products to one shared platform.

Start the story · SRead the case · R

Role in the project**Sole designer**Across teams · zvoove Germany

Collaboration**PMs, dev leads & Engineering**Feature teams in trios

Responsibility**Strategy to design engineering**Research · AI experience · Design system

As of September 2026 · Live since August 2026 · +3 · As of September 2026 · Live · Since August 2026: Launchpad, zAIn, Notification Center · 16 cloud customers, 12 active · Limited · 48 write endpoints behind the human gate, not shipped · Proof · Action Center and Decision Card, not shipped yet · Not measured · Impact: no before-and-after comparison · Launchpad · zAIn · Notification Center

## Lena’s morning · eight scenes

1. 01
   Another overview would not solve the problem.
   To make a decision, Lena has to find and combine information from several programs herself.
   Starting point

2. 02
   Monday, 05:52. Two shifts are still unfilled.
   Lena plans staffing for three branches. By six, she needs to know who can cover the shifts. The information is spread across different programs.
   Lena’s workday · reconstructed
   zvoove One · Staffing_×_zvoove One · Time tracking_×_Payroll_×_Absences – overview_×_zvoove Expert Mode_×_Inbox (37)_×_Shift plan wk 43.xlsx_×_zvoove Cloud · Applicants_×_Customer portal_×_Wiki – shift rules_×_Reporting_×_zvoove One · Orders_×_Time accounts_×_Inbox – North_×_+
   one.zvoove.example/dispo/plan?kw=43 · shell.zvoove.example/start · L
   **zvoove One · Staffing**Lena · Monday
   New · Edit · Delete · Export · Filter
   | Staff no. | Name | Customer | Assignment | Status |
   | --- | --- | --- | --- | --- |
   | 1042 | B., Jana | Logistics North | Mon 06:00 | sick |
   | 1077 | K., Marco | Logistics North | Mon 06:00 | open |
   | 1103 | S., Ali | Warehouse South | Mon 14:00 | planned |
   | 1119 | W., Eva | Warehouse South | Mon 14:00 | planned |
   | 1150 | T., Lukas | Logistics North | Tue 06:00 | open |
   | 1188 | P., Nora | Retail Centre | Tue 08:00 | planned |
   **Inbox (37)**Lena · Monday
   - North · Who covers Mon 06:00?
   - Customer South · Invoice October
   - J. B. · sick note (forwarded from WhatsApp)
   **zvoove One · Time tracking**Lena · Monday
   - M. K. · 38 of 40 hours
   - T. L. · 31 of 40 hours
   - 7 timesheets missing · wk 42
   **Absences – overview**Lena · Monday
   - J. B. · sick · from Monday
   - S. A. · holiday · 3–7 Nov
   - 2 requests open
   **zvoove Expert Mode · Shift plan**Lena · Monday
   - Mon 06:00 · 2 unfilled
   - Mon 14:00 · filled
   - Tue 06:00 · 1 open
   **Payroll**Lena · Monday
   - Payroll run October · waits for timesheets
   - Advances · 2 open
   - Statement wk 42 · locked

3. 03
   My first design solved the wrong problem.
   A shared interface looked tidier. In the customer session, the work of finding and checking information stayed the same. I dropped that approach.
   My decision · December 2025
   Staffing · Mon 06:00 · 2 shifts open
   Time · M. K.: 38 of 40 hours
   Payroll · Payroll waits for timesheets
   Absences · Sick note J. B. · from Monday
   Expert Mode · Shift plan: Mon 06:00 unfilled
   Mail · North: who covers Mon 06:00?
   zvoove · Launchpad_×_+
   one.zvoove.example/dispo/plan?kw=43 · shell.zvoove.example/start · L
   **zvoove One · Staffing**Lena · Monday
   New · Edit · Delete · Export · Filter
   | Staff no. | Name | Customer | Assignment | Status |
   | --- | --- | --- | --- | --- |
   | 1042 | B., Jana | Logistics North | Mon 06:00 | sick |
   | 1077 | K., Marco | Logistics North | Mon 06:00 | open |
   | 1103 | S., Ali | Warehouse South | Mon 14:00 | planned |
   | 1119 | W., Eva | Warehouse South | Mon 14:00 | planned |
   | 1150 | T., Lukas | Logistics North | Tue 06:00 | open |
   | 1188 | P., Nora | Retail Centre | Tue 08:00 | planned |
   **Inbox (37)**Lena · Monday
   - North · Who covers Mon 06:00?
   - Customer South · Invoice October
   - J. B. · sick note (forwarded from WhatsApp)
   **zvoove One · Time tracking**Lena · Monday
   - M. K. · 38 of 40 hours
   - T. L. · 31 of 40 hours
   - 7 timesheets missing · wk 42
   **Absences – overview**Lena · Monday
   - J. B. · sick · from Monday
   - S. A. · holiday · 3–7 Nov
   - 2 requests open
   **zvoove Expert Mode · Shift plan**Lena · Monday
   - Mon 06:00 · 2 unfilled
   - Mon 14:00 · filled
   - Tue 06:00 · 1 open
   **Payroll**Lena · Monday
   - Payroll run October · waits for timesheets
   - Advances · 2 open
   - Statement wk 42 · locked
   **+6**
   New interface · Search all apps …✦ Chat
   **zvoove One · Staffing**same app as before
   | Name | Customer | Assignment | Status |
   | --- | --- | --- | --- |
   | B., Jana | Logistics North | Mon 06:00 | planned |
   | K., Marco | Logistics North | Mon 06:00 | open |
   | S., Ali | Warehouse South | Mon 14:00 | planned |
   What happens here · J. B. is sick, but the shift plan still says "planned". The sick note has not been passed on.A nicer frame, the same apps underneath: they still do not talk to each other.

4. 04
   The programs stay. The case comes together.
   I focused the design on a shared task: connect the sick note, shift plan and hours without replacing the existing programs.
   My decision · connect the information
   Each program keeps doing its job
   Staffing · Time tracking · Payroll · Absences · Expert Mode · Inbox
   One shared platform
   ### Messages · What is happening
   Staffing · Mon 06:00 · 2 shifts open
   Expert Mode · Shift plan: Mon 06:00 unfilled
   Mail · North: who covers Mon 06:00?
   ### Details · What Lena needs to know
   Absences · Sick note J. B. · from Monday
   Time · M. K.: 38 of 40 hours
   Payroll · Payroll waits for timesheets
   Sign-in and access · Who can see and change what?
   Switch between programs · All programs in one place

5. 05
   Four messages. One task for Lena.
   In the target experience, the assistant spots the unfilled shifts. Lena receives the case with its supporting information before she has to search.
   Target experience · illustrative data
   Staffing · Mon 06:00 · 2 shifts open
   Absences · Sick note J. B. · from Monday
   Expert Mode · Shift plan: Mon 06:00 unfilled
   Mail · North: who covers Mon 06:00?
   4 messages → 1 task
   2 shifts on Mon 06:00 unfilled
   detected 05:51 · before anyone asked
   Staffing · Absences · Expert Mode · Mail

6. 06
   Two hours available. An eight-hour shift.
   I replaced the percentage with reasons people can check. M. K. does not have enough hours left. Lena needs to see that before approving anything.
   My decision · show the unresolved question · target experience
   Proposed by zAIn · Decision · Review staffing
   ### Cover two open shifts
   Monday, 06:00 · Logistics North · 8 hours per person
   **Not ready for approval**
   M. K. only has 2 hours available. The shift needs 8.
   **M. K.**6 hours short
   **2 of 8 hours** covered · 2 hours available
   **T. L.**Fits the shift
   **8 of 8 hours** covered · 9 hours available
   **Assessment: well supported**
   Based on**Shift plan****Hours balance****Qualification**
   The hours conflict is documented. A replacement for M. K. is still needed.
   **Find an alternative for M. K.**
   Keep T. L. Replace M. K. with someone for the full shift.
   Resolve in chat

7. 07
   Resolve the open question. Then approve.
   The chat picks up the prepared case. Lena asks for an alternative to M. K., checks the revised proposal, then approves both shifts.
   Target experience · execution not yet shipped
   ⌂Notification Center›**J. B.**_Monday, 19 October 2026_
   Tasks (9)
   Inbox · Done
   OVERDUE**3**
   _Shift_**Unfilled shift**J. B. — Logistik Nord◷ due Mon 05:30✦ Hand to zAIn
   _Time limit_**Assignment time limit exceeded**S. M. — BMW◷ 3 days overdue
   _Safety_**Safety briefing**T. F. — Siemens◷ 5 days overdue
   J. B. — Logistik Nord✓ Mark done
   _✦_zAIn can propose a replacement for J. B. on Mon 06:00.Hand to zAIn
   URGENCY
   Status_Overdue_
   Due**Mo 05:30**
   Type**SHIFT**
   **Chat with zAIn**
   Mon 06:00 · Logistics North
   Case completed
   zAIn
   The prepared case is here: two eight-hour shifts. M. K. only has two hours left. That proposal needs to change.
   Lena
   Keep T. L. Find someone else who can cover the full shift instead of M. K.
   zAIn Proposal from Staffing
   E. W. can replace M. K. Both candidates have the required licences and enough hours for the full shift.
   Prepared proposal · review and approve
   Proposed by zAIn · Decision · Review staffing
   ### Cover two open shifts
   Monday, 06:00 · Logistics North · 8 hours per person
   **Ready for approval****Approved by Lena**
   Both people can cover the full shift.
   **E. W.replaces M. K.**Fits the shift
   **8 of 8 hours** covered · Hours checked
   **T. L.**Fits the shift
   **8 of 8 hours** covered · 9 hours available
   **Assessment: well supported**
   Based on**Shift plan****Hours balance****Qualification**
   No conflict in the checked information. Lena makes the final decision.
   Only your approval changes the shift plan.
   Approve both shifts · Both shifts saved in the shift plan
   Mon 06:00 filled · T. L. + E. W.
   Lena
   Yes, approve both shifts.
   zAIn
   Both shifts are assigned and saved in the shift plan. M. K. remains excluded.
   zAIn
   This case is complete. The timesheets are a separate task and still need your approval.
   What happens here · zAIn prepares, Lena approves

8. 08
   One case done. The next stays open.
   In the target experience, all programs show the filled shifts. The timesheets still wait for Lena. Prepared does not mean completed.
   Target experience · Lena’s home page
   ⌂Home_Monday, 19 October 2026_
   Your morning at a glance
   The early shifts are covered. Next up: the timesheets.
   Prepared by zAIn · Decision · Needs your review
   ### Match 12 timesheets
   Week 42**Time tracking****Staffing**
   The matching proposals are ready. Check which assignment each timesheet belongs to.
   Already prepared: timesheets and assignments brought together.
   **Review matching**
   **Prepared matching · week 42**
   Scope · 12 timesheets
   Compared with · Assignments in the shift plan
   Your review · Do person, assignment and hours match?
   The proposal is open for review. Nothing has been approved or carried out.
   Only your approval changes the assignment.
   ### Done today
   **Mon 06:00 filled · T. L. + E. W.**
   Approved by you. Both shifts saved in the shift plan.
   **One shared state across programs**
   The updated shift plan is available to the other programs. No need to enter it again.
   **zAIn**Chat · Your dashboard
   zAIn
   **Good morning, Lena.**
   The shifts are covered, and I have prepared your next task. Would you like a quick status update?
   **Show status**
   **Completed**T. L. + E. W. cover the two shifts. M. K. remains excluded.
   **Your next step**Review the 12 timesheet matches. They are prepared, not yet approved.
   I prepare. You decide.
   What happens here · One case completed, the next still open

As the only designer, I designed _the platform_ and created the components other teams now build with.

02 · Starting point · 2 min.

## The opportunity was not another dashboard.

My first hypothesis was the obvious one: a shared shell with launcher, tabs, search, a small dashboard and a persistent AI chat. I built it alone in Figma Make in four days and showed it four days later to a staffing agency, a customer since 2018. They liked the interface. The value assumption did not hold.

Hypothesis

Hypothesis 1 · Value assumption

A shared shell creates enough value if it reduces navigation and context switching across products.

Discarded after the customer session, December 2025

![Figma Make prototype of the shared Shell, December 2025](https://max.does.systems/assets/cases/platform-experience-ai/shell-prototype-figma-make.jpg)

**Prototype · December 2025.** The shared shell from the customer session, deliberately rough: it was meant to test the hypothesis, not sell it.

> “A product world with bad interfaces doesn’t need to become a beautiful product world with bad interfaces — especially not if the beauty costs a lot of money again.”
> 
> Staffing agency, customer since 2018 · shadowing on the first Shell prototype, December 2025 · translated from German

The reason was not the interface but the landscape behind it. Fifteen years of acquisitions had left classic ERP, Expert Mode and cloud products side by side, and two earlier platform attempts had not brought them together. Users found the information but had to reconstruct the case themselves across tabs, systems and WhatsApp before they could decide. They spoke of eight to twelve context switches a day and two to three minutes to reorient after each. That is anecdotal, not measured.

A better frame reduces navigation. It does not reduce this reconstruction. So I discarded the prototype instead of defending it. And more than a prototype was at stake: Launchpad, zAIn and Notification Center are part of every zvoove One bundle, available from summer 2026.

### Twenty-four hours later, the next customer said the opposite.

Two customers, the same prototype, one day apart, opposite verdicts. That is not a research error but the evidence that one fixed entry point would have been the wrong answer.

Two customers, the same prototype

1. December 2025
   Small team · personal oversight
   The classic overview is seen as unhelpful day to day — with a manageable team size you already know the situation. What matters is fast access to a single fact.
   Chat · priority 1 · Dashboard · low

2. December 2025 · one day later
   Larger operation · Expert Mode
   Chat is rated “nice to have, but not essential”. The dashboard, by contrast, replaces the laborious search inside Expert Mode and is seen as the real gain.
   Chat · nice to have · Dashboard · priority 1

Context

The architectural resolution

Both are right, for their context. So the answer was not a fixed layout but a shell with configurable entry surfaces.

The contradiction became three variants instead of a compromise: task-first, AI-first and hybrid mix. A session in August 2026 favoured the hybrid mix. That is a signal, not a validation.

03 · Research as a corrective

## I tested every direction against real work.

At zvoove, research was not a phase before the work but the corrective for every direction. Six moments overturned or carried a decision.

Research in numbers

- ~20
   zAIn interviews, together with the PM

- 9
   interviews on the Safety Manager, 5 external, 4 internal

- 2
   customer shadowing sessions

- 11
   participants in the discovery workshop

- 3
   customers in the validation track

- 4
   personas from the user insights synthesis

Research trail · 6 stops

01 / 06 · 16 Dec 2025

1. 01 · 16 Dec 2025
   Shadowing at the first customer
   Enthusiasm for the unified interface, a clear no to cosmetics without solved interfaces.
   My part · Prototype and presentation
   → Decided · Shell discarded as the main value

2. 02 · 17 Dec 2025
   Second customer
   An on-premises customer saw no need for a shell.
   → Decided · Entry point configurable, not fixed

3. 03 · by Jan 2026
   zAIn interviews
   Around 20 qualitative interviews on chat and context.
   My part · Interviews, together with the PM
   → Decided · Chat patterns and context for zAIn

4. 04 · 14 Jan 2026
   Discovery workshop
   Eleven people from product, UX and engineering, one shared question: what does the entry point do, what does the chat do?
   My part · Author and facilitator, together with the research lead from Market Insights
   → Decided · Roles of Launchpad, shell and AI chat separated

5. 05 · Mar 2026
   User insights synthesis
   Four personas, a pain-point matrix and a code prototype with seven core modules.
   My part · Synthesis and prototype
   → Decided · Signal before prompt: recognise work before anyone asks

6. 06 · Jul 2026
   Innovation roundtable
   Customer tests of zAIn, Timesheet Automation and the onboarding agent.
   My part · Lightning rounds and persona interviews
   → Decided · Positive response before the August go-live

Further stops: Research Hub (July 2025), Safety Manager interviews, internal interviews with customer support, sales, key account and product, Master Unified Entry Experience, zCom, validation track with three customers.

### Four personas, one pattern

1. Dispatch
   - Sick notes via WhatsApp, calls or SMS, typed into the ERP by hand
   - context switches between ERP, phone, email and WhatsApp

2. Payroll and admin
   - statutory deadlines without proactive warnings
   - missing and faulty timesheets at the start of the month

3. Branch management
   - figures only through manual reports
   - no early warning on negative trends

4. Recruiting
   - applications across many channels without a central view
   - documents and work permits checked by hand

Shared pattern

All four roles reconstruct the case themselves before they can decide.

Personas · user insights synthesis, March 2026 · Time figures in the synthesis are estimates, not measurements.

> “If zvoove’s future product world looks LIKE THIS … it would be the first time in years that I’d even open a zvoove product again!”

> “We really need an AI chat like this that knows the zvoove context. It makes a lot easier.”

Staffing agency, customer since 2018 · shadowing on the first Shell prototype, December 2025 · translated from German

Context

How the direction got through

My concept for a shared entry point was on ice for a long time. Instead of waiting, I gathered evidence, internally and with customers, and talked the direction through with the VP of Engineering and the Head of AI of another unit. In the end, the product strategy for 2026 and 2027 was aligned with it.

Learning

Research overturned my own direction twice: the shell as the value and the percentage as trust signal. That is exactly what it was for: a corrective before engineering builds.

- Hypothesis 1: the shell as the value

- Hypothesis 2: the percentage

04 · Jobs to be done

## Four roles, one job: understand the case before you decide.

The personas say who works with zvoove. For the platform, what counted was when and what for. So the pain points from the synthesis appear here as job stories: situation, motivation, expected outcome. The role sits beside the sentence, not in it.

The four jobs to be done we identified share one core: understand the case before deciding, without piecing it together from several products. Per role, it looks like this:

- **Dispatch**Job story
   When · a sick note comes in via WhatsApp, a call or SMS,
   I want · it in the system without typing it in,
   so · I can restaff the assignment instead of switching between channels.

- **Payroll and admin**Job story
   When · the month starts,
   I want · to see which timesheets are missing or wrong and which deadline is next,
   so · payroll goes out on time and right.

- **Branch management**Job story
   When · a figure in my branch turns,
   I want · to know without building a report,
   so · I can act while there is still time.

- **Recruiting**Job story
   When · applications arrive across many channels,
   I want · them in one place, documents and work permits checked,
   so · no good application is left waiting.

Written from the pain points of the four personas, user insights synthesis, March 2026.

05 · Platform architecture & technical proof · 1 min.

## The platform had to carry context, not just navigation.

After the customer test the shared entry point still made sense, just not as the main value. The open question was technical: can the landscape carry context across product boundaries without migrating everything first? I built a technical proof in React and TypeScript on our production stack: two separate products in one shell, tested on the running system instead of in an architecture slide.

![Technical proof: two products on different tech stacks in one shell](https://max.does.systems/assets/cases/platform-experience-ai/technical-proof-two-stacks.jpg)

**Technical proof.** Two products on different tech stacks in one experience; it tested integration depth, permissions and shared context.

Architecture decisions

- Classic ERP · Legacy
   own domain logic

- Expert Mode · Legacy
   own domain logic

- Cloud products · Cloud
   own domain logic

Integration depth differs per product

Platform experience · Launchpad

- Identity

- Permissions

- Navigation

- Signals

- Relevant context

1. Domain logic stays where it is owned
   Only what is needed across product boundaries becomes shared: identity, permissions, navigation, signals, context.

2. Not every product needs the same integration depth
   Legacy and cloud connect at different depths; for users it stays consistent. Migration is not a precondition.

3. The proof became the central platform
   The platform experience went live with Launchpad and is set down in the product guardrail as the central platform.

My part

The technical proof in React and TypeScript on which these three decisions were tested.

Scope

- 152 → 25
   152 pages of legacy rights condensed into 25 roles.

- 1 day
   I built the login (16 screens in 3 layouts) in one day with Cursor, from the documentation and our design system.

06 · From UEE to UWX

## The shared entry became a work model.

With UEE I had defined the shared entry and prototyped it in code. It became Launchpad 2.0, where zAIn, the Notification Center and Timesheet Automation dock today. That settled where work arrives. What remained open was how people and AI get it done across product boundaries.

For that I developed UWX, the Unified Work Experience: the shared model for AI-supported work across all zvoove products. UWX is not a product of its own. Product, Design and Engineering align to it.

What UWX sets

- Context travels along.
   Moving from one product to the next does not mean rebuilding the case.

- AI where the work happens.
   In the background it sorts and prioritises signals; it shows up in the feature, in the Notification Center or in a conversation with zAIn.

- Autonomy in levels.
   Manual, assisted, prepared, autonomous. Which level applies depends on context, permission and confidence; the last one is the target.

- Approval is part of the model.
   Where a decision has consequences, a person decides, and that is designed in from the start.

I finished defining UWX in September 2026 in the 3/6/9 prototype, a vision system for the product managers. In it, PMs build live-like prototypes, and UWX is their operating system: anyone designing a new flow gets the model automatically.

**Illustration, rebuilt.** Launchpad with UWX: cases from every product, the urgent ones first, and zAIn with the context of the selected case. Sample data.

Today UWX is established. The product managers and their teams work with it, and it is anchored in the product strategy for 2026 and 2027.

See also:[UWX · Unified Work Experience](https://max.does.systems/en/work/uwx/)

07 · Agentic interaction · 2 min.

## AI had to start before the prompt.

A chat as the starting point comes too late. It assumes someone has already recognised, framed and phrased the problem, and according to the interviews that was the most expensive work.

So the AI had to start before the prompt: recognise relevant work, assemble context, prepare the next step. That is in operation. The assistant knows where in the platform someone is working and proposes from that context.

Interaction model · Target model · full chain

1. Signal
   An event, an exception or a deadline appears.

2. Context
   The system assembles only the relevant context.

3. Proposal
   The next step is prepared, not merely suggested.

4. Approval
   The prepared step becomes a deliberate human decision.

5. Action
   Executed, escalated or stopped — logged and traceable.

The workflow starts at the signal, not the prompt — and carries context through to the decision.

That led to a division of work between surfaces, instead of making chat the universal interface.

Division of labour between surfaces

**Notification · Makes relevant work visible before anyone goes looking. Ordered by deadline and consequence, not by timestamp.**

Notification Center · by deadline and consequence

- KH · Sick note — Karim Haddad · DB Schenker Hamburg · cover for the early shift · 3 replacement candidates prepared, message to the customer drafted · Critical · Due today, 12:00

- KK · Open shift — Klaus Klein · Audi GmbH · 11 May, 06:00 · Anna Berger suggested · approval needed · High · Due tomorrow

- Safety briefing due · 4 employees · Bosch AG · Dates suggested from the training calendar · Medium · in 5 days

Open case

**Chat · Clarifies and orchestrates. Strong on single facts, exploration and situations in motion. Weak on comparison and bulk handling.**

zAIn · Anna Berger

1. How many hours did Anna Berger work in May?A
   10:02

2. **38.5 hours** in May, 6 of them on the night shift. Shall I open the timesheets?
   10:02

3. Combined from 22 timesheets in May

Write a message …0/1000

**Structured UI · Carries evidence, approval and state. Many cases side by side, sortable, with a full history per decision.**

Cases · 4

| Case | Customer | Due | Status | zAIn suggestion | History |
| --- | --- | --- | --- | --- | --- |
| Early-shift cover | Audi GmbH | 11 May | Approval open | Schedule Anna Berger · Well supported |  |
| Sick note | DB Schenker | today | In review | Request the sick note · Uncertain |  |
| Extension | BMW Leipzig | 14 May | Prepared | Extend by 3 months · Well supported |  |
| Cancellation | VW AG | 16 May | Done | — |  |

The decision state must not fall apart when moving between these surfaces. That is why the prepared case needed its own persistent place.

That place is the Decision Object: a unit of its own with priority, context, proposal, rationale and open uncertainty, which persists after the decision.

Notification Center/Cases/Working-time conflict before payroll close · Schematic

Overview · Activity · Sources

Working-time conflict before payroll close

Recorded working time · 14 h · +4 h over limit

Daily limit · 10 h

zAIn · Why well supported?

Rule · daily limit unambiguous.

Context · case data complete.

Uncertainty · an earlier exception, but in a different context.

To be checked before approval.

Properties

Status · Approval required

Assessment · Well supported

Deadline · tomorrow · 12:00

State · Prepared — not yet executed.

Defer · Escalate · Approve 10 h, flag 4 h

Schematic with sample data: the same state, whether opened from a notification, the chat or a feature.

My part

The interaction model, the division of work between surfaces and the Decision Object. I designed zAIn in full, from research to chat experience.

See also:[zAIn · the assistant](https://max.does.systems/en/work/zain/)

08 · Experience principle · 2 min.

## Complexity in the analysis. Clarity in the decision.

My aim was to keep complexity in the system, not with the user. Decision compression condenses a prepared case to what a deliberate decision needs: what happened, what the system proposes, why, and what uncertainty remains.

![Prototype: prepared substitution process with approval](https://max.does.systems/assets/cases/platform-experience-ai/substitution-human-gate.jpg)

**Prototype.** A sick note via WhatsApp is recognised and the substitution prepared; nothing runs until someone approves exactly these steps.

The screenshot still shows a percentage. I walked it back. Google PAIR, the HHAI 2025 literature and the interviews showed the same thing: decisions need interpretable uncertainty, not apparent precision.

Hypothesis 2 · Percentage as trust signal

1. Withdrawn
   92 % confidence
   A percentage does not say what the assessment rests on or what would overturn it.

2. Evolved
   Assessment with rationale
   Interpretable uncertainty instead of apparent precision; colour orients, the popover shows what it is made of.

Principles · Decision compression

1. Decision relevance before completeness
   The flow shows only what affects trust and the decision; the rest stays reachable.

2. Interpretable uncertainty instead of a percentage
   Named qualitatively, with the rule basis, open uncertainty and the next check.

3. Bundle context, layer detail
   Evidence stays on the Decision Object; details appear on demand.

4. A standard case in about five seconds
   A design target for prepared standard cases, not a time limit for consequential decisions.

What the principles leave open

How far may the system push a case on its own, and when must human authority take over?

The next question

09 · Trust & human-in-the-loop · 4 min.

## Autonomy needed visible boundaries.

The closer a prepared case came to a real system change, the more convenience turned into liability. I needed three things that work together: autonomy by impact, escalation without loss of context, and an approval that stays visible in the interface.

### Autonomy follows the consequence of an action

The PM classified the 154 backend endpoints by read and write rights, and as a trio we decided the boundary together. 106 run read-only at launch. The 48 write endpoints sit behind the human gate, deliberately not in the first release. What counted was impact, permission and reversibility, not a score.

Autonomy boundary

154 backend endpoints by read and write rights

- **106** 69 % · read, enabled in the first release

- **48** 31 % · write, behind the human gate

1. Read
   Retrieve context, Join data, Check rules
   Allowed before approval

2. Prepare
   Form a proposal, Show consequences, Stage the action
   No system change yet

3. Change
   Write data, Change status, Execute action
   Human Gate

The boundary sits at the kind of action, not at a threshold someone can turn up.

### The score is a switch, not a display.

Confidence does not decide what the person reads but which route the case takes. Four bands, four routes.

Confidence · Sample dialogue, synthetic data

zAIn · Klaus Klein · Schematic

1. Take care of Klaus Klein.A
   10:01

2. I am not sure what you mean: the open shift on 11 May or today’s sick note?
   10:01

3. The open shift.A
   10:01

4. Refill shift · Approval required
   I will prepare a replacement. Nothing runs until you approve.
   Runs after approval · Decline · Approve
   10:01

1. Find a replacement for Klaus Klein at Audi GmbH on 11 May 2026.A
   10:01

2. The shift has two assignments: 06:00 and 14:00. Which one should I cover?
   10:01

3. One question instead of an assumption

1. Replacement for the early shift on 11 May, 06:00.A
   10:01

2. I found AB · Anna Berger · Best match · Experience at Audi GmbH · 8 km · available now · JW · Jonas Weber · Same qualification · 14 km · available · LD · Leyla Demir · Same qualification · 21 km · briefing needed with matching qualifications. Anna Berger fits best: experience in the same area, available now.
   10:01

3. Fill shift · Approval required
   Schedule Anna Berger for Audi GmbH on 11 May 2026, 06:00.
   Runs after approval · Decline · Approve
   10:01

1. Timesheet week 19 matched automatically

2. Done: 12 timesheets of week 19 matched to their assignments and logged in the Notification Center. Can be undone at any time.
   10:01

Write a message …0/1000

Beta · Please check all AI-generated content carefully.

My part

Engineering built the score. My work was the bands, their transitions and surfaces, and one rule above them: a higher band never opens a higher action class.

### Escalation without losing context

When rules collide or permission is missing, escalation must not be a restart: the most repeated finding from the customer interviews.

Handoff Packet · 3 rules

1. Boundary detected
   Hands over the prepared case instead of acting further. It stops before the action, not midway through it.
   System

2. Decision context
   - Trigger and relevant data
   - rules applied
   - proposal and rationale
   - open uncertainty
   Handed over

3. Judgement stays connected
   Check, adjust, decide — from the state the system reached, not from the state before the signal.
   Human

What is handed over: the situation, the next step and what remained open.

### Governance stays visible in the interface: in the chat as a three-layer pattern

Three-Layer Pattern · Anatomy of one answer · example

1. Understand
   Interpretation, clarification, handoff. What the system understood is visible before it plans anything.
   Message
   Replacement for the early shift on 11 May.A
   Understood: a replacement for the **early shift on 11 May, 06:00**, customer Audi GmbH.

2. Plan
   Sequence and steps. For write actions the approval sits here — the action is visible before execution.
   Intent
   Plan · 3 steps · Approval required
   Check candidates → schedule Anna Berger → inform the customer
   Runs after approval · Decline · Approve

3. Decide
   Output, confirmation, next step. The result stays attached to the case and remains traceable.
   Result
   Plan · 3 steps · Done
   Check candidates → schedule Anna Berger → inform the customer
   Anna Berger is scheduled and logged on the case. Can be undone at any time.

Autonomy, handoff and approval are written down as reusable rules.

![Design mock: prepared IBAN change requiring approval](https://max.does.systems/assets/cases/platform-experience-ai/human-gate-design-mock.jpg)

**Human Gate · design mock.** A sensitive change can be prepared; execution stays explicitly approval-bound.

10 · How I work & AI enablement · 3 min.

## Not being the bottleneck.

As the only designer in the German unit I had to solve two things at once: turn product questions into testable artefacts quickly, and create a way of working in which product and engineering do not wait for design.

My principle is fact first: make hypotheses visible early, test them and discard them when the evidence says so. Code is part of my design medium; I build into production code. Engineering decides architecture and operations. I bring the user case, the working proof and the question of what the system owes the person.

Working modes · the question decides the medium, not the process

1. Explore
   Make the product thesis testable
   While the experience is still moving. Four days to customer validation.
   - Figma Make

2. Prove
   Test behaviour in the real system
   When behaviour is part of the hypothesis: on the real stack.
   - React/TS
   - Cursor
   - design system

3. Enable
   Make what worked reusable
   What held up becomes a component, token or rule.
   - Design system
   - skill repository

As close to code as makes sense, as cheap to change as necessary.

### The design system carries all of it.

The modes only work on the same substance. So that prototype, proof and product don't drift apart, all three draw on UnityUI, zvoove's existing design system. I developed its Figma library further and, with the frontend lead, made the system AI-ready: prototypes in Figma Make, Claude Design or Cursor are built on real components.

UnityUI · one source, three consumers

UnityUI

I own UnityUI on the design side. One source, one version: deviations surface in the pipeline, not in review.

Live · dual publish

As of September 2026 · being rebuilt

Code, AI tools and Figma are to share one state; the Figma library is being rebuilt for it.

- Code · package registry
   Products build on the published package.

- Design tool · Figma
   The same version as in code.

- AI tools
   Defined with the frontend lead: prototypes are built on real components.
   Figma Make · Claude Design · Cursor

Context

And one retraction that belongs here

Direct-to-repo worked technically, but for complex flows handover costs rose. With the VP of Engineering we redrew the line: flows through Figma, components straight into the repository.

My part

Then came zvoove Brain, built on my own initiative: a skill set that now lives in the AI steering group (Value Builders), with reviews per department and an owner per skill. It no longer depends on me alone.

See also:[zvoove Skills · knowledge that lets others build](https://max.does.systems/en/work/zvoove-skills/)

11 · System & delivery evidence · 2 min.

## What actually held up.

The shipped core is broader than the prototype it came from, but not everything in the target picture is live. Keeping those states apart is part of the case.

The team built the pilot product, with me in the middle of it in production code with Cursor: Launchpad 2.0 as the shell, SSO connected to Rights & Roles, platform settings, search, zAIn with its integration and the Notification Center with AI enrichment.

![Context-aware assistant with case context, live](https://max.does.systems/assets/cases/platform-experience-ai/notification-assistant-live.jpg)

**Context-aware assistant · live.** It takes over the case context and proposes from it.

![Notification center with handover to the assistant, live](https://max.does.systems/assets/cases/platform-experience-ai/notification-center.jpg)

**Notification Center · live.** A case arrives with deadline and context and goes to the assistant as a prepared entry point.

In production · Live · 4 entries

1. Platform experience
   Shell, app launcher, rights and roles, settings; anchored as the central platform.

2. Signals · Notification Center
   Deadline- and risk-relevant signals, prioritised by impact. I built the frontend and shipped it with a backend engineer. Eight weeks were planned. After two I was moved to other priorities; the frontend was done, the backend half done.

3. Context-aware assistant
   Proposes from the context of where someone works; context can also come from a notification.

4. Enablement
   Skill repository, UnityUI guardrails and tool setup run in production.

Design & proof · 2 entries

1. Human gate and confidence
   Limited
   Patterns designed; the write endpoints are not shipped.

2. Action Center and Decision Card
   Proof
   Technical proof started, not shipped yet.

Pending · 1 entry

1. Impact
   Pending
   No before-and-after measured; that would need a baseline and telemetry.

As of September 2026

- Live since August 2026
   Launchpad, zAIn and Notification Center. Rollout in progress.

- 16 cloud customers · 12 active
   First GA wave. Usage, no measured business impact.

- 106 read · 48 write
   154 endpoints classified; write access outside the first release.

Action Center and Decision Card: technical proof, not shipped yet. Fewer context switches and shorter handling times have not been measured as impact.

The shift

At the start, the human was the integration layer between products. The target picture inverts that relationship: the system carries signal and context through to the decision — the human brings judgement and authority.

12 · Transfer & open

## What carries over, and what remains open.

The patterns are not tied to staffing. They apply wherever AI prepares decisions with real consequences: money, people, compliance.

Four patterns

Pattern 01

Authority stays with the person when the decision moves money

The system recognises, prepares and explains. Approval stays with the person who carries the consequence, from the start, not as a safety layer added later.

Pattern 02

Read before write, opened in stages

What the system may not do is designed as deliberately as what it may do. It opens by impact, permission and reversibility.

Pattern 03

Autonomy is a level per decision type, not a switch for the product

Filling a routine shift and changing someone’s pay are not the same decision.

Pattern 04

Escalation is a handover format

Whoever takes over should see that someone carries on, not that something starts from scratch.

Research at a glance

1. Research Hub
   Central repository and method templates set up anew
   July 2025

2. Safety Manager interviews
   9 interviews, 5 external, 4 internal
   July–Aug. 2025

3. Shadowing
   2 customers, prototype shown on site
   Dec. 2025

4. Internal interviews
   Customer support, sales, key account and product
   from Dec. 2025

5. zAIn interviews
   about 20 qualitative interviews, together with the PM
   to Jan. 2026

6. Discovery workshop
   11 participants from product, UX and engineering
   Jan. 2026

7. User insights synthesis
   4 personas, pain-point matrix, jobs to be done
   March 2026

8. zCom
   quick validation with early prototypes
   May 2026

9. Validation track
   3 customers, Launchpad and Notification Center
   June 2026

10. Innovation Roundtable
   Customer tests of zAIn, Timesheet Automation and the Onboarding Agent; lightning rounds and persona interviews
   July 2026

Role and craft · four proof points

Alongside the running product business, UWX took shape: the AI experience across all products. There was no brief and no established pattern; as the only designer I defined the area from the ground up, from contextual AI and signals to human-in-the-loop. Today UWX is established and in use.

- **Four days to customer validation**
   The first hypothesis was in front of a customer after four days as a running prototype. One session later I discarded it rather than defend it — the tangible thing first, the insight second.

- **Shipped beats complete**
   The Decision Center was designed but too large to ship. Cut back to a notification experience — with the data structure underneath that later becomes decision objects. The slice went live.

- **Run the interviews and steer by what they say**
   Interviews, shadowing and support conversations run and analysed myself — and the product direction re-aimed twice because of them. Research as a corrective, not as confirmation.

- **Others build without waiting for me**
   Product and engineering produce testable prototypes themselves. What comes out is consistent because the foundation is given — not because I saw every file.

### What remains open

1. Does the Action Center carry decision state in the real system?
   The technical proof is running. The test is whether the same case stays understandable across notification, chat and feature.

2. Does the system actually reduce reconstruction work?
   Not measured yet; that needs a baseline and telemetry.

3. Where can autonomy grow without diluting authority?
   For write actions on personnel data, purpose limitation, data minimisation and a traceable record become design requirements.

My assessment, not evidence

1. Assumption 01
   The level of work shifts
   Pre-sorted signals arrive on one surface. The person reads, approves or thinks along briefly. Not less work, but less reconstruction before it.

2. Assumption 02
   Trust moves from the interface to the model
   Then what is bought is not the tool but the confidence, and pay-per-call pricing becomes the obvious form.

3. Assumption 03
   The uncomfortable consequence belongs to the design
   What is relief for the decision-maker is loss for the role below. That has to be part of the design thinking.

## TL;DR: My contribution to zvoove

I led the shared Platform Experience from the first prototype into production, with Product and Engineering. Launchpad, zAIn and Notification Center are live; UWX is embedded in the 2026/27 strategy as the AI experience across products.

### Each product on its own

- Shift planning

- Working time

- Absences

Human **searches and reconciles**

The case comes together in their head.

### One connected workflow

- Shift planning

- Working time

- Absences

**Shared platform**Identity · Permissions · Context

**Prepared case**

Signal · Context · Proposal

Human**reviews and decides**

### Strategic goals

One shared platform · Connect existing and new products without rebuilding their domain logic.

AI that prepares the work · Carry signal and context through to the decision. Approval stays with the person.

Scale product development · Enable teams to build independently with shared patterns, UnityUI and AI skills.

fin (for now)

**Case notes · How to read this case.**

Evidence · Working prototypes, production interfaces and conceptual models are identified as such throughout the case.

Data · Names and data in the product interfaces are synthetic.

Scope · The case shows my product-design decisions and collaboration with Product and Engineering — not a technical architecture specification.

Last updated · September 2026. Later developments are not part of this case.
