Decision intelligence in UX

Decision intelligence is an approach that uses data and AI to put decisions on a solid, systematic footing, so teams can make more reliable decisions faster and with a clearer trail. Sometimes it even makes sense to automate simple, recurring tasks completely. Responsibility always stays with people.

In UX, we focus on decisions about the concept, development and optimization of products and services.

Over the past few months, we’ve kept asking ourselves how to get the most out of the possibilities digital progress offers. Some of our AI initiatives have found their way into day-to-day project work and have already proven themselves in practice.

Whatever the project phase, whether Discover, Define, Develop or Deliver, decisions always have to be prepared, made and carried out. That hasn’t changed. How those processes run for us in each phase today has:

Most projects start with plenty of open questions, loose ends and the need to build up context. Essential information is spread across large volumes of documents, presentations and notes, and anyone who needs a specific answer for a decision first has to find it. On top of that, these files often contain sensitive data that has no business being in Gemini, Claude and the like.

That’s why we built Delphi. It’s a local application that lets us bring everything for a project together in one place and build a dedicated knowledge base. We search it, ask it questions and analyze connections. With AI’s help, we keep an overview, find the answers we need and still protect confidentiality. Delphi also converts everything to Markdown automatically, so the AI model can work with it more efficiently. It came out of a project with large volumes of documents, where it helped us work through them efficiently.

What other teams can take away: Everything about a project belongs in one fixed place where you can ask targeted questions and find answers. If you want to work with AI on that knowledge base without second thoughts, set it up locally and securely from the start. That way, decisions can always be backed up with facts in a traceable way.

As early as 2025, we built a custom GPT for simulated user tests with AI personas and later turned it into a prototype. On one client project, it let us discuss directions without immediately launching elaborate testing, and then dig deeper based on that first assessment. We found that simulated tests can map the predictable early and fast, which is enough to form first hypotheses for decisions that set a direction. They don’t replace real user research, though. Simulations deliver the average. Real people, with rare, unexpected voices and real experiences, deliver variance.

A recent study by the Nielsen Norman Group confirms this. 120 synthetic participants and 604 professionals from tech and design answered the same question. The core message was the same in both groups, but the real respondents used more than 50 different terms, the synthetic ones about a dozen. Rare answers were missing entirely, and differences between professional groups looked smaller than they really are. NN/g therefore recommends synthetic users, as we do, as a starting point for hypotheses, not as a substitute for user research.

In practice, our clients benefit most. Simulated user tests save project budget and working time in the early phases, time that can go into more complex decisions where human variance and in-depth user research are needed.

What other teams can take away: Simulated user tests work well for a first orientation and for hypotheses that give a project a focused start. Before using them, define a clear goal and check whether the method fits it.

Prototypes are the most important tool for communication and validation when working together. Vibe coding has become almost indispensable here. All the more reason for us to make our vibe coding processes flexible and secure.

That’s why our engineering team built Vibedock, a reliable foundation for our prototypes on our own servers. The team is also responsible for the security and maintenance of this infrastructure.

In Vibedock, any team member can set up a new project on their own in a few steps, complete with a suitable template and access through their company account. The code lands in our own repository, and every change is deployed automatically.

For each project, we can also decide whether it stays internal, is shared with clients behind a password, or is made public. This lets us test hypotheses quickly and without friction. A real prototype exposes misunderstandings and wrong assumptions, reflects what people actually expect, and can then be tested with real users to answer specific questions about how it’s used.

What other teams can take away: Anyone building prototypes with AI needs a well-maintained, secure infrastructure with clear owners who have the necessary expertise.

User tests and surveys generate large amounts of data. Evaluating and preparing it by hand has cost us a lot of time in the past. It also gave us the experience we needed to build Plato, our AI-powered all-rounder, which now takes a great many steps off our plates in this area while leaving us in full control at all times.

Plato reads survey data, detects question types automatically, compares groups, maps questions to hypotheses, runs statistical analysis, validates and falsifies hypotheses, and presents results in customizable charts. Analysis of interview transcripts is also in beta. On client projects, we already use Plato to evaluate and prepare user test results, and we’re significantly faster as a result.

Results can also be turned directly into a presentation for stakeholders. Raw data becomes substantiated insights and tested hypotheses that later decisions can build on.

What mattered most to us is that everything can be adjusted at any time. Plato works step by step, every intermediate result is visible, and if an analysis ever misses the mark, a few quick edits fix it.

What other teams can take away: A decision is only as reliable as the analysis behind it. If you automate analysis, make sure you can follow, check and, if necessary, correct every intermediate step.

Decisions are rarely made alone. They emerge in meetings, workshops and conversations with stakeholders. Expectations and assumptions often sit between the lines of a confidential conversation, and its content doesn’t belong in common AI software. And weeks later, someone still asks why something was decided the way it was.

That’s why we use Transcriber, which we built ourselves. It’s a local application that records conversations, recognizes who is speaking, produces concise summaries and checks them against internal objectives.

Provided everyone involved agrees, even sensitive decisions and statements can be captured automatically, without manual typing. That lets us focus fully on the conversation and still show later what was decided. And because the entire conversation is on record, we can pick up on the hints between the lines afterwards and take them into account in future decisions.

What other teams can take away: Record decisions together with their reasoning in a fixed log and keep the full conversation. That way, you can show later what was decided and revisit the nuances. This requires the consent of everyone involved and local processing of confidential conversations.

A good decision only takes effect when it’s implemented as agreed. That’s where quality assurance comes in.

We’re currently running an internal experiment to find out at which checkpoints a human eye is needed and which are worth automating with AI.

Looking back, our role in UX design has changed profoundly. We carry out fewer manual steps ourselves, hand them off to AI more deliberately and review the results. Three principles run through all our examples: confidential material stays local, every step stays visible and correctable, and simulation never replaces real people.

We invest the time we gain where years of experience count: in rare, unexpected voices, in curating and interpreting results, and in taking responsibility for decisions. For our clients, that means a broader, better-substantiated basis for decisions in the early phases and more focus where people make the difference. In the end, responsibility always rests with people.

Nielsen Norman Group: Synthetic Users vs. Human Respondents (16.10.2026), https://www.nngroup.com/articles/synthetic-users-vs-human-respondents/