Six separate initiatives — AI training, live data connectors, a design-system build and a two-week heuristic audit — unified into one repeatable workflow that surfaced, prioritised and prototyped UX wins for a high-traffic agent task UI.
Ignition is the platform energy-supplier agents live in all day. Collecting a customer’s meter reading is one of its highest-frequency tasks — and the flow looked fine. Agents completed reads; nothing was visibly broken.
But “nothing visibly broken” is exactly where behavioural evidence earns its keep. The organisation was standing up a new behavioural-analytics capability and, in parallel, I was formalising an AI-assisted design workflow. The opportunity was to prove both at once: could we move from opinion-led tweaks to a defensible, prioritised set of UX changes — in two weeks, as a team of one?
Rather than a one-off audit, I built the audit as a system — a structured Claude Code project where context, data and design all live together and every decision is logged.
A top-level repo with shared business context and per-domain folders, plus persistent instruction files so every AI session starts already knowing the role, rules, personas and goals.
Global and project instruction files, a weekly decisions log and a learnings log — the difference between a clever one-off and an operating model.
Two MCP connectors wired straight into the workflow: FullStory for behavioural journeys and funnels, and Figma for design tokens and component awareness — with a research insights backlog alongside.
Flow diagrams uploaded as source; then App 1 (a constrained fix that respects today’s layout) and App 2 (a greenfield optimal flow) briefed against the same evidence.

The temptation with funnels is to invent the steps you expect and measure those. I inverted it: use behavioural journeys to discover the routes agents actually take, then build funnels only for the top paths. Counting by unique users, not sessions, because heavy agents log dozens of sessions each and would skew everything.
The finding that mattered: everything downstream of “Add Read” converts at 95–98%. Data entry wasn’t the problem. Intent was being lost in the gap between opening the meter tab and starting a read — a single, fixable bottleneck converting at just 38–47%.
I prototyped both a constrained and a greenfield response — deliberately — so the team could weigh a quick, low-risk win against the strategic target state, and see the trade-offs in pixels rather than in a debate.


Every finding was evaluated against a combined framework — Nielsen’s 10 usability heuristics cross-referenced with a seven-principle model — then annotated directly in-app. Findings were split by how to act on them: quick wins to ship now, and systematic changes to route through the roadmap and change management.

Beyond the specific fixes, the audit changed how the work gets done. It replaced opinion-led debate with a shared, evidenced basis for prioritisation — and drew a clean line between cheap improvements to ship immediately and structural changes to manage through a roadmap.
It also established a repeatable, AI-assisted method that a single designer could run and a whole team could trust: the operating foundation for maturing a business-design effort into a scalable research & design function.