← All work

Case study · AI Product Design · Tex

Beyond the chatbot.

Tex V1 is a chat on our tax knowledge base. While it ships, I keep the team on the next question: what if the assistant knows your clients, and then your working day? Each user test writes the brief for the next horizon.

ProductTex, the AI assistant in Nextens tax software (RELX)
RoleDesign lead: horizons, V2 test, thesis supervisor
ResearchThree user tests: 10, 8 and 5 users
HorizonsThe law → your clients → your day
Period2025 to 2026
01 · Three horizons

Ship V1. Design what comes after.

V1 looks like the AI chat users already know. My designers struggle to ship what feels like a copy. So I split the work into three horizons. V1 needs only basic UI, which frees time to research and explore V2 and V3.

Trade-off: V1 ships with basic UI, on purpose.

NowLive since 2025

V1 · The law

Answer tax questions with sources, in a chat users already know.

NextAlpha done, in beta

V2 · Your clients

Connect Tex to client data. Ask across all clients, or zoom into one.

LaterResearch done

V3 · Your day

An assistant across the working day. Research first, with a thesis.

Weekly jam session
Every week the designers jam together in FigJam: open brainstorming on what comes next.
02 · Listen past the product

Users ask Tex to remember their clients.

We test V1 with 10 customers. It is easy to use, and the sources build trust. But the best ideas point past the chat.

EaseEasy to use. It feels like ChatGPT.
TrustThe sources make me trust the answer.
FixThe feedback buttons are too small to notice.
Sources5 of 10 want more sources: tax authority, trade register, case law.
FilesLet me attach a document.
ClientsRemember my clients' data, so the whole process goes faster.
The V2 brief
Connect Tex to client data.

Users ask for it before we build it.

Findings from the V1 test, paraphrased.
03 · Your clients

One question across 200+ clients.

Advisers compare returns by hand. Five years means five screens. A designer on my team designs a query layer on client data, under my supervision: ask across all clients, then zoom into one.

All clients

One question across all your clients

Which of my clients have income above €100,000?
23of 212 clients match: a list to start advice from.
One client

Ten years of returns, side by side

Income per tax year
Box 320242025
Bank balances€52,400€54,100
Other assets€48,200€4,820
This field differs from last year. Probably a typo: check it before you file.
Illustration with invented data. The concept, not the product.
04 · Test and respond

Eight users see the value, and the gaps.

I lead the V2 test as a team. Users want it. They also need to find it, and they ask how Tex gets its answers and what happens to client data.

Trade-off: more reach means more to explain.

8

Users in the V2 test. They see clear value: advice opportunities in their own client data, time saved and better client service.

FindingThe feature is hard to find.
ResponseFour variants of a mode switch you cannot miss
FindingUsers want AI answers explained.
ResponseReasoning you can follow, step by step
FindingUsers worry about the privacy of client data.
ResponsePrivacy returns in V3 as one of five conditions
Condition
05 · Your day

Design for the whole working day.

For V3, an intern on my team writes his thesis on AI in tax work. We set the research question and scope together, and I am his company supervisor. He maps one working day and gives each part its own design principle.

09:00 to 10:00

Planning and intake

Mail, team, first client questions. Information sits in four places.
Design principleContextual relevance
10:00 to 16:00

Core tax work

Returns and calculations. About 15 minutes per client just to collect the data.
Design principleHuman control
In between

Client contact

Advice and questions. Clients research it themselves with ChatGPT, which strains trust.
Design principleCognitive alignment
16:00 to the end

End of the day

Is everything filed? The fear of missing something.
Design principleExplainabilityCognitive alignment
Rebuilt from the thesis journey map: the average tax professional who sometimes uses AI.
06 · Conditions for trust

Tex prepares. The adviser decides.

The research includes a usability test of the first concept. Adoption depends on five conditions. After his thesis, Nextens hires the researcher as UI Engineer. How I grow people.

5

Professionals test the first concept in May and June 2026. They want Tex to do the prep work and leave the decisions to them.

VisibleAI features are easy to find and to use.
ExplainableUsers see how Tex reaches a conclusion.
Fits the workTex follows the process that already exists.
Human controlThe professional makes the final decision.
Privacy, explicitHow client data is protected is said out loud.
07 · Where it stands

Every horizon moves.

V1 is live and V2 is in beta. The V3 research is done.

3
User tests, one for each horizon
200+
Clients in one question
Beta
V2, after a finished alpha
5
Conditions for adoption, from the V3 research
08 · What travels

Two lessons I take to any AI product.

Let the last test write the next brief.

Users point past the product they have. Their ideas are your next research question, if you listen for them.

Give designers a horizon.

A team that ships a copy needs a future to work on. Once mine had one, the problems stopped coming up in our 1:1s.

Related: making the design system readable for AI, in Operationalising the design system.