What work gets finished with your new AI tool?

What do we want done when the build is finished?

It is easy to start with the possibility. We can connect systems, generate material and build a new view of the numbers. In the episode, my first question is where we want to get to. Which task should improve? What is the result of the work?

There needs to be direction before the next prompt. Otherwise there is almost no limit to what can be built, but no clear basis for saying something is finished either.

Apply the same requirements to AI work as to a product launch: who will use the result, what needs to work around it and who decides it is ready? That helps define what the build needs to include.

A page can make sense in code and none to a person.

Fredrik describes receiving dashboards made of boxes inside boxes, lots of text and several prominent buttons. The information is there, but it is hard to see what matters or what the user should do first.

That is a useful distinction. Structured data does not mean a presentation that helps someone. A technically functioning page can still place a heavy burden of interpretation on the person at the screen.

In his reasoning, design rules and concrete examples therefore matter. What makes a clear hierarchy? Which action takes priority? Which information needs showing immediately, and which can wait until requested?

An explanation may sometimes be necessary. But if every part of an interface needs long instructions, the design itself needs testing too. Otherwise we have built something that looks finished but does not yet help the user.

Count the route to something useful, not just the cost per step.

We also discuss the value of a specialist tool built on a general AI model. The simple objection is that the same model can be used directly. But the user also pays for the context around the model.

That can mean instructions, a clearer interface and several connected parts of the workflow. If reaching a useful result takes fewer retries, it can make a difference even when an individual run costs more.

Ask how much work remains after the model’s answer. Who needs to change the material, connect the parts and check the delivery? Include that time when comparing a specialist tool with a general model.

AI can keep you busy with the wrong things too.

Another part of the conversation concerns focus. Fredrik describes days when working with AI felt intense but the original task still remained unfinished. The tool had found more questions, more errors and more possible improvements.

Each detour could sound reasonable. Together, they took over the day. That is a concrete problem to recognise when building your own solutions: more suggestions also mean more decisions for the person.

The task and priority therefore need to stay visible. What are we trying to complete now? Which errors need solving for it to work? Which ideas can be put aside?

Let the chosen task lead. A new error or improvement suggestion only needs attention now if it prevents delivery. The rest can wait until what you started is finished.

An automated flow needs clear decision points.

Later in the episode, I describe work in a new poster company. At the time of recording, the team has built a flow from trend research and new material to drafts of product information and ads. The drafts are reviewed before publication.

It is worth pausing on the word drafts. A system producing quickly does not mean it should publish everything. Someone still needs to assess how the material looks, whether it fits the range and what customers should actually see.

The same goes for follow-up. We discuss using e-commerce sales data and separating it from what an ad platform claims. They are different questions: what the platform reports and which products actually sold.

More production also creates range-management work. If new designs arrive constantly, someone needs to decide what stays, what goes and how the customer finds her way among the options. Otherwise automation just builds a larger problem faster.

The next market exists outside the screen too.

The conversation does not end with everything now being automatic. We discuss quality, rights and what would be required for a local printer to serve customers in other markets. Those questions need real owners and evidence.

Someone needs to choose production and delivery arrangements and check that the customer receives what was promised. Put that responsibility in the plan even as more content work becomes automated.

Then try the flow from new design to reviewed product and order. Note where the team needs to intervene and what the customer sees. That gives us a more useful list to work on than yet more possible AI features.