Case study
Tender
A workspace for an AI procurement agent. You describe what the business needs, the agent plans the job, finds and contacts suppliers, and checks in with you at the moments that matter.
Client
Concept product
Role
Product design, prototype
Platform
Web app, Figma, code
Year
2026
Status
Self-initiated
Contents
Process
01
Understand
Trusting an agent with real money
02
Design
Brief, plan, suppliers, emails
03
Build
A working prototype in code
04
Launch
Simplified and shared
05
Outcome
What it shows and what is next
Tender
AI agent / 2026

An agent is only useful if you can see what it is doing and stop it when it matters.
01 Understand
The problem
The brief
A self-initiated product: a workspace for an AI agent that handles business purchasing, from a request to signed-off quotes. Built to show how I design for AI that acts, not just answers.
What I looked at
Where people hesitate to hand work to an agent. Spending money and contacting suppliers on your behalf are the moments trust breaks, so those needed the most design.
The decision
Let people choose how much the agent does alone, and make every step it takes visible, explainable and reversible.
Insight
People don’t need the agent to be clever. They need to know what it will do next.

Levels of autonomy

Design principles
02 Design
The flow
Brief
Requests start as a plain-language brief. The agent turns it into clear requirements you can check before anything happens.
Plan
Before acting, the agent shows its plan step by step, so you approve the approach, not just the outcome.
Suppliers and emails
It shortlists suppliers, drafts the outreach and tracks replies, with every message visible before it goes out.
Principle
Show the plan before the work. Approve the approach, not just the result.

Brief

Plan
03 Build
A prototype in code
Coded, not mocked
I built Tender as an interactive prototype in code rather than static frames, so the agent’s states, waits and checkpoints could be felt, not just shown.
Simplified twice
The first dashboard version was too busy. I stripped it back to an editorial layout with fewer panels and more room, then arranged that calm style back into a working dashboard.
System
Inbox, brief, plan, suppliers and emails share one set of components, so every stage of a request feels like the same product.
Why it matters
Agent products live in their in-between states. A coded prototype is the only honest way to design them.

Suppliers

Emails
04 Launch
Showing it
Checkpoints
The moments the agent pauses for approval are designed as clear decisions, with the evidence it used laid out beside them.
Shared on Contra
Presented as a set of 1600 x 1200 frames for Contra and as a full case study here.

Supplier scoring

Approval checkpoint
05 Outcome
Results
5
Stages from request to quotes
Code
Interactive prototype
2x
Simplified from the first version
Every
Agent action visible first
What I learnt
With AI products, the interface is mostly about trust. The busier it looked, the less people believed it.
What is next
Testing the autonomy levels with real buyers, and a mobile view for approving checkpoints on the go.
Next project
EchoMind →