At our recent event, the QED team hosted The Anatomy of AI. Not a product demo. Not a panel of predictions. An evening with the engineers who build these systems every day, talking about what actually happens when AI meets a real organisation.
The question we started with: when you are building AI, where do you land on cost, accuracy, and privacy? Not as abstract values. As real constraints that shape every decision in the build.
You cannot maximise all three. That is not a limitation of the technology. That is the nature of the problem. The job at the start of any project is to understand which one matters most, and design from there.
The evening was built around three case studies. Three industries. Three very different places to land on that triangle.
Education: safeguarding 110 UK schools
Before this project, safeguarding worked on keyword alerts. Every match triggered a manual review. At that speed, a real problem could escalate before anyone reached the school.
The AI layer we built analyses full context, not keywords. Text, images, behaviour patterns. It removes the noise before anything reaches a human reviewer, so auditors spend their time on the cases that actually matter.
- 0 known critical misses across 1.8 million alerts processed
- 42% of alerts auto-cleared, up from 24% at launch
- 500,000+ benign alerts cleared before reaching a human queue
- Eight-minute end-to-end SLA on around 40,000 captures a day
Privacy was not a feature. It was a design constraint. Identifiers are anonymised before anything is sent to a model.
Read the full case study here.
Energy: cutting consumption without new sensors
Large commercial buildings are usually optimised reactively: systems respond to current conditions instead of anticipating demand. Many also lack the occupancy sensors that would let them do better, with no capex to add them.
The system we built infers occupancy from signals the buildings already produce - air quality, HVAC, usage - with no new hardware, and forecasts heating, cooling and ventilation demand instead of reacting to it.
- 28% reduction in energy consumption
- 35% improvement in HVAC efficiency
- 94% air quality consistency, held within target throughout operating hours
Read the full case study here.
Defence: institutional knowledge that stays secure
Cost estimation in defence depends on expertise that lives in people. When those people leave, the knowledge goes with them. Proposals took around four weeks to get right.
The constraint here was privacy, not accuracy. No sensitive data could leave the client's environment. We built the system entirely on self-hosted language models. The institutional knowledge now lives in the system, not only in people.
- Proposal turnaround: four weeks to five days
- 35% more accurate cost estimates
- 60% reduction in the gap between estimate and final cost
Read the full case study here.
What the evening was really about
Three industries. Three completely different constraints. The same underlying question every time: what does this organisation actually need, and what are we building within?
The conversations after the presentations went further than we expected. People were not asking what AI can do. They were asking whether they can trust it, afford it, and control it. Those are the right questions.
That is the conversation we want to keep having.
If you want to understand where AI could move the needle for your organisation, start with a diagnostic.
Published