When we talk about artificial intelligence, we almost always end up talking about productivity: how many hours it saves, how many tasks it automates, how much code it can write, how many processes it runs without human intervention. That way of measuring is easy because it rests on numbers.
There is, however, a far more interesting way to assess its impact: asking not how much faster something can be done, but what it lets someone do that they could not do before. At that point the conversation changes completely.
The World Health Organization estimates that around 1.3 billion people, roughly 16% of the world's population, live with a significant disability. Many of the barriers they face do not come from the disability alone, but from the environment. A staircase is an obstacle if climbing it is the only way into a building; a visual interface is one if there is no other way to interact with it; and a written explanation can be perfectly accessible for one person and unnecessarily hard for another.
Until now, most accessibility technology has been dedicated to compensating for those barriers.
AI opens up a different possibility: the system adapts to the person.
A simple example is Be My AI. For years we have had software that recognises objects or reads text through OCR, but multimodal models can go further: they interpret a scene and hold a conversation about it. A blind person can take a photograph and ask what is in front of them. The difference between getting an answer like "There are three products on a table" and being able to ask "Which of them contains nuts?" or "Where is the power button on this device?" is enormous. The first only recognises information; the second uses it, and that is autonomy.
The same idea applies to transcription, voice generation, text simplification and any tool that changes how we present information. Where the impact gets especially interesting is in the cognitive load we take on when using everyday tools.
The cognitive load nobody measures
Think about how much enterprise software is built:
- Remembering what is still pending.
- Prioritising tasks.
- Not losing context.
- Understanding a quickly written instruction.
- Finding the right document.
- Remembering where you left something before you were interrupted.
For some people that process takes very little effort; for others it consumes a considerable amount of energy. This is where AI can do something traditional automation never handled well: turning a meeting into decisions and concrete next steps, breaking an ambiguous task into five specific actions, or holding on to the context of something you left half-done and bringing it back hours later.
We are not inventing the concept. We have practised cognitive offloading for centuries through diaries, notes, calendars, alarms and lists that take part of the work off our memory. The problem is that all of those tools require us to remember to use them. AI can start taking on part of that maintenance.
Adapting information, not replacing learning
There is already research exploring how to genuinely adapt information. A study published in Frontiers in Education (2026) used AI-generated visual explanations to work on reading comprehension with three children aged 9 to 11 with dyslexia. The sample is tiny and the results cannot be generalised, but the idea is far more interesting than the usual "AI for studying": it does not aim to replace learning, but to change how the same content is presented.
Another review, also from 2026, looks at the use of AI systems within Universal Design for Learning, with the goal of offering different ways of accessing the same knowledge instead of creating separate education for particular groups.
That leads to a powerful design question. For decades we have built products with an average user in mind and then added an "accessibility" section. What happens if we invert the process? An application could show a whole task to someone who wants to see it at once and split it into steps for someone who prefers to work that way; it could rephrase an explanation, strip out secondary noise, change the presentation format or restore context when it detects that context is missing. This is not about switching on dark mode or increasing the font size; it is about adapting the interaction.
Subtitles are essential for deaf people, but we all use them when watching a video on the metro. Designing for different needs usually produces better products.
Dictation removes a motor barrier and is simply convenient when your hands are busy. History shows that many of these solutions end up benefiting everyone.
Technical capability is not real impact
Something similar happens in medicine. There are AI systems today that analyse medical images, triage studies or support clinical decisions; the FDA maintains a list of AI-enabled medical devices, particularly in radiology. But it is worth separating technical capability from real impact. A study published in npj Digital Medicine (2026) evaluated a commercial intracranial haemorrhage detection system across more than 100,000 scans in 17 healthcare centres. That kind of evaluation matters more than a benchmark, because the real exam starts when the model leaves the lab and meets real patients, real professionals, imperfect processes and cases that may look nothing like the training data.
We should apply the same standard to many other AI applications: not only asking whether the model has better metrics, but what changes for the person using it. Can they do something they could not do before? Do they need less help? Do they understand something that used to be out of reach? Has a barrier been removed?
Automation will remain one of the great applications of artificial intelligence, but it would be a shame to measure this technology only in hours saved. Some of its best applications may not be the ones that let us do more things, but the ones that let more people do them.
References
- World Health Organization. Disability and health.
- Be My Eyes / OpenAI. Multimodal AI use cases applied to visual accessibility.
- Alsamani, O. A. & Alsamiri, Y. A. The effects of AI-based visual instruction on the reading comprehension of students with dyslexia in Saudi Arabia: a single-case experimental study. Frontiers in Education, 2026.
- Ram, A. M. et al. The AI scaffold and engagement spectrum as a novel UDL aligned system for supporting students with dyslexia. Education and Information Technologies, 2026.
- U.S. Food and Drug Administration. Artificial Intelligence-Enabled Medical Devices.
- Real-world performance evaluation of a commercial deep learning model for intracranial hemorrhage detection. npj Digital Medicine, 2026.