I fixed an AI console by removing before adding
Seven areas of an enterprise AI console, redesigned on one rule: remove before adding.
The console is where enterprises govern their AI: who can reach which models and agents, what data they see, what it costs, and what every agent did.
After the design system pilot succeeded, I was given ownership of the product, and redesigned its pages as a new version beside the old one, so the product kept running.
- Senior Experience Designer
- Role
- Sole designer · 3–5 developers
- Team
- Jul–Aug 2026
- Timeline
- In beta
- Status
01 / The challenge
Hard to find, and different every time.
Things were hard to find, every flow behaved differently, several graphs showed the same data, and badges and colours competed on every page. The brief: miss no function, upgrade the UI, and change nothing in the backend.
The opportunity
How could every page say only what an admin needs to act on?
Every page should
- Show what needs a person first
- Say a rule in one sentence
- Use a badge only where it means something
No page should
- Show the same data twice
- Use colour for decoration
- Make you hunt for a function
02 / Dashboard
From averages to what to act on.
The old dashboard was a mess of graphs and repeated data. The new one opens on one sentence about the state of things, then a strip of what needs attention: ROI status, the review queue, the error rate and the daily cost.
03 / AI Registry
A table you can sort and filter.
The registry of workflows, agents, applications, connectors and models was a grid of cards with gradient thumbnails. It became a dense table with a tab for each type and a search on name or id, because admins sort and filter: they don't browse.
04 / Access rules
A rule you can read out loud.
The rule dialog was a very long list of options, with an allow or deny switch on top of the actions. Selecting an action already allows it, so the switch went. Choosing "Any" ticks and disables the rest, and the detail view writes the rule as a sentence. A planned policies phase was dropped too: adding a policy was the same task as adding a rule.
Fixed once
Long resource menus got clear group headings and a search in the field itself. I tried it on one menu first, then it went everywhere.
05 / PII
One badge per row.
PII has four parts: detection providers, classifications, masking rules, and rules for who may see data unmasked. I asked for each to be explained in plain language before designing anything. Then the badges came down to one per row, row actions moved into the detail dialog, and the tabs and filters moved into the table's own container: controls that change the table belong with the table.
06 / Traces
What an agent did, step by step.
A trace is one run of an agent: the workflow, the agents it called, the tools and models they used, and what each step logged. The first detail view was cramped, so I took things away: the card around it, the hierarchy's title, a floating control. Personal data stays as tokens until someone allowed to see it reveals it.
07 / Teams and roles
A matrix that can't contradict itself.
A role is a matrix of screens and actions. An "All" column ticks and disables the four actions, so a role can't say "all" and "not delete" at once. Teams moved onto the same layout as Users, so every list-and-detail page behaves the same way, and breadcrumbs were fixed once, in the shared component, to start with the page you're on.
| Screen | All | Create | Read | Update | Delete |
|---|---|---|---|---|---|
| Users Management | |||||
| AI Registry | |||||
| Audit Logs | |||||
| PII Policies |
08 / Collections
An agent that calls borrowers.
For a lending deployment, an AI agent calls borrowers before an instalment is due and asks for a promise to pay; reviewers check what was agreed. The call list shows only what needs a reviewer. An instalment is fixed, so a promise must match it, on a firm date: two error rules replaced a vague "partial promise" warning, and confidence scores went everywhere.
Made to be seen
A flagged field gets a bright red border: at 1px "it reads as decoration and reviewers walk past it". Wrong numbers and refusals escalate automatically.
09 / Impact
In beta, and backed.
The redesigned console is in beta. The CXO and the client liked it.
10 / Takeaway
Clarity is what makes it safe.
“In a console that decides who may do what, clarity is what makes it safe to act on.”
Every page got there the same way: remove before adding, say it in a sentence, and design from the user's job, not a generic pattern. When something was wrong everywhere, I fixed it once, in the shared piece.