Sasol is an instructive case for African industrial AI. JSE-listed, deeply integrated petrochem and energy, exposed to South African operating conditions, with the budget for serious deployment but the procurement discipline of a regulated industry. The 10 GenAI pilots they disclosed in 2025 are running at Secunda and Sasolburg in 2026.
I cannot describe specific pilots in detail — Sasol's operations technology team is rightly careful about what is public — but I can describe the architecture, the discipline, and the patterns. African mid-tier energy operations (renewables IPPs, distribution-grade infrastructure, standalone gas-to-power) can copy most of this directly.
The portfolio shape.
Ten pilots spread across three categories: (1) operations efficiency (cogeneration tuning, heat-integration optimisation, water management), (2) maintenance + reliability (predictive models on critical turbines, compressors, exchangers), and (3) safety + compliance (autonomous inspection routing, incident triage, regulatory reporting). The deliberate diversification is the discipline: Sasol is not betting on one application class, it is testing the spread to find which applications recover capex fastest at their operational scale.
For mid-tier African energy operations, the portfolio shape translates with one modification: the safety + compliance category should be heavier-weighted, because mid-tier operations cannot absorb a major regulatory incident the way Sasol can. Diversification first, but skew toward incident-prevention investments at smaller scale.
The architecture.
Sasol's operations technology team runs a layered architecture: AVEVA PI as the historian backbone (decades of investment), Microsoft Azure for the data platform layer (Synapse, Fabric, OpenAI Service), and a custom application layer Sasol built internally for the GenAI pilots. The architecture is opinionated: PI is the source of truth, Azure is the compute, the application layer is sovereign.
What I found notable: Sasol does not run GPT-4 directly. They run Azure OpenAI Service with data-residency guarantees, finetuned domain models for the operations technology context, and a strict policy of never sending operational data outside the Azure private endpoint. For African operations with energy-sector regulatory exposure, this is the right pattern. The data residency conversation matters in 2026 in a way it did not in 2024.
The discipline.
Three patterns I observed worth copying. First, the pilots have hard kill-criteria. Each pilot has predefined success metrics (efficiency gain, downtime reduction, incident avoidance) and predefined kill thresholds. If the pilot does not clear the threshold by the agreed quarter, it is killed. The discipline prevents the "eternal pilot" problem that plagues a lot of industrial AI investment.
Second, the pilots are sized to a 12-month horizon — long enough to get real operating data, short enough to maintain executive attention. Sasol is explicit that anything claiming 24+ months to value either has the wrong scope or the wrong sponsor.
Third, there is a quarterly cross-pilot review. Lessons from one pilot — particularly integration patterns, vendor performance, and unexpected failure modes — propagate immediately to the others. This compounds. The 10-pilot portfolio is producing more shared learning than 10 independent projects would.
What African mid-tier energy can copy.
Most of it. The architecture pattern (historian + cloud data platform + sovereign application layer) scales down. The diversified-pilot portfolio approach works at any scale — three pilots at a R500M operation works the same way as ten at a multi-billion operation. The kill-criteria discipline costs nothing.
What does not translate: Sasol's internal application layer build. A R500M IPP cannot justify a custom application platform team. The Foundry Cortex recommendation for mid-tier African energy is to use SymphonyAI IRIS Flows for the application layer (300+ pre-built industrial agents covering most of the categories above) rather than custom build. Save internal engineering for the integration layer between IRIS and your specific operations technology estate.
What to skip: the GenAI hype-cycle naming. Sasol calls them "GenAI pilots" partly because their board wants the AI exposure narrative. The technology underneath is mostly classical industrial AI — predictive models, optimisation, anomaly detection — with LLMs playing supporting roles for natural-language interfaces and report generation. If you find yourself architecting around LLM-first patterns for industrial operations, you are probably optimising for the wrong story.



