Last 40Networking Pass tickets — then sold out
Design Strategy

Od Pomysłu do Realizacji: Jak zaprojektować i zaimplementować Generative AI w swoim produkcie

September 20, 2024 12:05 PM
Long Lecture
🇵🇱 Polish
Wisła

About

Designing the very first Generative AI for a PR SaaS can be a challenge. Recently, we developed an AI Assistant that helps our users write good quality press releases. I will share our process of designing Generative AI - starting from defining business requirements and hypotheses, then going through the design process and collaboration with other stakeholders, ending on what went well and what didn’t. I will also share some tips and tricks for designers - how to navigate through the unknowns around AI and how to design when good practices are not yet set.

Talk by

Photo of Aleksandra Wancisiewicz, Product Designer at Prowly
Aleksandra Wancisiewicz
Product Designer at Prowly

Watch the full talk

Watch this WaysConf session, then continue with related talks or explore the current programme.

Explore the 2026 agendaMore WaysConf recordings
From the recording

Talk in brief

A product case study on adding a generative-AI writing assistant to a PR application. The feature helps users move from a key message to press-release ideas, draft content, and an evaluation against eleven editorial criteria, while keeping the human author in control. The talk covers supportive feedback language, prompt iteration with a subject-matter expert, token costs, refresh behavior after edits, usage limits, unpredictable output length, and multilingual testing. Adoption and conversion signals improved, but experienced users applied fewer suggestions, reinforcing that the assistant was never intended for everyone. The lasting guidance is to explain outputs, design for trust, and use AI to support creative judgment rather than replace it.

Key takeaways

  1. 01

    Design AI around a concrete user struggle

    The team began with blank-page anxiety, slow drafting, and uncertainty about quality rather than adding generation for its own sake.

    Watch from 2:01
  2. 02

    Keep the author in editorial control

    Users can choose ideas, edit the draft, hide assistance, and decide whether to apply suggestions or refresh the evaluation.

    Watch from 4:00
  3. 03

    Explain recommendations and soften judgment

    Reasons, criteria, and constructive language make automated criticism more understandable and less threatening to professional confidence.

    Watch from 8:01
  4. 04

    Budget for iteration and inference costs

    Prompt experiments, repeated evaluations, manual edits, and token usage all affect both delivery time and the viable interaction model.

    Watch from 12:03
  5. 05

    Validate usefulness by user segment

    Strong adoption by new users coexisted with lower suggestion use among paying experts, showing that aggregate success can hide different needs.

    Watch from 18:01
Read edited transcript highlights

These concise notes were edited from automatic captions and checked against the talk structure. They are not a verbatim transcript.

The feature started with the writing process

The team identified three recurring obstacles: users did not know how to begin, a polished press release could take hours, and less experienced writers could not judge its quality. An early prompt asked questions a journalist might ask, turning a blank canvas into a structured conversation. Positive user sessions showed that this direction addressed a real task rather than merely showcasing generation.

Watch from 2:01

Automated critique should preserve confidence

The assistant assesses a draft against concrete editorial criteria and shows what is working as well as what needs attention. The interface explains the issue, points to the relevant passage, and can propose a revision, but it avoids framing the author’s work as simply bad. This wording matters because professional feedback can be experienced personally even when it comes from software.

Watch from 8:01

Prompt quality depends on domain review

The first prompts followed common advice about assigning the model an expert role, yet a product marketer with public-relations experience found that this made the output more generic. Prompts then moved through design, implementation, testing, and repeated subject-matter review. The lesson is that plausible instructions are not enough; domain experts must judge whether the result is actually useful.

Watch from 10:00

Every refresh has a product cost

Allowing continuous editing created a difficult question: when should the product pay to reevaluate the draft? A suggestion applied through the assistant can trigger a clear refresh prompt, but ordinary manual changes are harder to classify because a full rewrite and a punctuation fix should not be treated alike. Token costs therefore shaped limits, notifications, and explicit user control.

Watch from 12:03

Adoption did not make the tool universal

New users adopted the assistant and users who created releases with it showed stronger conversion behavior, while a smaller share of paying users applied its recommendations. The team’s interpretation was that experienced practitioners may need less guidance. This supports a flexible design in which assistance can be hidden and the product does not assume every professional wants the same level of intervention.

Watch from 18:01
Explore WaysConf 2026

View the agenda

Get tickets