Over the past two years, our UX team has experimented with AI across four disciplines: UX Copy, Localization, Research, and Product Design. What started as scattered manual processes has evolved into an integrated, AI-native workflow where our team orchestrates and AI executes.
Three Eras of UX Copy
Our UX Copy discipline evolved through three distinct eras.
In the pre-AI era, we handled everything manually — iterating on copy in design and documentation tools, creating localization keys one by one (each requiring a description, key name, character limit, and screenshot), and requesting translations via our ticketing system. It worked, but it was scattered and slow.
The AI-assisted era kicked off in 2023 with a custom copy assistant built in ChatGPT, and then accelerated in 2024 when trivago Copilot (our internally built AI tool, running multiple models from different providers) arrived alongside dedicated agents for each step: a UX writing agent (iterated over 30 times!), a key builder agent, and a translation task agent. The real game changer? Being able to create localization keys directly from the chat interface — no more manual spreadsheet exports and imports. The result: turnaround time dropped by 75% while output increased by 41%, all while maintaining high quality scores.
Now in the AI-native era, having shifted our stack to Claude, we’ve consolidated our agents into reusable skills that flex across tools like Cursor. Up next: a single orchestrator that knows which skill to pull in at the right moment, plus a “fast track” Claude project that lets stakeholders create copy and keys themselves — with proper guardrails and a QA step by a UX writer, of course. On top of that, we use automations to manage admin: blocking incomplete tasks automatically, centralizing A/B test wins and product contributions, and nudging stakeholders for feedback after 24 hours. The payoff: rework has stayed close to zero over the past two months, and task acknowledgment time is down significantly.
Localization: From Agency Dependency to AI Teammate
Previously, every translation was fully human and involved a chain of manual handoffs between our ticketing system, our localization platform, and external agencies.
In 2023, we ran our first AI tests with machine translation plus post editing (MTPE) — where AI handles the base translation and humans fine-tune it. Rolled out across all languages in 2024, this model meaningfully cut costs. From there, we built two AI agents: a localization knowledge base and a translation agent stocked with approved translations.
Our current setup takes it much further. A single workflow now runs the entire process through one central skill, which detects ticket types, matches keys, creates jobs across all supported locales plus parallel jobs for the agency, and drafts translations in-house. A dedicated QA teammate skill then checks every translation against glossaries, formality guidelines, character limits, and placeholders. And on the finance side, we’ve automated monthly invoice processing, tracking spend against forecasted budgets, and even drafting updates for budget owners.
Research: From Scattered Reports to a Virtual Research Teammate
Research used to work like this: a stakeholder would request support, and a researcher would spend hours — sometimes days — reading, investigating, formulating questions, recruiting users, interviewing them, and compiling findings. Reports varied in format depending on who wrote them, and ended up scattered across different documentation spaces.
We built over eight specialized agents — from an objectives formulator to a competitive comparison agent to, most notably, synthetic users: AI personas built on years of real trivago user behavior data that can simulate user interactions and generate reports.
Now, everything lives in one place: a virtual research teammate. Unlike the earlier agents, it isn’t a rigid, instruction-following bot — it engages in back-and-forth dialogue, identifies what you’re asking, and routes your request to the right skill automatically. Reports now come in a standardized format regardless of who initiates them, are stored in a centralized location, and — here’s the real kicker — are translated into machine-readable insights that continuously feed back into our AI tools. Research is no longer just alongside the workflow; it’s embedded in it.
Product Design: Building an AI Prototyping Orchestrator
Before AI, prototyping meant working with very complicated design files that didn’t always match what eventually shipped. We mapped our entire design process to find where AI could help.
The breakthrough for us: code as the source of truth. Rather than designing in Figma and hoping it matches reality, we built an orchestrator that takes a scenario prompt and a reference — a ticket or a design file — and generates a working prototype in code, handling the dev environment, fixing errors in the background, and keeping everything safely in a sandbox branch. Under the hood, it coordinates nine sub-agents running in parallel, each matched to the right LLM for its task, and even delegates to our other AI teammates, like the research assistant and copy assistant.
The results speak for themselves: real examples showed dramatically more polished, production-ready prototypes compared to earlier attempts. Medium-complexity design tasks have gone from 5–8 days down to 1–3 days. We also generate two reports after each prototype — one for stakeholders and one internal learning report covering efficiency and optimization suggestions.
Key Learnings
Takeaways we think apply well beyond UX:
- You orchestrate, AI executes. Bring your domain expertise and let AI follow your lead — not the other way around.
- Spend tokens wisely. Choose the right model for the right job, set up your technical foundations upfront, and don’t be afraid to test across different tools.
- Invest in reusable systems. Whether it’s a design system, a style guide, or a glossary — these shared foundations are what make AI teammates effective instead of expensive.
What’s Next
We’ve been laser-focused on the delivery side of our work so far, but the next frontier is product discovery — the messier, more ambiguous space where problems aren’t yet defined and opportunities are still emerging. If what we’ve accomplished on delivery is any indication, that’s going to be an exciting journey!
Follow us on