The Hardest Part of African AI Is Still the Paper

On October 9, 2026, Disrupt Africa reported on Entelon, a Kigali startup founded in August 2025 by Alain Barsime Shema and Don Durkheim. The description of what the company actually does is worth sitting with: it takes over work organisations run on paper forms, hand grading, unanalysed phone calls, and data entry, across health, research, education, customer experience, and travel.

Most coverage of African AI startups treats the word AI as the interesting part. It rarely is. The interesting part of Entelon's description is the word paper, and what it implies about where the engineering effort actually sits.

Two different starting points

Compare Entelon to Bujeti, the Y Combinator backed Nigerian fintech that just made its AI teammate platform, Brain, publicly available. Bujeti's four agents find invoices, process payments, run approvals, and reconcile accounts for finance teams already running on corporate cards, vendor payment automation, and a mobile app. More than two thousand finance professionals in Nigeria and Kenya use the platform. That number is possible because the underlying data was already structured before any agent touched it. A card transaction is a clean record the moment it happens. An invoice uploaded to a payment platform already has a predictable shape.

Entelon starts somewhere else entirely. A hand graded exam script, a paper patient intake form, a phone call that was never transcribed, none of these are data in any usable sense until someone builds the step that turns them into data. That step is not a model call. It is a decision about what a clean record of that event even looks like, made by someone who has gone and looked at the form.

The work nobody names

This is the part of building AI in this region that gets skipped in most advice aimed at founders here, because most of that advice is written by people whose reference point is a market where the digitisation already happened years ago. The actual engineering risk in a product like Entelon's is not whether the model can summarise a transcript. It is whether the system can take a handwritten form filled in a local dialect, with a field left blank and another contradicting the register it is meant to match, and produce something a clerk or a clinician can trust enough to act on. That is a harder problem than running an agent against an API that already returns clean JSON.

CIO Africa reported this month on how AI is reshaping Rwanda's global business services sector, which is itself evidence that there is now a market for exactly this layer, the unglamorous work of making an institution's records legible before anything smarter can run on top of them. That demand did not exist five years ago because the organisations running on paper were not yet being asked to digitise fast enough to need outside help doing it.

Isn't this just OCR with a better name

The fair objection here is that this is old automation wearing a new label. Document extraction, handwriting recognition, and data entry tools have existed for decades, and calling them AI products does not make the underlying engineering novel. If Entelon is simply running an OCR pipeline and calling it an AI product, the market should treat that claim with the same scepticism it treats any vendor dressing up a simpler tool.

The answer is that extraction is not the hard part, and never was. The hard part is judgment under incompleteness: what does the system do when the form is wrong, when two fields disagree, when the handwriting is ambiguous, when the record simply does not map cleanly to any schema. A pure extraction tool returns garbage or a low confidence score and stops. A system built for this condition has to decide, consistently and defensibly, what counts as resolvable and what gets flagged for a human. That decision logic, applied across a health clinic's intake forms or a school's exam scripts, is closer to the judgment a front desk clerk exercises every day than to anything a transcription API does on its own.

This is also where the advantage that rarely gets named actually sits. A team building on top of an already digitised American or European dataset inherits someone else's assumptions about what a clean record looks like, baked in long before they arrived. A team building from the paper up has no such inheritance. They are forced to define the schema themselves, against the actual mess of the actual institution, which means the resulting system is built to survive contact with conditions that an imported tool was never designed to meet.

The next time a vendor pitches an AI product for an East African institution still running on paper, the question that matters is not what model they use. It is what their system does the first time a form arrives half filled, contradicting itself, or written in a hand nobody can quite read. If they have an answer for that, they have built something. If they do not, they have built a demo that will survive exactly as long as the data stays clean, which in this region is not very long at all.

Sources

Rwanda’s Entelon is building AI for organisations still running on paper (Disrupt Africa, 2026-10-09)

YC-backed Bujeti launches AI “teammates” to take manual work off African finance teams (Disrupt Africa, 2026-10-08)

How SA’s deedy is letting homeowners sell property without an estate agent (Disrupt Africa, 2026-10-07)

How Pakistan’s KangarooCare is using AI to improve neonatal care (Disrupt Africa, 2026-10-09)