On this page
Applied AI, screenshot-led documentation and a practical improvement to software-release workflows.
I joined Zegeba in Norway as an AI Product Manager in August 2025, bringing with me experience in product development, systems architecture, user experience and technical documentation.
One problem that interested me was familiar to anyone involved in maintaining a substantial software product: the work required to keep customer-facing documentation aligned with changes in the application.
A release changes more than code. It can affect screenshots, explanations, instructions, support materials, release notes and the questions customers are likely to ask.
Each of those outputs may require attention, and inconsistencies can accumulate when they are prepared separately.
I designed and developed an AI-assisted documentation workflow to reduce repetitive preparation, improve consistency across release materials and keep people responsible for the final information.
A recurring operational problem
Product documentation is rarely a one-off exercise.
Features evolve, interfaces change and release information must be communicated to people who have very different needs. A support colleague may require a concise explanation of a change. A customer may need revised instructions. A release manager needs to describe what is different and why it matters.
All these materials should reflect the same underlying product behaviour.
When the information is maintained in separate places, the effort of keeping it consistent grows. People also need to spend time checking wording, terminology and presentation.
For an organisation preparing English-language materials in a multilingual working environment, the risk of avoidable inconsistencies in language adds another practical consideration.
My interest was in improving the process rather than simply generating more text.
Start with what is already known
I developed an AI-assisted documentation and product-enablement approach built around an organised knowledge baseline.
Instead of repeatedly asking an AI system to describe a feature from scratch, the workflow began with the product information and established explanations already available.
Screenshots formed part of the process. They provided a practical way to examine visible product behaviour and changes, supporting the comparison between an existing application and a newer release.
The aim was to identify what had changed, consider which material was affected and produce consistent proposed updates across several types of documentation.
These included knowledge-base articles, release notes, frequently asked questions and support-related materials. I also explored ways of preparing content for video and other explanatory formats.
The important architectural decision was to treat the existing product knowledge and human review as authorities, with AI assisting in analysis, organisation and drafting.
An AI-generated description is not proof of how a product actually behaves. It must remain subject to review against the software and the information maintained by the people responsible for it.
A workflow rather than a collection of prompts
I was interested in making the approach repeatable.
That meant thinking about how source material was organised, how changes were identified, how terminology and tone were preserved, and how the resulting outputs could be prepared in useful, implementation-ready formats.
Consistency mattered as much as speed.
A release note, a support explanation and a knowledge-base article may differ in length and purpose, but they should not disagree about what the product does.
The approach was designed to reduce the repeated effort of rediscovering the same information for every output.
It also offered a way to reduce avoidable language and formatting problems, while preserving the organisation’s established terminology and voice.
Moving beyond an experiment
I began with a practical prototype and developed the approach into a repeatable process used to help prepare materials for a subsequent production release.
The workflow was used in a live release-preparation process, connecting existing product knowledge, screenshots, documented changes and review.
My manager reported substantial savings in documentation preparation time. This was practical feedback from the work, not a quantified time study.
The result was a repeatable workflow used in a real software release, making documentation easier to prepare consistently and reducing repetitive work for the people involved.
That is the sort of applied AI improvement I find interesting.
The human part remains essential
I don’t regard this as an argument for publishing AI-generated documentation without checking it.
Product knowledge has context. A screenshot can omit behaviour that matters. A change can have implications that aren’t obvious from the interface. A generated explanation may sound plausible without being correct.
Those limitations make review essential.
The goal was to make the work easier to perform well, not to remove the people who understand the product from the process.
In that respect, the project was consistent with the way I’ve approached other systems: reduce unnecessary effort, preserve the important decisions and make the outcome easier to examine.
What I took from the work
The technology used to assist documentation will continue to change. The operational problem will remain recognisable.
Software products need clear, accurate and current explanations. Preparing those explanations should not require the same information to be manually reconstructed for every audience and format.
The Zegeba work showed me how AI could support a repeatable process built around existing knowledge, visible product changes and human validation.
It also reinforced my preference for judging new technologies by their contribution to a real problem, rather than by the novelty of the technology itself.
Useful AI is not merely the ability to produce an answer. It is the ability to support a dependable process in which people can understand, verify and use the result.
Project context
Research programme: From intention to implementation: human-directed AI-assisted engineering
Topic: AI-assisted engineering
Explore further
Further reading is listed below.