Case study · VitalHub UK · SHREWD Resilience

AI summaries NHS teams can check.

LISA is a one-click summary for SHREWD Resilience dashboards. I designed how it appears, how it earns trust, and how every sentence points back to the data behind it.

Phase 1 designed · In development · Not yet launched

01 · Overview

A dashboard full of answers still has to be read.

SHREWD Resilience shows pressure across an NHS system on one dial. Reading it still means scanning dozens of indicators across sites before a system call. NHS clients had asked for a faster way to get the headline, and product wanted dashboards to offer an AI summary.

Product

A feature inside the SHREWD Resilience web dashboard

Problem

Getting the headline from a dense dashboard quickly, without trusting an answer you cannot check

Role

UX / UI Designer. I turned the requirements and wireframes into the full interaction design and spec

Worked with

Product owner, engineering team, NHS clients

Scope

Interaction design, UI, component states, spec, prototype

Status

Phase 1 designed, in development, launch date to be confirmed

1 click

from dashboard to summary. Phase 1 has no typing and no prompt picking

10 min

refresh cadence, matched to the dashboard’s own auto-refresh

0

predictions. The summary describes current metrics only, and says so

02 · The challenge

The concept was a chat panel docked to the dashboard.

The SHREWD Kent and Medway Overview dashboard, with sample data

Generic AI assistants could already summarise a screenshot of the dial. The answer sounded confident, but nothing in it could be traced back to an indicator. The March 2026 concept brought that same pattern into SHREWD. When I read it against the requirements, three things stood out:

  • The panel took a large share of the screen and sat over the indicator tables, so the data you were asking about disappeared.
  • The answer was long prose. Nothing linked a statement to the indicator behind it.
  • No cue for when it was made or what it was based on. In a clinical setting, that decides whether anyone acts on it.

“Can I trust this summary, and where did it come from?”

03 · Design decisions

Easy to ignore. Easy to verify.

I held the design to four principles. Phase 1 was the smallest feature people could trust. Typing and picking prompts was already parked for phase 2.

01 · You ask, it answers

A summary appears only when someone clicks. Nothing arrives uninvited.

02 · Describe, don’t predict

It reports the present state. The footer says “Based on current dashboard metrics only.”

03 · Always checkable

Each statement points at indicators, and the dial and tables highlight them.

04 · Stay out of the way

A 58px orb at rest. One layer when open, closed with one click.

Decisions, and why

Orb bottom-right, where the brief said bottom-left

→

Agreed with the product owner. The left edge is already the navigation sidebar and its footer links.

One click, no prompt box

→

Typing is a phase 2 idea. A single action keeps Phase 1 buildable and easy to learn.

A floating toast, not a docked panel

→

The whole dial stays visible. The cost is that it can cover a table, which the next iteration addresses.

The toast is always white

→

The same readable surface on every dashboard, whatever its theme.

Highlights on the dial and tables

→

They reuse the dial’s own red and amber at lower opacity, so there are no new colours to learn.

04 · Solution

A small orb that opens a summary you can trace.

The orb carries its state through light and motion, not labels. It is 58 × 58px with no text at rest, so it never competes with the dial.

  • Hover. A pill expands to say exactly what a click does: “Generate summary”.
  • Click. A ring spins and a skeleton shows while the summary is built.
  • Summary. Rows are grouped under the dial segment they come from, so the summary has the same structure as the dial.
  • Highlights. The dial segments and tables it talks about are tinted red or amber to match.
  • Close. The × or the orb again.
Built to be scanned in seconds

A context strip names the dashboard and how much was read. Each row has a label and one plain sentence of figures from that area. A multicolour stripe marks the content as AI-generated.

Checked in one more

The same colour links a sentence to its place on the dial. The toast shown at the top of this page is built from the real Phase 1 design.

05 · Next iteration

Five refinements before it ships.

Reviewing the Phase 1 designs against the brief, I found five things to improve. These are proposals. They are not yet in the engineering spec.

Phase 1The toast can cover the tables it describes. Close button only.
→
Proposed · movable and collapsibleA drag handle, arrow-key move, and a collapse to the header and chips so the highlights stay in view.
Phase 1“Generated 14:47” and nothing else.
→
Proposed · stale state“12 min ago”, an amber banner with Refresh, and a dot on the orb.
Phase 1Tint only, with 9px labels in light grey.
→
Proposed · accessibilityA “Red · High” label on every row, 10.5px minimum, in a darker colour that passes contrast.
Phase 1Segments and rows tinted when the summary opens.
→
Proposed · linked highlightsHover or focus a row and its dial segment and table outline together.
Phase 1The same placeholder sentence under two rows.
→
Proposed · contentArea name, a severity label, and one plain sentence of figures from that area.

06 · Feedback and what’s next

“This is what we asked for.”

I took the designs to a workshop and design review with our NHS clients. An AI summary was one of the features they had requested, and they loved it.

Understood straight away

No explanation of the orb or the toast was needed.

Easy to use

One click from the dashboard to a summary.

Backed by clients

It answers their own request, so it went into development with their support.

How success will be measured
  • Share of dashboard sessions that generate a summary
  • Clicks from a summary row to its indicators
  • How often people refresh a stale summary
  • Reports of a summary that disagrees with the dial. The target is none
Before release and phase 2
  • A wider usability round in staging: can people find the source of each statement?
  • Check the toast position on the densest dashboards
  • Design the admin toggles
  • Phase 2: summaries for a single segment and a table, and typed or selected prompts inside the same trust rules

The feature is in development with no release date yet, so there are no results to report. The dashboard figures shown are sample data.

The chat panel made AI the star. Linking every sentence to its indicator made the dashboard the star.

Dropping prompts and notifications from Phase 1 made the feature smaller, easier to build and easier to trust. Read the full case study →

The full case study is password protected. Ask me for the password and I’ll send it over.