AI-Powered Operational Knowledge

The Problem

A UK utility operating safety-critical infrastructure manages a large estate of sites and operational documentation: SOPs, site-specific instructions, asset manuals, permits, spread across systems that don't connect. The safety risk is real: when something goes wrong at 2am, finding the right guidance isn't a search problem. It's a pressure and context problem. Staff call colleagues, rely on memory, or make judgment calls under uncertainty. In a utility, that inconsistency has regulatory and operational consequences.

The client's working hypothesis was that better search would fix it. Our job was to test that, and determine whether AI could do more. By week three, the more important question was whether we were solving the right problem at all.

My Role and Scope

I was the only Senior Product Designer on a six-person team: Delivery Lead, Technical Solution Architect, Senior Innovation Consultant, Utilities Innovation Consultant, and Research & Innovation Consultant.

My responsibilities spanned the full discovery arc: building research artefacts, shaping how insights were communicated to stakeholders, driving concept design, and building the prototype that made the product vision tangible. I owned the visual quality of all client-facing outputs: pitch decks, synthesis deliverables, concept presentations.

• 0119 operational roles researched
• 0239 workflows mapped
• 03107 needs synthesized
• 0414 concepts evaluated
• 051 funded direction

The Key Insight, and the Pivot It Forced

In the first three weeks, I helped conduct and synthesise interviews across 19 roles: water, wastewater, bioresources, and central operations. One pattern was impossible to ignore: people don't think in documents, they think in context. An operator dealing with an alarm doesn't search for an SOP. They think: this asset, this condition, what do I do next?

Better search doesn't answer that question. It retrieves documents faster. The question itself remains unanswered.

Design Constraints, Surfaced Through Research

These weren't in the brief. I surfaced them through research and built them into concept criteria before ideation began.

• 01Traceability. Every AI response had to link back to a verified source document. Guidance without citation wasn't acceptable in a safety context.
• 02Governance. The concept had to reinforce existing approval chains, not create workarounds around them.
• 03Field conditions. Many users work in low-connectivity environments. The mobile experience had to function under real operational conditions.
• 04Trust calibration. Frontline staff are trained to be cautious. An assistant that felt overconfident would be ignored or mistrusted. I treated tone and confidence signalling as explicit design constraints with the same weight as layout decisions.

From 14 Concepts to One

We generated 14 concepts across priority opportunity themes. Rather than presenting all of them, I ran a structured prioritisation session scored against user impact, feasibility within the client's existing tech estate, and strategic alignment with their AI programme.

We converged on a single concept: Clarity, a contextual AI assistant that lets field and central operations staff ask open questions, follow guided flows by asset or process, and access site-specific information with full traceability to source documents.

Prototype Build

The hardest constraint of the final week: three to five days to turn a product concept into something a client could experience, not a slideshow.

I built a high-fidelity working prototype in under five days, using Figma for the interaction system and v0 to accelerate implementation. The goal was not to showcase a tool. It was to shorten the distance between product thinking, validation, and executive decision-making. Stakeholders could ask open questions, follow asset-based guided flows, and see source-linked responses in context.

I also produced the video content used to present the prototype to leadership, scripting the scenarios, capturing the flows, and editing into a format that could stand alone without a presenter in the room. The prototype became the centrepiece of the business case presentation and the artefact that moved the client toward MVP commitment.

1. 01Site and asset context grounds guidance in the operator's current situation rather than a generic document query.
2. 02Guided actions support the next decision while preserving escalation and human judgment.
3. 03Source transparency was treated as a core trust requirement, not an optional AI detail.

Outcome

By the end of six weeks, the client had a validated product vision with a working demo. MVP development was funded. The discovery also produced research assets that extended beyond this project, 19 role-based personas and 39 documented as-is processes that became a reference point for the client's broader digital programme. The Clarity framing, contextual, AI-powered, source-transparent, was the direction I defined in week two. It shipped into the MVP phase unchanged.

Reflection

The project reinforced that the hard part of enterprise AI UX isn't generating an answer. It's helping someone understand whether the answer applies, why they should trust it, and what they should do next.

That pushed the design away from a generic chat interface and toward contextual guidance, source transparency, and clear decision support.