I recently gave a TEDx talk in Berlin. You can watch it here.
Because the TED talk itself is barely 15 minutes, interested readers may want to read the two pieces that inform the core points in the talk. The first half of the talk is about how AI grows smart on the “civilisational substrate” and could thus get dumber if human social interactions become less generative and we fail to amplify latent genius in places like Africa cut off from formal economies. It draws on this essay. The second half about how AI can lull Africa into a false sense of “shallow advancement” and thus undermine the continent’s sense of urgency about its industrialisation is elaborated further here.
The two tables below expand on a key element of the second essay (Why AI can hold back Africa’s industrialisation and what to do about it.) Readers arriving from the essay will find in the tables the sector-level evidence behind the warning of a potential honeypot trap that keeps Africa focused on facading with chatbots rather than progressing towards the types of “spatial-abductive” AI critical for industrial transformation. Readers landing on this page first should treat the two tables as a companion exhibit to a broader argument, summarised in the two paragraphs after the table.
Sectoral decomposition of AI Value Map for Africa over the medium-term horizon
| Sector / function | Compositions | McKinsey value signal (US$ bn/yr) | Readiness today | High-value genAI use cases | Binding complement / infrastructure sensitivity | Relevance to Thesis |
| Retail | Customer intent, SKUs, clicks, promotions, product attributes | 6.6-10.4 sector value | HIGH | Shopping assistants; marketing content; hyper-personalized campaigns; customer insights; store-associate copilots | LOW-MODERATE Needs product and customer data, e-commerce/app layers, and CRM integration; physical supply chains still matter but are not the first bottleneck. | Highest sector value in McKinsey Exhibit 1. Most early gains are front-office and commercial: information is rearranged faster before new physical assets are required. |
| Telecommunications | Calls, tickets, customer records, airtime/product data, network-event data | 6.0-9.6 sector value | HIGH | Call-center copilots; chatbot assistants; network ticket resolution; personalized offers; legacy-code support | MODERATE Needs CRM, knowledge bases, manuals, and network-data integration. Existing digital infrastructure carries much of the early value. | A classic early-win sector: the product is already data/connectivity, and the first genAI layer improves service, sales, and ticket flows. |
| Banking | KYC/AML records, credit memos, risk/legal documents, code, customer interactions | 4.7-7.9 sector value | VERY HIGH | Credit-memo drafting; document vetting; risk monitoring; relationship-manager prompts; self-service and servicing assistants; software engineering | MODERATE-HIGH Legacy core systems and regulation complicate scale, but the core inputs are still documents, transactions, and rules rather than physical instrumentation. | A rules-heavy, data-rich industry can capture genAI quickly: the readiness barrier is workflow and compliance integration, not sensors or heavy capex. |
| Customer operations / customer service (cross-sector function) | Conversations, cases, call summaries, scripts, chatbot flows, service decisions | 15.8-26.6 function value; not additive to sector rows | HIGH | Self-service bots; agent copilots; call summarization; state helplines; claims and complaint triage | LOW-MODERATE Requires curated knowledge bases, CRM/workbench integration, quality controls, and a human-in-the-loop for higher-risk interactions. | The clearest “moves information” pool. It is large, legible, and deployable because the work product is text, voice, routing, and decision support. |
| Sector / function | Compositions | McKinsey value signal (US$ bn/yr) | Readiness today | High-value genAI use cases | Binding complement / infrastructure sensitivity | Relevance to Thesis |
| Mining, heavy industry, and energy (core operations) | Ore bodies, plants, conveyors, turbines, substations, field crews, safety/environmental events | 5.3 – 8.5 combined sector value | LOW to LOW-MODERATE | Predictive maintenance; production copilots; yield optimization; fleet management; root-cause analysis; safety and environmental compliance | VERY HIGH Requires sensor historians, FMEA/work-order libraries, SCADA/OT systems, calibrated devices, field tablets/headsets, broadband, OT security, and mixed engineering teams. | The value range is close to telecoms and banking, but readiness is not. The missing complement is the physical data loop, not model access. |
| Energy grid and transmission/distribution operations (within the combined energy range) | Outages, meters, grid balance, substations, maintenance crews, customer interruptions | Included in 5.3 – 8.5 | LOW-MODERATE | Disruption ticket resolution; self-service; servicing assistants; outage and maintenance support; grid and asset optimization | VERY HIGH for operational gains Customer-service cases resemble telecoms, but grid productivity needs smart meters, PMUs, SCADA, asset records, reliable power and connectivity. | Energy can win early at the customer-service edge, but the structural productivity prize is infrastructure-sensitive and capital-intensive. |
| Agriculture and precision farming | Soil, weather, pests, machinery, plots, yields, post-harvest flows | 1.6 – 3.3 sector value | MIXED Advisory higher; precision low | Advisory chatbots/IVR; planting and fertilizer advice; pest alerts; drone/remote-sensing crop monitoring; variable-rate application | HIGH-VERY HIGH Precision gains require drones, soil sensors, remote sensing, variable-rate equipment, connectivity, trained operators, and plot-level economics that can amortize capex. | Advisory AI scales cheaply because it is informational. Yield-changing precision agriculture remains gated by physical instrumentation and operating models. |
| Manufacturing operations / advanced manufacturing (proxy for the broader physical production base) | Production lines, tools, defects, quality data, machines, robotics, throughput | 2.1 – 3.6 advanced manufacturing | LOW-MODERATE | Line maintenance; QA/QC support; digital twins; robotics/control; SOP retrieval; supplier and production planning | VERY HIGH Needs machine sensors, MES/ERP/PLC integration, QA/QC data, maintenance discipline, digital twins, and technicians able to interpret shop-floor signals. | Current LLMs can retrieve SOPs and draft reports; structural productivity requires world-model/digital-twin infrastructure in the engine room. |
The tables relate to the part of the essay that contend that the headline estimate of Africa’s generative AI opportunity, put by McKinsey’s QuantumBlack at between $61 billion and $103 billion in additional annual value, conceals a lopsided benefit structure. Early wins from the current generation of AI (especially the large language models – LLMs) cluster in retail, telecommunications, banking and customer service. These are sectors whose raw material is language, transactions, customer records and rules-checklists, precisely the terrain where Large Language Models excel.
On the other hand, sectors like Mining, energy, agriculture and manufacturing register comparable headline value, yet the continent’s readiness to harness AI for these sectors is far lower. Mostly because converting model capability into productivity there depends on sensors, calibrated instruments, industrial control systems, maintenance regimes and trained field crews, the very “infrastructural complements” the continent lacks.
The essay describes the first group of ready-as-you-are opportunities as “verbal-deductive” and the second as “spatial-abductive” (in between these extremes are “logical-inductive” opportunities clustered around the “agentic workflow” segment of the spectrum where the risks and opportunities for Africa are moderate). The categorisation borrows a taxonomy of reasoning modes developed at greater length in the main text.
The table tests the analytical contrast sector by sector. Panel A covers the early-win sectors, while Panel B covers the gated ones. For each entry, it records the sector’s core informational structure, the McKinsey value estimate in dollars per year, an assessment of readiness today, the high-value generative AI use cases, the binding complement or infrastructure whose absence throttles value capture, and the bearing of each row on the essay’s thesis.
Careful reading should reveal a consistent pattern. In the early-win sectors, value is created because rearranging information is relatively easy, deployment is quick, and the complements are cheap. In the gated, hard-to-crack, sectors, value is created by mastering the physical world, and the missing ingredient is rarely “AI model access” and almost always the data loop connecting algorithms to machines, grids, and sensors. Yet, without mastering this evolving generation of spatial-abductive AI, the true mega-opportunity of AI in such areas as precision agriculture, mining exploration democratisation, manufacturing yield maximisation, and water stress (including drainage) management cannot be realised.
Quick notes on the methodology behind the table:
Value figures are drawn from McKinsey, while readiness and complement assessments are the author’s.