Backstage · October 2026

Design principles for AI-supported products

AI can take on a lot of the work in a product. People still need to understand what is happening, trust what they get and stay in charge of what matters.

These are the principles I design AI support by. Each one comes with a small example of what to avoid and what to do instead.

At the end, I show how I use them in practice when I design and build with AI.

The examples are design principles and illustrations from a product I have built with AI.

Make the next step obvious

The user should understand straight away what to do, without being overwhelmed by information or choices.

Show the right information at the right time.

Don’t

New analysis

Configure your workspace, choose a template, set the output format and review the advanced options before you begin.

UploadImportTemplatesSettingsFormatAdvancedHelp

Start your analysis

Add the interviews you want to analyse.

Add interview files

Open a saved analysis

Make the process visible and understandable

The user should know what the AI is doing, how far the work has come and when they need to act.

Show the work steps and real progress, not technical logs.

Don’t

Processing…

POST /api/analyse 202
job 7f3a… running
chunk 3/7 · 4 812 tokens
retry 1 (format)

Interview 2 of 5

Reading the transcript
Finding observations
Checking the quotes

You’ll be asked to review when all interviews are done.

Make the value and the result clear

The user should understand what the AI will deliver, what the result can be used for and what its limitations are.

Set expectations the product can meet.

Don’t

Unlock insights instantly

✨ Generate insights

You’ll get

✓Observations with quotes and their source
✓Findings for each research question

Not included: a decision on what to build. That stays with you.

Start the analysis

Build trust through honesty

AI should clearly separate what it knows, interprets, suggests and cannot decide.

Show uncertainty where it affects the user’s decision. Never claim something is done without verification.

Don’t

✓ Analysis complete

Users want a faster checkout.

12 insights found

Draft · not yet reviewed

Said4 of 6 respondents describe giving up when the form resets.

InterpretationRecovering from errors may matter more than speed.

Can’t tellWhether this applies to new customers. None were interviewed.

Give the user control where it matters

The user should be able to review, adjust, approve and change direction at important decisions.

AI suggests and carries out. The user steers the goal and the key choices.

Don’t

✓ Done

AI chose concept B and created 14 tasks in your backlog.

Suggestion

Concept B

Why: it is the best supported by findings F2 and F4.

ApproveAdjustChoose another

Let the AI do the work

The AI should take responsibility for the complex work and minimise unnecessary questions, repetition and manual steps.

Only ask when the answer is needed to move forward in a good way.

Don’t

Before we start

Number of interviews

Interviewer’s name

Language

File format

Date of each interview

Speaker labels

I found 5 interviews in Swedish and recognised who the interviewer is.

What should we call the people you interviewed?For example: Teacher

Start

Make it easy to change and recover

The user should be able to correct something, go back or continue after an interruption without losing their work.

Keep the context, earlier choices and finished results.

Don’t

Session expired

Your analysis could not be saved.

Start over

Welcome back

3 of 5 interviews are done, and your choices are saved.

ContinueReview earlier choices

Make complexity optional

Those who want to move on quickly get a simple experience. Those who need to understand or check the result can go deeper.

Simplicity in the interface must never mean shallowness in the analysis.

Don’t

Finding F2

confidence 0.82 · sources I1, I2, I4, I5 · obs O3, O7, O9, O12, O15, O21“…” · “…” · “…” · “…” · “…” · “…”model run 7f3a · 4 812 tokens · v3

Recovering from errors matters more than speed.

Based on 4 of 6 interviews

Show the evidence (6 quotes) ⌄

Three complementary quality requirements

Beyond the design principles, I would always check:

AccessibilityClear language, understandable controls and working navigation for different users.
PrivacyThe user understands what information the AI uses and how sensitive material is handled.
TraceabilityIt is possible to understand the basis for the AI’s conclusions when it matters for the user’s decision.

How I use the principles when I build with AI

I designed and built Research Synthesis, a tool that takes interview transcripts to observations, findings and decisions, with AI as part of the workflow. The AI wrote much of the code. I decided what the product should do, tested each version and gave feedback.

The principles work as my review checklist. When something feels wrong in a test run, naming the principle it breaks turns a vague feeling into feedback the AI can act on.

A few moments from building it:

Principle 2 · Make the process visibleOne interview at a timeThe first progress view ran several interviews at once and counted them up while the last step spun for a long time. I asked for one interview at a time, with its own steps, then “1 of 5” and start over. Details like “part 2 of 2” were removed: they described the system, not the user’s progress.
Principle 6 · Let the AI do the workOne question instead of a list of namesAn early version listed every name it found and asked the user to sort them out. It was hard to understand. Now the tool asks one question, what should we call the people you interviewed, works out who the interviewer is and removes names itself.
Principle 1 · Make the next step obviousThe interface follows a flow that already workedThe synthesis prompt already had a user flow that felt easy to follow when run directly in Claude or ChatGPT. Instead of inventing a new flow, I built the interface around that one, step by step.
Principle 5 · Control where it mattersResearch questions first, decisions lastThe first setup asked for the purpose and the decisions up front. I moved it to start from the research questions and leave the decisions to the end, where the user makes them with the findings in front of them.
Principle 4 · Trust through honestyFindings rest on observationsUsers can start from observations they already have, but the synthesis is always built on observations, never straight from transcripts. That keeps every finding traceable to what was actually said.

AI makes building fast. The principles make sure the speed goes towards what users actually need.