Impact Hook
Pipeline controllers held $8M-per-hour infrastructure together with real-time monitoring screens, hydraulic modelling tools, and institutional memory. As sole designer, I led the Business Intelligence application that gave them one place to see the system, catch the anomalies that mattered, and act with confidence.
As the sole designer and Principal Design Lead, I led vision, strategy, and design for an energy operator's first AI-powered Business Intelligence application. The application replaced a fragmented, multi-step daily planning process with a single real-time operational summary and a Human-in-the-Loop anomaly workflow. It was recognized with the 2022 Pipeline & Gas Journal Best Digital Transformation Award.
My Role
Principal Design Lead. Sole designer on the engagement. Owned the end-to-end design outcome across Aspirational Discovery, Product Discovery, and MVP delivery. Partnered directly with a five-person core delivery pod (Data & Analytics Lead, Business Analyst, Solution Owner, Solution Architect, Design Lead — me) and with the client's Operations Planning, Gas Control, and Executive stakeholders.
Deliverables
- Product vision
- Design strategy
- Ecosystem map
- Task model
- Service blueprint
- User journey
- Personas
- Guiding principles
- POC prototype
- MVP high-fidelity design
- UX metrics strategy
- Executive summary deck
Challenge
Pipeline controllers made time-critical decisions across ~70 data sources and roughly 90,000 operational data points, moving between real-time monitoring systems, hydraulic modelling tools, historical data systems, engineering drawings, email, and verbal handoffs. Anomaly reviews took anywhere from 30 minutes to weeks. Missed anomalies surfaced later as outages — with delivery delays, customer impact, and revenue loss on infrastructure generating over $8M per hour at peak.
Business Goals
- Equip Operations Planning with an AI-powered tool to process vast data volumes and support decisions in near real time.
- Streamline anomaly detection to lift throughput without new pipeline capex.
- Keep the human operator in the loop — auditable, safe, trusted.
Solution
An AI-powered Business Intelligence application built on AWS, integrating ~70 data sources into a single operational summary. The application surfaces anomalies by risk; controllers verify, act, and log the outcome through a structured record-keeping flow. The MVP replaced a multi-step manual process with a workflow designed around the controller's decision moments, not the underlying data structure.
Final Outcomes
- ~90,000 data points moved from manual review to automated analysis.
- $8M+ per hour in operational value tied to the application
- Weeks → days, sometimes hours for insight generation, depending on the scale of the problem the operations team is solving.
- 2022 PGJ Best Digital Transformation Award.
Key goal title copy
Give a small team of operators one place to see the whole system, catch the anomalies that matter, and act with confidence — without hiding the reasoning behind the AI.
The problem was cognitive load, context switching, and the absence of a decision-support layer around the data.
Anomaly detection was reactive and expensive. Controllers switched between 20+ tools, relied on senior tribal knowledge, and could only spend about an hour a day on surveillance. The junior half of the team — 75% of headcount — struggled to make decisions without a senior engineer nearby. When an anomaly slipped, it later became an outage.
Approach & Craft title copy
I ran a four-phase design lead-out: Frame → Learn → Model → Ship. It compressed what looks like a linear waterfall diagram in most consulting decks into a single loop the team could run inside Agile sprints.
I anchored the team on one How-Might-We: "How might we enable ops planners to confidently and accurately estimate maximum operational capacity?" Twenty stakeholder interviews, eight follow-up interviews, and a stakeholder-influence map surfaced three user groups. I made the deliberate call to design for the Junior Ops Planner — 75% of the team, and the operator most exposed to knowledge gaps. Designing for the median user, not the expert, is what made the AI valuable.
I mapped the current-state ecosystem — every human, system, channel, and asset touching the Operations Planning team. The map made the invisible visible: disconnects between real-time and modelling systems, verbal handoffs with no audit trail, no execution tracking for management. I then ran 14 primary-user interviews with the core team of operations planners, and used affinity mapping to synthesize the qualitative data into themes for product ideation. The Anomaly Detection user journey and a task-model breakdown of the detection-to-resolution flow gave the team a shared object to design against.
With the Data & Analytics Lead and Solution Architect, I ran co-design sessions on the behavior of the AI in the workflow: what the application would flag, how risk and status would show up in the UI, when to defer to the human, and how the operator would log outcomes through record-keeping. The intelligence layer was shaped here — not later, in delivery. I designed the POC against two guiding principles the operators wrote with us: Keep it simple and Make it thoughtful — with role-based access and streamlined visualizations of large, complex data.
I converted validated concepts into an MVP: application and user flows, digital wireframes, and a high-fidelity design library aligned to global accessibility standards. I set the UX metrics strategy (task success, anomaly detection accuracy, adoption, user satisfaction, system optimization) so the team could measure the AI's impact on human performance, not just model performance.
The design principle: augment the operator, don't replace them.
The application surfaces anomalies by risk. The operator verifies, acts, and closes the loop through a six-step record-keeping flow that captures Anomaly Status → Model Feedback → Resolution Effort → Resolution Benefit → Lessons Learned → Benefits Realized. Three design choices made this work:
Check how to compine with closing sentence: That is what "AI in a high-stakes enterprise workflow" actually means: a design where the model's job is to make the human faster and more confident, not to be right on its own.
Every action the operator takes on an anomaly becomes an auditable record. The screens that let the operator resolve an event are the same screens that let a manager see what was done and why.
Anomalies carry type (for example, volatility shift), status, and a short plain-language summary — enough context for a junior operator to decide without a senior looking over their shoulder.
The AI never auto-resolves. The interface is built so a controller moves from notification to decision in a small number of steps, with clear traceability of what the application contributed and what the operator did.
The client pushed for immediate build. I held the line on 14 primary-user interviews. The extra two weeks saved months of rework and revealed the knowledge-transfer problem the executive team hadn't seen.
The instinct in operations tooling is to expose every signal. We designed for the smallest sufficient view at each decision point, and pushed depth behind clear affordances.
We deliberately did not automate resolution. The near-term efficiency loss was worth the trust the application earned on a safety-critical workflow.
Designing AI for a safety-critical enterprise workflow is 20% model and 80% workflow. The team that owns an AI product is not just design and engineering — it is data, operations, safety, and the executive layer that has to defend the outcome.
The next horizon for AI-powered enterprise applications like this is agentic operator support — models that can propose a resolution path with cited reasoning, and interfaces that let the operator accept, edit, or reject with the same audit rigor. The design work I'm most interested in now is exactly that: how a Human-in-the-Loop workflow scales when the AI's contribution grows.
Being the sole designer sharpened this. There was no design team to distribute the ambiguity across. The choices I made in Frame — who we designed for, what we surfaced, what we deliberately did not automate — became the product. That is the leadership lesson: on 0-to-1 AI work, the design decisions are the product decisions.
The last learning is one I carry into every AI engagement now: insisting on qualitative research inside a Discovery meant to move fast was the highest-leverage move I made. The user data reframed how the executive team saw its own operators, and it turned the work from "an ML pilot" into an application Ops relied on daily.
