Teslim Bello, specialist in human-artificial intelligence interaction, leads product design at ContextQA. His work focuses on how AI behaves when people need to trust, challenge, and correct its outputs. In this interview by KINGSLEY ALUMONA, he speaks about his work in AI, among other issues.

Tell us about your work in Artificial Intelligence (AI) and tech in general.

I am a product designer specialising in artificial intelligence, specifically human-AI interaction: designing how people interact with, inspect, and stay in control of AI systems. I currently lead product design at ContextQA, an agentic AI software-testing company based in the US.

My path into AI was a bit unusual. I studied Computer Science, worked as a product designer for the Lagos State Ministry of Science and Technology, and later designed products across real estate, hospitality, and fintech. I then spent several years at Anchor Group building healthcare operations systems for care agencies in New York before moving fully into AI design in 2024. Later on, I moved to the UK for a master’s degree in User Experience and Interaction Design, where my research explored how humans and AI work together. Everything I do sits at that intersection between people and intelligent systems.

Companies are rushing to add AI to everything. What is the biggest mistake they are making?

They start with the technology instead of the task. They add a chatbox, count how many responses it generates and call that adoption. But volume is a weak metric. It tells you the system produced something, not that a person accomplished anything. The better questions are: Did users complete the workflow? Did they come back? Could they inspect the evidence, correct mistakes and recover from failure? Did the feature actually improve a real decision? The interface also has to communicate uncertainty honestly. The goal is not to make AI appear smarter. It is to make the product more dependable.

What do you mean when you say AI should “show its work”?

I do not mean exposing the model’s private reasoning or overwhelming people with technical detail. I mean providing the practical evidence people need to judge an output: which records were analysed, what time period was used, where the source can be inspected, and whether important information was missing. Fluency is not the same as reliability. An answer can sound certain while still being incomplete. In a professional product, users should be able to move from the summary to the underlying evidence and clearly distinguish retrieved facts from the system’s interpretation. Confidence without inspectability is not trust. It is presentation.

Is AI making product designers less important?

It is doing the opposite, although it is changing the job. AI dramatically reduces the time needed to produce screens and prototypes. What it does not replace is judgment: deciding what should be automated, where uncertainty must remain visible, when a human should intervene and how people recover when the system is wrong. If anything, AI raises the cost of poor product decisions because AI features can act, not simply display information. That is also why the boundary between design and engineering is becoming less rigid. I prototype directly in code now so teams can experience AI behaviour before building it. The designers who matter most will be the ones who understand technology well enough to design for failure, not just for demonstrations.

What opportunity does AI create for Nigerian and African designers?

A significant one, and probably bigger than many people realise. Our environment trains us to design for unreliable connectivity, cost sensitivity, multiple languages and different levels of digital familiarity. Those constraints force exactly the questions AI products increasingly need to answer. Can someone understand what happened? Can they continue after an interruption? Does the system remain useful when conditions are not ideal?

Those are not local problems anymore. They are global design challenges. The field itself is still young. Almost nobody has a decade of experience designing AI interfaces, which means the opportunity is genuinely open. My advice is simple: choose difficult problems rather than template work. Publish your thinking because talent that stays silent often stays local. And contribute publicly through open source, communities and mentoring.

What principles guide you when designing AI that people should be able to challenge?

Four principles: provenance, boundaries, correction and recovery.

READ ALSO: Artificial Intelligence: Will Nigeria embrace or ignore it?

Provenance means users can see where an output came from. Boundaries mean the product is honest about what the system did and did not examine. Correction means people can fix an interpretation or add context without restarting the task. Recovery means failure leads to a useful next step rather than a dead end. I have also published an early open-source checklist and audit tool that turns those ideas into practical review questions.

One thing I feel strongly about is that “human in the loop” cannot simply mean adding an approval button after the system has hidden its assumptions. Control only becomes meaningful when people have enough context to make an informed judgment.

AI is now appearing as a chatbot. Is chat enough?

Chat is a good starting point because people can express intentions in ordinary language. But most professional work does not end with receiving a paragraph of text. People need to compare information, inspect records, edit details, share conclusions and take action. When every complex task is squeezed into a conversation, important structure disappears. A user can receive an impressive answer without knowing what information was examined, what was excluded or how to correct the result. The stronger model combines conversation with structured reports, editable workspaces, visible sources and meaningful controls. Chat can open the door, but it shouldn’t become the whole building.

How has your work in AI-assisted software testing shaped that perspective?

Software testing is an excellent training ground because every claim can be checked against concrete evidence: test runs, requirements, failures and historical records. If a system says a release carries risk, you should be able to trace that claim back to the underlying data. At ContextQA, I design AI-assisted workflows across a platform teams use to create, run and understand software tests. The biggest lesson has been that generating an answer is only the first layer. The product also has to help people understand what happened, judge whether the output is useful and decide what to do next.

You recently worked on natural-language reporting. What problem were you trying to solve?

Traditional dashboards work well when you already know which metric you’re looking for. But many real questions don’t start that way. Someone simply wants to know why a test run failed, which tests are unstable or how performance changed over the month. I worked on a dynamic reporting system where users begin with a question, and the product pulls live data into a structured answer. The important design decision was that the report could not just be polished prose. It had to make its scope visible, separate data from interpretation and provide a path back to the underlying records. An AI report should not be a decorative summary floating above the product. It should be an inspectable layer connected to the work itself.

You have conducted extensive user research. What has it taught you about trust?

People rarely talk about trust in abstract terms. You see it in behaviour. They hesitate because they do not know what an action will change. They search for information the interface has hidden. Sometimes they accept an AI response while holding a completely inaccurate mental model of what the system actually did.

In research sessions, I am less interested in whether someone likes a screen than in whether their understanding matches the system’s behaviour. Those insights have shaped navigation, onboarding and how I have organised complex information. The value of research isn’t the transcript. It is the product decision that follows.

What is next for you?

I will continue leading design on AI products, expanding the open framework and publishing more of my thinking. I am also interested in having more conversations with international product and technology communities about this work. Long-term, I want to build a company of my own in this space.

I have spent years studying how humans and AI should work together, and I still have ideas I want to build. Whatever form that takes, the mission stays the same: making artificial intelligence something people can genuinely trust and control.