The first AI UI is usually wrong

Our first M&A-native AI interface was an "Assign to AI" button.

The idea was simple. A user would create or open a task, then assign that task to AI the same way they might assign it to a teammate. The system would analyze the relevant documents and produce an answer.

It had the shape of a product. It also had the wrong center of gravity.

The buyer did not want to manage AI labor. The buyer wanted an issues list.

The first AI interface usually exposes the machine's action instead of the user's desired work product.

That is the mistake to look for.

"Assign to AI" was a product smell

The button made sense because it mapped AI onto the workflow we already understood.

Colabra had projects. Projects had tasks. Tasks had owners. AI could become another owner.

That mental model was tidy. It was also inherited from the old product. Scientists and R&D teams had reasons to create, organize, and assign work item by item. M&A buyers had a different job. They did not want to compose hundreds of tasks before the system became useful. They wanted to upload or sync a room and get a structured first pass.

The old interaction model made the user supervise the AI before the AI had created value.

That is backwards.

In a high-volume document workflow, the AI should do the first sweep automatically, then ask the human to review, correct, escalate, assign, and export.

The human should spend judgment, not ceremony.

Chat boxes can make the same mistake

The chat box is the most common AI interface because it is flexible and familiar.

It is also overused.

A blank chat box asks the user to know what to ask. In diligence, that is often the wrong burden. The user may not know which files matter yet. They may not know which issues exist. They may not know which request-list items are uncovered. They may not know which workstream needs attention.

If the product's first move is "ask me anything," it has pushed discovery back onto the user.

Sometimes that is fine. For exploratory work, chat is useful. For repeated workflows with known outputs, chat is usually a secondary interface. The primary interface should be the work product.

For diligence, that work product might be an issues list, request-list coverage map, missing-item list, workstream report, or source-cited answer table.

The user can chat with the system after that. The first screen should show what the product already knows.

Design from the output backward

The better design question is simple:

What artifact would make the user's next meeting, review, or decision better?

Then work backward.

If the artifact is an issues list, the product needs document ingestion, classification, retrieval, severity calibration, citations, reviewer state, and export. If the artifact is request-list coverage, the product needs a request parser, file mapping, missing-item detection, partial coverage logic, and a way to show confidence. If the artifact is a workstream summary, the product needs folder context, stakeholder-specific views, accepted issues, open questions, and materiality.

The interface should reveal those objects.

This is why AI product design is usually less about inventing a new surface and more about finding the old work product that already carries trust.

In M&A, people already understand issues lists, trackers, request lists, red flags, summaries, and diligence reports. The AI product earns adoption by making those artifacts faster and better, then showing enough evidence that the team can rely on them.

The review step is part of the interface

A lot of AI demos end when the model produces text.

Real AI products start their hardest UX work there.

What can the user do with the output?

Can they open the source? Can they see the clause? Can they lower the severity? Can they reject the finding? Can they assign it to legal? Can they mark it reviewed? Can they ask why it was flagged? Can they see what evidence is missing? Can they export it into the format the team already uses?

Those controls are not polish. They are trust.

The user interface has to make correction cheap because correction is how the system becomes safe enough for real work. A product that cannot absorb human review will either overstate the AI's authority or force the user to redo the work somewhere else.

Both outcomes break adoption.

Magic is a weak default

Founders like magical demos. Buyers like clear control.

That tension matters in AI UI. The more expensive the failure, the more the interface has to show its work. Diligence is full of expensive failures: missed clauses, false-critical issues, privacy mistakes, clean-team boundary problems, unsupported conclusions, and confusing handoffs.

A magical interface hides too much.

A useful interface shows coverage, gaps, evidence, confidence, ownership, and next action. It gives the reviewer handles.

This does not mean the UI should be cluttered. It means the product should be honest about the workflow. If the AI has not seen a document, say that. If the citation is weak, show that. If the issue needs human judgment, make that visible.

Trust grows when the user can see the boundary.

The pattern I would use now

For an AI workflow product, I would design in this order:

1. Name the user's desired work product.
2. Identify the source material needed to produce it.
3. Define the state the product must preserve.
4. Generate a first pass automatically where the task shape is clear.
5. Put human review directly beside the evidence.
6. Capture corrections as structured feedback.
7. Export into the format the team already uses.

The interface may still include chat. It may still include manual actions. It may even include an assignment model later.

Those should support the workflow. They should not be the workflow.

The broader lesson

The first AI UI is usually wrong because it starts from the model.

What can the model do? What button invokes it? What prompt should the user type? What animation makes it feel intelligent?

The better starting point is older and less glamorous.

What job does the user already have? What artifact do they trust? What evidence do they need? What control do they require? What would make the next human decision better?

Build from there.

AI should reduce the distance between messy inputs and trusted work.

Everything else is theater.