Lloyds Banking Group, AI hack

A working AI analyst, built by a designer in a day.

Hacks often split into people who build and people who write the pitch deck. I asked to build.

The Northgate analyst console, asking three clarifying questions about a vet practice in Bristol, with sheep grazing on a lawn around the message box.
The brief didn’t ask for an interface. This is the one I built, in funny mode, with sheep.
4AI agents working together on one report
42test questions, rescored after every change
<4 hfrom first commit to the multi-agent report

The challenge

Business clients ask their relationship managers things like “should I open a second vet practice in Bristol?” A good answer takes hours of searching. The brief, set at a fictional bank called Northgate, was an agent that gathers the facts and never invents a number.

My role

I was the designer in a team of five. Four of us built: the two engineers by hand and later with Copilot, and a business colleague and me with Claude Code.

What changed

By the end of the day my version answered from live Companies House data and ONS statistics, and wrote a short market entry report with sources, caveats and next steps. We finished three of the four challenges.

How it works

An orchestrator

Talks to the relationship manager, asks what it needs to know and hands out the work.

A competitor analyst

Searches live Companies House data through a connector we built, for example to count active vet practices in Bristol.

A market analyser

Reads the Office for National Statistics’ local data on business births and deaths.

A strategist

Turns the findings into a short report, so a relationship manager can see where every number came from.

What I learned

AI didn’t do the work for us

We still made the decisions, weighed up technical choices and tested whether things worked. Claude helped me write the requirements and the tests, and I checked the results.

Testing finds what matters

The first test run caught the agent stating rules of thumb as if they were facts. One rule fixed it, and the cases we re-tested went from 1 to 4 or 5 out of 5.

Keep going when things break

Our model quota ran out mid-hack, so we switched models, rescored and carried on.

A designer in the room shows

Funny mode made the whole thing less serious. The team’s verdict: “Ah, I can spot the designer in the room.”

The demo replays answers recorded during testing, so it only answers the test questions listed beside the chat. The cloud project from the hack has since been removed.