As artificial intelligence systems become more capable and autonomous, product designer Bello Teslim Olasubomi believes the industry’s greatest challenge may no longer be building more powerful models, but designing products that enable people to understand, verify and confidently use AI-generated outputs.

While much of the AI industry focuses on improving model performance, Bello argues that the user experience surrounding those models plays an equally important role in determining whether people ultimately trust the technology.

“The AI industry measures progress in capability, but a more immediate question receives far less attention: what happens when the system is wrong?” he said.

Bello leads product design at ContextQA, an AI-native software testing platform where artificial intelligence assists engineering teams in generating, executing and maintaining software tests. His work has given him first-hand experience of designing professional products in which AI-generated outputs directly influence critical workflows.

According to him, many AI products are still designed as though machine-generated responses are always correct.

“Most AI products are still designed as if the model is always right,” Bello said. “But these systems are non-deterministic. They can be confidently wrong, incomplete or inconsistent. If the product pretends otherwise, users either over-trust the output and get burned, or they lose confidence and stop relying on it. Both are product failures.”

Rather than viewing trust solely as a characteristic of AI models, Bello argues that trust is largely shaped by product design.

He believes users need interfaces that clearly communicate what an AI system has produced, how it reached its conclusions and where uncertainty may exist. Just as importantly, people should be able to review, edit and correct AI-generated work without restarting an entire workflow.

“The goal is not to make the AI look infallible,” he said. “The goal is to make it legible. A person should be able to understand what the system produced, judge whether it is right and correct it without having to start again.”

Bello describes this philosophy as designing AI as infrastructure rather than spectacle.

That principle informed his work on ContextQA’s AI-assisted test creation workflow, where he identified a recurring problem involving how user-provided information flowed through the system.

Although users supplied important instructions early in the process, that context was not always fully reflected in the AI-generated output, even though the final result appeared polished and complete.

“The more difficult failure is when the output looks intelligent but has quietly dropped something important,” Bello explained. “If a user gave the system a constraint three steps earlier and the final output ignores it, the product has failed even if the language looks perfect.”

To address the issue, Bello redesigned the workflow so that important user context remained visible throughout the process. He also introduced opportunities for users to review what the system had understood before generating the final output.

Instead of limiting the redesign to interface mock-ups, he developed an interactive prototype that allowed product and engineering teams to observe how users, AI-generated content and system behaviour interacted across an entire workflow.

According to ContextQA, the redesigned experience was later deployed into production.

The company said monthly active use of the AI-assisted test creation workflow approximately tripled in the months following the redesign. Users also increasingly reviewed the system’s understanding before generating tests, reducing instances where important context was discovered to be missing only after completion.

“It made the generated output something people were willing to trust and act on, which is the hard part of any AI feature,” the company said.

For Bello, the experience reinforced a broader principle that extends beyond software testing.

He argues that human review should not be viewed as a safety mechanism added after AI systems have been built, but as an integral part of the overall product experience.

“People sometimes talk about human-in-the-loop systems as if the human is a safety feature added on top,” he said. “I think that is backwards. In many professional AI products, review is part of the job the product is helping the person perform. You should design for that from the beginning.”

Bello’s approach reflects his interdisciplinary background. He holds a Bachelor’s degree in Computer Science and a Master’s degree in User Experience and Interaction Design from Glasgow Caledonian University.

He says this combination has shaped how he thinks about the relationship between system behaviour and user interaction, particularly as AI products become increasingly sophisticated.

Rather than relying solely on static interface designs, Bello frequently develops coded and interactive prototypes to examine how trust develops over multiple interactions.

“With AI, you are designing a relationship between the system’s behaviour and the user’s judgement,” he said. “A flat frame can show what a page looks like. It cannot always show whether the system earns trust over five or six decisions.”

He believes these questions will become even more significant as AI systems are increasingly deployed across sectors such as healthcare, finance and enterprise software, where inaccurate or misleading outputs could have serious consequences.

Although advances in model capability will continue to reduce certain categories of error, Bello remains unconvinced that better models alone can eliminate the need for thoughtful product design.

“Trustworthy outcomes come from giving people the ability to verify and steer what the system does,” he said. “If an AI product cannot be checked and corrected by the person using it, raw capability will only take it so far.”

He argues that organisations building AI-powered products should treat the human control layer as a fundamental architectural component rather than an interface refinement added late in development.

“The interface is where trust is won, lost and repaired,” Bello said. “The strongest AI products will not be the ones that pretend the model never fails. They will be the ones that help people understand the system, recognise when something needs attention and stay in control of the outcome.”