Quando conversar com a IA dá mais trabalho

When talking to AI creates more work

You want to change a delivery address. You open the app, find your order and are greeted by a chat box. “How can I help?”

So far, so good. You explain what you need. The system asks for the order number, asks which address to use and returns a summary for confirmation. That might do it. Or you might need to explain that the new address applies only to this purchase.

The task was small. The conversation grew.

I’ve been thinking about this whenever I see proposals to add AI to digital products. A text box looks simple: there is little on screen, it accepts everyday language and it promises to understand what someone needs. But some of the complexity may simply have moved out of the interface and into the user’s head.

The work of explaining

A screen with options shows you some of what is possible. There is a field for the address, another for additional details and a save button. It might be poorly organised, of course. But there is a structure that helps you understand the task.

That structure is not always visible in a conversation. The person typing has to decide what to include, how much context to give and how to check whether they were understood. They also have to discover the limits: can they change the delivery? Cancel part of it? Restore the previous address?

That freedom is useful if you know how to phrase your request. If you are still trying to work out what you can do, it can feel like a very polite blank page.

This strikes me as an important design problem. An interface that looks cleaner may take more effort to use. Counting fields or buttons does not capture that difference.

Some conversations help us think

If I need to make sense of an idea that is still unclear, a conversation can be a great place to start. I can explore alternatives, change the question and realise along the way that I was looking at the wrong problem.

Imagine someone planning a trip. They have not chosen a destination, have a limited budget and will be travelling with someone with reduced mobility. A conversation can help organise those constraints and uncover possibilities that an initial set of filters might not reveal as well.

Now imagine that same person, after booking, wanting to download their confirmation. At that point, a button could make their life much easier.

The product is the same. The need has changed. So have their willingness to chat, their sense of urgency and how much they already know about the result they want.

That is why I am less interested in debating whether the future of interfaces is conversational than in understanding when conversation helps. We can have a useful conversation during discovery and a direct action when it is time to get something done.

What happens after the answer?

There is another part that tends to receive less attention: checking what was done.

If AI organises a task list, how do I correct just one item? Do I have to write another message and hope it leaves the rest alone? If it prepares a purchase, can I compare the items before confirming? If it changes a setting, is it clear how to undo it?

When everything happens through text, even a simple correction can become a negotiation. “Keep the first one, change the second, but use the condition I explained earlier.” By then, a small editable table starts to look like a rather advanced invention.

A combination seems more promising to me: conversation helps shape the request, while the interface presents the result in a way that lets people act on it. Fields to edit, options to compare, confirmation before an important action and a clear way to undo it.

These are familiar features. They remain useful when AI is part of the process.

Autonomy includes choosing how to get there

Not everyone will feel comfortable typing a request. Some people prefer speaking, some need visible options and others simply want to repeat a task they already know.

A conversation can open up possibilities for some people and create difficulties for others. It depends on the language, the device, the conditions of use and how the alternative was designed. Voice, for example, helps in some settings but may be impractical somewhere noisy or when the subject is private.

I would avoid making conversation a compulsory step. When someone’s intention is already clear, let them act directly. When they need to explore, help them formulate and review their request.

That choice also shows respect for the user’s time.

Whose effort are we reducing?

When evaluating an experience like this, I would start with a specific task. I would ask someone to complete it and observe where the work appears: understanding what to ask, explaining the context, waiting for the answer, checking the result or correcting an error.

The number of messages alone tells us little. A long conversation can be a valuable exploration. It can also be someone trying, for the fourth time, to get the system to understand a simple instruction.

What matters is whether the person got what they needed, understood the result and could correct their course without another blind attempt.

AI greatly expands what we can build. That is precisely why I think we need to be more deliberate about where we use it and how we present its possibilities. Sometimes its best contribution will be helping someone work through a difficult decision. At other times, it will be organising what happens behind a very simple action.

The person who only wanted to change their delivery address will probably be happy if they can do that and get on with their day.

Photography: Tim Mossholder / Unsplash.