Applied AI

Two years ago the enterprise AI conversation was "which chatbot should we license?" In 2026 it is "which capabilities run inside our walls, on our data, under our control?" Open-weight models — the Llama, Qwen, DeepSeek, Mistral, and gpt-oss class — now land within a few points of frontier systems on the benchmarks that predict real work, at a fraction of the serving cost. At the same time, the regulatory floor is rising: the EU AI Act’s transparency obligations bite from August 2026, the African Union has a continental AI strategy, and Kenya has moved from a National AI Strategy to a draft national AI policy in public consultation. Sovereignty is no longer a philosophical preference. For regulated data, it is the deployment requirement.

Meanwhile the industry’s open secret is the pilot-to-production gap: by most published counts, barely one in ten enterprise AI agent pilots reaches production at scale. The failures are almost never about model quality. They are about data nobody joined, evaluation nobody built, monitoring nobody owns, and a rollout that skipped the boring questions — access control, fallback behaviour, cost ceilings, and who gets paged when the model is wrong. Organisations fail at AI for operational reasons, not model reasons.

That is why we treat AI as an infrastructure and operations discipline. A GPU node is a server with an aggressive power budget. A model is a workload with failure modes, capacity limits, and a patch cycle. Inference is a service with latency targets and a bill. We have run production estates for two decades under exactly those constraints, and we bring the same discipline — sizing, deployment, observability, change control, and training the team that stays behind — to every AI engagement in this section.