More capacity. The same rigor.

As infrastructure, it sits behind the work and never shows up in the interface. It summarizes a round of interviews, writes down a system exception so engineering does not depend on my memory, and drafts the note in which business and design see the same decision. The point is to reach the decision sooner, and for the team to share the criterion.

As a design material, AI is what the person uses. Without it the product does not exist: there are no transactions to classify, no contract to flag, no text to draft from an offer. Designing there means deciding what the model does, what the person confirms, and what happens when the model does not know. The projects below sit in this second category.

It comes in when I frame the question, not once the deliverable is already closed. I use it to organize what I heard, to test whether an argument holds, and to get a first version of a text, a flow, or a specification.

What I do not hand over is judgment. What stays, what gets cut, and what cannot be claimed is my decision. Leave it out of that path and the mechanical work takes back the time that should go to the decision.

Research synthesis, copy iteration, design system documentation, and stakeholder communication. In research I move from loose notes to patterns without losing the quote that supports them. In copy I try several forms of the same argument until one can be defended in a room. In a system, components, exceptions, and criteria get written down so engineering can build them without reinterpreting them. With stakeholders I prepare the decision before the meeting: what is proposed, what is left out, and what evidence supports it.

The result is that judgment and attention go to what actually adds value. AI shortens the path to a version that can be discussed. Deciding whether that version is the right one, and what it must not say, is still design work.

There are two positions worth distinguishing. The first is AI as an intermediate layer: it works between systems or processes, reduces operational friction, and brings forward decisions that currently take human time. It works off the screen the person uses: it reads what used to be read by hand, orders the signal, and leaves the action where it already was.

The DORA platform is a clear example. The AI does not replace the auditor: it compresses the time between a risk signal and a human decision. The model flags the risk in a contract and puts a traffic light ahead of the detail. The auditor validates and signs. The detail arrives after the light, not before. Designing that layer means deciding what gets automated, what has to be shown in order to trust it, and what stays outside the model because a signature cannot be delegated.

The other position is a product with AI at its core. In Contable AI Pro, the person uses a solution that would be impossible without it. There is no earlier process for the AI to speed up. The product is the classification: thousands of raw transactions become a dashboard because the model understands the context of each one.

What I learned by building it is where the model has to stop. It does not reason from scratch every month. Each real error becomes a rule, and what it does not know it leaves blank for a person to complete. The next month, that correction is already precedent. The mechanical part is handled separately. Only the ambiguous cases reach the conversation.

Telling which position a product needs, and designing the interaction between the person and the AI, is one of the most interesting design problems of this decade: deciding who acts, who confirms, and what happens when the machine does not have enough signal.