AI-assisted project feasibility for Africa: satellite data, structured research, and the verification stack

The standard objection to investing in early-stage African development projects is that the cost of pre-investment due diligence is disproportionate to the deal size. A solar-plus-storage project of 50–100 MW in a sub-Saharan jurisdiction generates a feasibility cost stack — site assessment, grid studies, regulatory mapping, off-taker analysis, EPC tendering, environmental and social impact, financing structuring — that has historically run into six figures before any investment decision is taken. That cost is the friction that keeps capital out of projects that, on the underlying merits, would clear an institutional return threshold.

AI-assisted feasibility is changing the shape of that cost stack. Not by replacing the on-ground work — which still has to happen — but by collapsing the time and the cost of the desk-research layer that surrounds it. This piece is what we've learned operating this stack across pre-investment work for African development opportunities. It is sector-agnostic by design — the same pipeline serves renewable energy, agrovoltaics, telecoms infrastructure, and integrated rural development.

What pre-investment desk research actually involves

Strip the buzzwords away and the desk-research layer of an African project feasibility involves answering the same set of questions, in roughly the same order, every time:

1. Site characterisation. Where exactly is the project? What does the land look like? What's the topography, the soil, the water access, the road access, the proximity to existing infrastructure?
2. Resource assessment. For energy projects, what's the solar irradiance, wind regime, hydrology? For agriculture, what are the climate and soil conditions? Multi-year time-series data is the gold standard here.
3. Regulatory mapping. Which national, regional and local regulations apply? What permits are required, in what sequence, with what timelines? Which agencies own which decisions?
4. Stakeholder identification. Who controls the land? Who controls the off-take? Who controls the grid connection? Who controls the regulatory clearance? Who, locally, has standing to object?
5. Financial structuring. What financing instruments are likely? Which DFIs and impact funds are active in the sector and jurisdiction? What are typical terms and timelines?
6. Risk profile. Currency, political, regulatory, off-taker creditworthiness, construction, operations.

Each one of these has historically been an expert-led research task, often farmed out to specialist consultants. The aggregate cost is what makes pre-investment expensive. The aggregate timeline — usually six to twelve months — is what makes deal flow slow.

Where AI changes the cost curve

The leverage points are specific:

• Satellite data analysis. Modern open satellite imagery, combined with computer-vision and time-series analysis, can produce site characterisation outputs in hours that used to take weeks. Land cover, slope, water bodies, infrastructure proximity, agricultural potential, change detection over decades — all programmatic.
• Structured regulatory research. Language models that can read national and sub-national regulatory texts, extract permit pathways into structured data, and cross-reference precedents from comparable projects. The output is a structured regulatory roadmap rather than a hand-rolled memo.
• Stakeholder mapping at scale. Public-data analysis on land registries, corporate registries, government appointments and recent media coverage produces a structured stakeholder map without the equivalent expensive desk-research effort.
• Comparable-project analysis. Embedding-based search across project databases, DFI portfolios and sector reports surfaces comparable transactions for benchmarking.
• Multi-language synthesis. A non-trivial share of relevant regulatory and academic material exists in French, Portuguese, Arabic, Swahili and other regional languages. AI translation and synthesis bring this material into the desk research layer at near-zero marginal cost.

The combined effect is that a feasibility-grade desk-research package — the kind of document that used to be three to six months of consultant time — can be assembled to first draft in under two weeks. The expert layer still reviews, refines, and validates. The drudgery layer has collapsed.

The satellite layer in practice

Satellite analysis is the highest-leverage component because the data is genuinely free, the resolution is genuinely good, and the historical archive is genuinely deep. For any site in any African country, openly accessible satellite imagery archives provide:

• Multi-year cloud-free composites at sub-10-metre resolution.
• Time-series of vegetation indices, water extent, surface temperature.
• Synthetic-aperture radar for cloud-pierced topology and structural data.
• Night-lights data as an indirect proxy for grid presence.

Combined with publicly available datasets — climate reanalysis, hydrological flows, road networks, population grids — you have the inputs for a remote site assessment that's competitive with anything a consultant produces from a desk in London. The pipeline runs the imagery through a sequence of processing steps, generates structured outputs, and feeds them into a project brief. Total compute cost: pence per project. Total wall-clock time: hours.

The honest caveat: remote site assessment is not a substitute for on-ground site visits. It is a powerful filter. It tells you which sites deserve the on-ground budget and which do not. That filtering function alone changes the unit economics of pre-investment, because every site visit you don't have to do is real money you didn't have to spend.

Structured regulatory research

The hardest part of African development pre-investment, in many sectors, is the regulatory mapping. Each jurisdiction has its own permit hierarchy. Authority is sometimes split between national, regional, and local levels in ways that aren't obvious from the texts. Sectoral regulators often have their own approval processes that overlap with the general planning system. Foreign investment regimes interact with sectoral regimes. The whole thing changes annually.

The pattern that works is to build a structured representation of the permit landscape per jurisdiction — what permits exist, who issues them, in what sequence, with what statutory timelines, with what typical real-world timelines, and what the cost structure looks like — and to maintain that representation actively. AI does the heavy lifting on the maintenance: ingesting new regulations as they're published, comparing them to the existing structure, flagging changes for human review.

For any new project, the desk research then becomes querying the structured representation rather than starting from scratch. The output is a regulatory roadmap with specific permit names, agency names, statutory and realistic timelines, fee structures, and known-precedent mitigations.

The verification stack underneath

This entire approach only works if the desk-research outputs are verifiable. AI-generated research with no provenance is worse than no research, because it gives a confident answer to a question the model couldn't actually answer. The verification stack we run alongside:

• Citation discipline. Every claim in the desk-research output is tagged with a source, ideally a primary source. The structured output makes this easy to enforce.
• Cross-source corroboration. Important claims are corroborated against at least two independent sources. The pipeline can run this check programmatically.
• Confidence flagging. Output that the model is uncertain about is flagged for human review. "Uncertain" is itself a structured state, not a vibe.
• Reviewer-in-the-loop. A human expert reviews the desk-research output before it goes into a decision package. They are reviewing rather than originating, which is where the productivity gain comes from.
• On-ground validation. For any opportunity that progresses past the desk filter, on-ground validation closes the loop. The on-ground reality either confirms the desk research or surfaces what the desk research missed; both outcomes are useful.

What this changes for capital allocation

The strategic effect of this stack — when run by a serious operator — is that the cost of saying "no" to a project drops to near zero, and the cost of saying "yes" drops by an order of magnitude. That changes which projects get looked at. Sub-scale projects that previously couldn't bear the desk-research cost become reachable. Speculative early-stage opportunities that would have been declined at the qualifying stage now warrant a closer look.

For the African development context specifically, this matters because the median project size is smaller than the institutional capital allocators typically engage with. Aggregating credible feasibility on a portfolio of smaller projects has historically been the bottleneck. Closing that bottleneck unlocks meaningful new capital flow — not by replacing the institutional layer, but by feeding it qualified opportunities at a rate it can actually deploy against.

What it doesn't change

The stack does not change the on-ground reality. Land disputes don't resolve by satellite. Off-taker creditworthiness doesn't improve because a model summarised the off-taker's accounts. Currency volatility doesn't go away because the regulatory roadmap is structured. The work that actually has to happen on the ground — partner identification, community engagement, regulatory follow-through, construction supervision, operational ramp — is unchanged.

What changes is what reaches the on-ground layer. The filter is finer. The opportunities that reach the institutional desk are better-qualified. The pre-investment pipeline runs at higher tempo with lower cost. The work that wins or loses at the on-ground level still wins or loses there.

The realistic timeline

Standing up the satellite + structured-research + verification stack is its own engineering project. From a cold start, it's six to nine months of focused work to get to first production output, and another six to twelve months of refinement to reach the quality bar where the outputs can drive institutional decisions. The longest pole is the regulatory representation — building it out per jurisdiction is the work, and there are no shortcuts.

For operators willing to do the build, the payoff is durable. The structured representation is an asset that compounds — every project run through it improves the representation, which improves every subsequent project. It is the kind of moat that doesn't show up on a single project's P&L but shows up clearly across a portfolio.

AI-assisted project feasibility is not magic. It is a careful integration of satellite data analysis, structured regulatory representation, multi-language synthesis, and disciplined verification. The combined effect is that the desk-research layer collapses in cost and time, the on-ground layer is fed better-qualified opportunities, and the capital allocators at the top of the stack see better deal flow. None of this changes the fundamental difficulty of African development project execution. It changes the cost of finding out which projects are worth executing.

For the operators willing to build the stack, the next decade of African development opens up. The cost curve has bent. The discipline still has to come from somewhere — the technology lifts the floor; it does not raise the ceiling.