Case study · Self-initiated concept

Predict.Control.Verify.

How would you make people trust an AI coworker with real work? A concept for “Nomu”, an AI teammate that lives in team chat, built from public signals and a clickable prototype.

Product designProduct managementAI UX

The problem

Handing work to a colleague is easy. Handing it to an AI is not.

People do not delegate real work to an AI coworker because they cannot tell what it will do, what it will cost, whose data it will use and what it will remember.

A human colleague says “ok, I’ll do it like this”, tells you when something blocks them and remembers your correction. In chat, an AI often does none of that visibly. The concept turns those four gaps into four design principles.

Public signals

What people say about AI coworkers, and how much to trust it.

I did not interview anyone for this concept. I read public material: research write-ups, security research, product reviews and community threads. Each signal is rated by how strong the evidence is.

What this is not. There were no user interviews and there is no usage data. Several sources are vendor blogs or review sites that may earn commission, and some community comments come from people who work at competing products (they said so themselves). I treat all of it as a hypothesis to test, not as proof. The “Validation plan” below says how.

Product thinking

Who it is for, what they hire it for, and what I left out.

Assumed user (not interviewed)

Marta, finance lead

Owns a recurring report for leadership. Has a small team, little time, and is accountable for every number in it. Already works in team chat all day.

Job to be done

Hand off the report, trust the result

“When I need a recurring report for leadership, I want to delegate it and trust the result without re-checking everything, so I can spend my time on analysis.”

Goals
  • More real tasks delegated, not just questions
  • Less rework after the first run
  • Predictable spend per task
Non-goals
  • Improving the model itself
  • Adding more integrations
  • Redesigning the pricing model
Metrics
  • North star: tasks completed without rework per active user per week
  • Guardrails: cost variance, permission incidents, approval abandonment
Sensitive data

Careful by default

The data in these tasks is often sensitive, and in fintech even more so. People are right to be careful, so the concept starts from strict rules: read-only access with the requester’s own rights, totals instead of raw records where possible, personal data masked by default, nothing posted anywhere without approval, and nothing kept after the task unless the user saves it. Sensitive fields are also protected before they reach the model: masked with reversible codes, or hashed one way. This lowers the risk but does not remove it, because detection is not perfect and pseudonymised data can still count as personal data. I would involve security and compliance before designing the details.

Definitions (assumption to verify)

A semantic layer, where possible

Where it is technically possible, Nomu reads metrics through a semantic layer: shared, named definitions such as “Net revenue” or “New customer”. The agent does not guess what a word means, so the results are more precise. If there is no semantic layer, Nomu asks, and the report says which definition it used.

Risks and trade-offs

Every guardrail adds a step. Asking, planning and approval slow the first runs, so the concept earns speed back: after a task has run several times inside its budget, the plan can be approved once and reused. The riskiest assumptions are about feasibility: can the system estimate tokens and cost honestly before it starts, and is there a semantic layer to read definitions from? I would test both with engineering first.

Principles

Ask when a request is vague. Predict before it runs: sources, definitions, access, cost. Control while it runs: pause, scope, approval. Verify after it runs: sources, receipt, memory you can review.

Clickable prototype

One conversation, seven decisions.

Marta asks Nomu for the weekly board report. Click through the chat. Each step has a design note on the right: the problem, my decision and the option I rejected.

Illustrative data. Not connected to any real tools. The layout imitates a team chat app.

    # weekly-reports4 members · Nomu (AI) added

    Decision log

    What I traded off, and how I would measure it.

    Edge cases

    What the prototype covers, and what stays open.

    Validation plan

    How I would find out if I am wrong.

    None of this has been run yet. This is the order I would do it in.

    Reflection

    Trust is built in small moments.

    What I would do next

    Talk to six or seven people who already delegate recurring work to AI, and watch them do it. Then test the prototype against a plain chat baseline, with engineering in the room to check which numbers the system can honestly show.

    The biggest open question is not visual. It is whether cost ranges and source links are accurate enough to be worth showing.

    What I learned

    Every guardrail costs the user a step, so autonomy has to be earned, not switched on. And a correction is a product decision: it needs a scope, a way to see it and a way to undo it. Building it as a prototype made that obvious, because the correction step needed more states than any other.

    Built with AI-assisted tools, and edited and decided by me.

    “Nomu” is a fictional product and this is a self-initiated concept. It is not affiliated with, or based on private information from, any company.