How AI is changing requirements engineering

Artificial intelligence is gaining ground in many areas of business. Unsurprisingly, it is also rapidly changing requirements engineering. Today, it supports tasks such as analysis, structuring, formulation and quality assurance of requirements, thereby taking over activities that were previously performed entirely manually. At the same time, the question arises as to how reliable these results are and under what conditions AI can really prove its worth in requirements engineering.

This is precisely what I will examine in the following article.

Why AI requirements are compelling and where to look closely

For many years, I have been intensively involved in requirements engineering and the use of AI in this field. As a manufacturer of relevant software, a regular speaker at the IREB [1] Special Interest Group AI, and a supervisor of two master’s theses on this topic, I gain broad insight into current developments and practical challenges. In doing so, I repeatedly encounter aspects that are surprising in practice or cause discussion.

This becomes very clear when evaluating AI-generated requirements. In recent studies, they consistently performed better than those formulated by humans. Developers rated the acceptance criteria of AI as superior in all four dimensions examined:

• readability,
• comprehensibility,
• delineability and
• technical accuracy.

The differences were evident not only in quality but also in speed. One participant in a study summed it up aptly: ‘We’re talking minutes versus hours.’ [2] Overall, around 80 per cent of AI-generated requirements were accepted, while those formulated by humans came in at 54 per cent. Depending on the task, efficiency gains ranged between 50 and 80 per cent. [3]

One key reason for this difference is consistency. Humans write creatively, but often unevenly, imprecisely or with their own interpretation of structure. Criteria formulated by humans therefore often show variations in structure, style of wording and linguistic quality. AI, on the other hand, provided that sufficient examples are given, reliably adheres to defined patterns, thus ensuring a level of structure and consistency that is difficult to achieve consistently in everyday life. Simple AI-based quality control also prevents many avoidable errors, such as obvious spelling or grammatical mistakes.

Despite these advantages, one point remains important. Language models often generate texts that appear clear and convincing, but whose content may be questionable or inaccurate. The critical eye of a human being is therefore still essential.

Why AI in requirements engineering absolutely needs context

Just because AI can formulate requirements well does not mean that it can reliably grasp the technical context. Requirements thrive on context, and this is precisely where the greatest weaknesses become apparent in practice. If the context is missing, unclear or lost in the workflow, results quickly become imprecise or mix information that does not belong together.

This is particularly evident in widely used systems such as ChatGPT. It is representative of all generative AI models because it is frequently used, easily accessible and serves as an informal benchmark in many teams for formulating initial requirements or acceptance criteria.

The central problem with working manually with ChatGPT & Co. is the lack of context or imprecise context. Without reliable contextual information, even the latest model cannot produce consistent results. To get around this, you would have to copy the entire product documentation, the architecture description and all relevant requirements into the chat again for each new user story. Obviously, this is more than impractical and quickly leads to further difficulties.

If you write story after story in the same chat, the AI can easily become distracted. It takes details from previous stories, mixes up information or ignores parts of the current task. The longer the chat becomes, the more pronounced these effects are. At some point, the model responds more slowly or not at all. The reason for this is the way large AI models work. Since they have no memory, the entire chat history, including all attachments, is reprocessed with each individual request. [4]

One might assume that large context windows would solve this problem. Modern models such as Google Gemini or Claude Sonnet can process a million tokens, which is more than all three Lord of the Rings books combined. [5] And indeed, the needle-in-the-haystack problem has now been largely solved. AI can find specific hidden information in long texts. In the reverse direction, however, many models continue to reach their limits. This so-called reversal curse means that AI can find individual requirements such as ‘The system must be GDPR-compliant,’ but cannot reliably identify all the stories that are affected by this. [6]

The technical background is well known. The advances made in recent years are based on the mechanism described in the paper ‘Attention is all you need’. [7] In this process, each token is compared with every other token. This increases accuracy, but leads to a quadratic increase in computational load. The longer the text, the more difficult it becomes for the model to work reliably.

For requirements engineering, this means the following. Large AI models can only reliably assess whether requirements meet defined criteria up to a certain text length. A global question such as ‘Do all stories meet the INVEST criteria?’ may seem practical, but it quickly leads to inaccurate results when many stories are processed simultaneously. [8] Models perform significantly better when the same question is asked individually for each story or epic. This significantly increases the hit rate because the AI evaluates the respective content separately instead of having to process many objects at once.

The search for the right tool in requirements engineering

When requirements need to be precisely formulated, clearly structured and thoroughly checked, it’s not just AI and methodology that matter, but also the tools. Many teams reflexively reach for Word or Excel when they’re not using ChatGPT or similar systems. These tools are ubiquitous and familiar, but they are only partially suitable for requirements engineering.

Word was not developed for requirements engineering. It does not recognise user stories or epics, cannot track semantic dependencies and does not offer practical link management. Even small changes can shift the entire layout and cause unnecessary rework. Furthermore, Word does not synchronise with project management or development tools.

Excel often allows requirements to be structured better, but this also presents new challenges. Tables quickly become confusing, difficult to read and are only suitable to a limited extent for communicating content to stakeholders or customers. At some point, the question also arises: what happens after Excel? After all, requirements must be transferred to systems that do more than just manage data in cells.

Word and Excel are one of the main reasons why requirements engineering appears chaotic, inconsistent or error-prone in many organisations. Not because requirements engineers do a poor job. And not because Word or Excel are bad tools. They were simply developed for other tasks and can only map the specific requirements of requirements engineering to a limited extent. Even complex, formula-based Excel documents therefore only achieve a fraction of the traceability, structure and support that a professional RE tool offers.

So which tool could be a good fit?

Conclusion

AI has arrived in requirements engineering. It formulates requirements quickly, consistently and often more clearly than humans can in everyday life. Research and practical experience show that many teams benefit from this: less manual routine work, fewer inconsistent formulations and more structure in their daily work. AI does not change requirements engineering, but it does make it more precise and efficient.

At the same time, it is clear that advantages only arise if the technical context is clearly defined. Models such as ChatGPT are powerful when it comes to handling text, but their architecture reaches its limits when too much information has to be processed at once or when inferences across many objects are required. Unfortunately, mixing, omissions or incorrect connections are the result of how large language models work.

This puts the tools more in the spotlight. Office applications are not designed for working with complex requirements. They offer neither structure nor traceability nor clear context management. This circumstance inevitably draws attention to specialised solutions. Here, too, context is decisive: if you work in a regulated environment and have been using established tools for years, you will probably not replace them. However, if you have the freedom to redesign or specifically improve your requirements engineering, it is worth looking at tools that use AI to help you formulate requirements better, version them more easily and control them more reliably throughout the entire process.

Notes (partly in German):

Are you interested in talking to Simon Jiménez? Or would you like to try out the storywise tool? Then simply write to him on LinkedIn or visit his excellent website.

[1] IREB stands for International Requirements Engineering Board. It is a non-profit organisation that establishes professional standards in requirements engineering.
[2] ‘Using artificial intelligence to create effective and efficient software offerings’ (Knuplesch S., FH Campus 02, 2024)
[3] Improving acceptance criteria with AI – product data as a knowledge base for higher requirements quality‘ (Kainer P., FH Campus 02, 2025, publication 2026)
[4] This also explains why even a short message like a simple ’thank you” requires computing power. To generate the response, the model must first analyse the entire chat history.
[5] AI News 24: Claude Sonnet 4: 1 million token context – the new AI memory for developers?
[6] arxiv: Towards a Theoretical Understanding of the “Reversal Curse” via Training Dynamics
[7] arxiv: Attention Is All You Need
[8] What is the INVEST principle?

Here you will find an article about the changing role of software engineering in the AI age.

Would you like to discuss the possibilities of AI in requirements engineering as a multiplier or opinion leader? Then share this article in your network.