The foundations
A design system with conversation components, the first interfaces and the first agentic flows.
My workAI-native SaaS
An AI assistant that handles invoices and paperwork for freelancers over WhatsApp, and the web platform their accounting firm uses to follow along. I’ve been on the product since day one: from the design system to the agent’s behaviour and the code.
Client and product names are anonymized. Visuals are schematic recreations of the real interfaces.
TLDR
Context
Freelancers and small businesses use the product to send invoices and handle tax paperwork from WhatsApp, and to stay in touch with their accounting firm. The firm works from a web platform connected to the same conversations.
The challenge
It started as a menu-driven bot. As it grew into an agent that understands requests, uses tools and carries out multi-step tasks, it raised a new design question:
How should an AI behave when it can decide, ask, act, and sometimes get things wrong?
I built the design system, including conversation-specific components, and designed the agentic flows that took the product from menus to natural conversation.
Evidence
Every week we run a query that pulls every conversation in production. The figures in this case don’t come from a sample: they cover all the activity on the platform, comparing before and after each change.
The other designer and I set up a Friction Radar. Every friction is cross-checked against Jira to separate what already has a ticket from what’s new. Between January and August 2026 alone it covered 595 users, 1,245 sessions and 36,212 messages.
I also audited the prototype of the mobile app for freelancers: I checked its 15 screens against real API data and found 48 issues, 10 of them blocking, such as a result that showed €69.08 too much.
I review agent traces in LangSmith to see how it uses its tools, and we used AI agents to build a test suite that checks what we see in the conversation logs against the actual code.
“It seems to happen” became “it happens in 29 out of 29 attempts.”
What we found
Decision 01
In some situations, the agent silently applied 0% VAT. It was technically convenient, but users couldn’t see it, and for accounting firms it was a real risk.
The agent tells apart fixed rules, company preferences and one-off decisions. It applies the first two. When a one-off decision has consequences, it explains and asks, and the user can always leave it to their accounting firm.
An agent that always asks becomes tiresome, and one that never asks becomes dangerous. It only asks when it really matters.
Decision 02
Hitting the plan limit in the middle of an invoice interrupted the task. And sign-up asked for tax details before the product had delivered anything: usage analysis showed many customers stopped there without trying it.
If you hit the limit, the draft is saved and issued automatically after payment, on WhatsApp and on the web. Tax details are requested right before the first invoice.
It’s the same pattern across the product: warn in time, keep what the user was doing and pick it up afterwards.
Decision 03
Answers to official questions were long and technical: the main point was buried, jargon went unexplained and sources got lost along the way.
A composition step that answers the actual question first, translates jargon and groups sources by organisation, with a next step at the end.
Since this is tax information, answers had to get clearer without losing accuracy or inventing figures. That’s why we also analysed which sources each answer used and how reliable they were.
Decision 04
Some invoice errors didn’t show on any screen: a €515 total that, according to the Tax Agency, should be €550, VAT exemptions without their legal reason, or tax IDs with the wrong letter.
Before switching on Verifactu, check what the platform produced against the Tax Agency’s rules and the invoicing regulations, and take each error to engineering with the correct calculation and the rule that requires it.
With Verifactu, every invoice reaches the tax office the moment it’s issued, and one with errors would be flagged as “accepted with errors”. Better to find them first.
Figures come from the Jira tickets and the test session, checked against production between September and November 2026.
The firm’s side
For accounting firms, the home screen starts with “Ask your question here”, ahead of the menu. Each client query opens with full context next to the conversation, and a tax notice turns into a plain-language summary and a client message in one click.
Transactional mode
Messages from the accounting firm wait in a queue instead of breaking a critical flow.
Campaigns
After finishing a task, the agent reminds the user what’s still pending (“14 of 17 questions”) and when it’s due. The firm sees who’s missing, not just who received the message.
Engineering
Besides designing, I write code in the product: 126 commits across the backend, the platform and the website, in 28 merge requests the team has already reviewed and merged. I built a new invoice-header feature end to end, from database migrations and backend changes to the Angular interface. I also write the sprint tickets that reach engineering.
GitLab data as of 10 November 2026, excluding merge commits.
Impact
We changed when it asks, when it explains, when it waits and when it lets a person decide.
Upgrading without losing the draft, onboarding that asks for tax details later, and campaigns are live. And every friction reaches engineering with its frequency, its traces and a ticket, so priorities are set with data.
What it solves
How it grew
I’ve been on the product since day one. These are the stages it has gone through.
A design system with conversation components, the first interfaces and the first agentic flows.
My workThe agent understands what you ask in your own words, or in a voice note, and gets it done.
My workEvery message goes to the right specialist: invoices, quotes, expenses, official queries or paperwork.
Queries with full context, campaigns with tracking and tax notices turned into plain language.
My workUpselling that keeps the draft, and onboarding deferred until the first invoice.
My workA friction radar, agent traces, recomposed answers and the mobile app prototype.
My workSame agent, two contexts
Less screen, more conversation.
Full context, without switching tools.
What I learned
The changes with the most effect were in what the agent says and when it asks, more than in the screens.