The second Pan African AI & Innovation Summit ended on Wednesday 23 September at the Kempinski Hotel Gold Coast City in Accra. Over two days it brought together public authorities, technology companies, universities, continental organisations, investors and teams trying to turn artificial intelligence into products. The official theme proposed scaling an ethical African AI ecosystem through youth, policy, partnerships and skills.

Media coverage focused mainly on the summit’s political level. Ghana News Agency highlighted Ghanaian minister Samuel Nartey George’s proposal to regulate risks and harms without preventing lower-risk experimentation. It also reported the warning from Arnold Kavaarpuo, Executive Director of the Data Protection Commission, about a form of technological power that already spans communications, cloud infrastructure, digital identity, payments, AI models and market access and which, in his view, no African regulator can govern alone.

The second day revealed another layer. Wednesday’s opening marked the shift. After the previous day’s discussions about policy, data, partnerships and investment, it was time to move ‘from frameworks to people’ and ask how AI could be turned from an exciting idea into real impact.

Over the following hours, words such as sovereignty, talent, infrastructure and innovation came up against basic phones, schools without connectivity, cocoa waste, banking data that a start-up cannot access, medical devices that have not yet reached a clinic, and models whose outputs need to be integrated into institutions capable of acting.

Co-ordination without uniformity

Thelma Quaye returned at the start of the second day to expand on what she had raised on Tuesday. Smart Africa currently brings together 42 member states and is pursuing the creation of a single continental digital market. That figure refers to the alliance, not to the Africa AI Council. The latter was formally established on 17 November 2025 and has 15 members: seven ICT ministers and eight independent representatives from business, academia and other sectors.

The previous day, Quaye had summarised the problem in a phrase reported by Ghana News Agency: ‘We need coordination, not uniformity’. Incompatible identity, payment, regulatory or data systems may formally preserve each state’s sovereignty while at the same time preventing a solution developed in one country from operating in the others.

In Wednesday’s presentation she brought the argument down to its material conditions. Access to GPUs, training, markets and finance appeared as parts of the same architecture. She also returned to the US$60 billion target of the Africa AI Fund. The figure comes from the Africa Declaration on Artificial Intelligence adopted in Kigali in April 2025. The declaration states that the fund should combine public, private and philanthropic capital to finance AI infrastructure, African companies, professional training and research capacity. Months later, the Carnegie Endowment noted that details of the fund had still not been made public.

In Accra, Quaye also stated that US$18.7 billion in AI financing had already been reached in Africa. We have not found a public breakdown that would allow that figure to be treated as a consolidated data point. We therefore retain it as a statement made during her intervention rather than as a verified estimate of capital mobilised.

The distinction matters. An AI strategy without compute limits what can be built. Talent without buyers or investment can remain little more than training. And a company able to operate only within its own country’s technical or regulatory borders will struggle to access the continental market that is supposed to justify its growth.

After training

George Opare Addo, Ghana’s Minister for Youth Development and Empowerment, devoted much of his intervention to questioning another familiar metric. Access to training programmes, he argued, is only the beginning.

Rather than counting only how many young people receive training or funding, he proposed tracking what happens next. Among the indicators he proposed were formally registered companies, survival at twelve and twenty-four months, revenue growth, jobs created, reinvestment, and the participation of young women and people with disabilities.

The University of Ghana’s Digital Youth Village presented a similar sequence. Olivia Frimpong Kwapong, the project’s director and Dean of the School of Continuing and Distance Education, described a pathway that starts with training and ideation but continues through prototyping, incubation, acceleration, investment and market access before it makes sense to speak of business growth.

Talent infrastructure therefore appeared as something considerably broader than a course.

On the education panel, G. Ayorkor Korsah of Ashesi University distinguished three kinds of knowledge needed by those teaching AI: understanding what it is and is not, knowing how — and when not — to use its tools, and understanding how it works well enough to create with it. She added a warning. Critical thinking, creativity and empathy should not disappear precisely when tools make it possible to delegate a growing share of cognitive work.

Ayo Jones, education strategist and creator of the SHiFT System, brought the discussion down another level, to classrooms that do not always have electricity, water, devices or connectivity. ‘You can teach curiosity without Wi-Fi’, she said. Her argument was not to deny the technology gap, but to avoid allowing missing infrastructure to suspend, in the meantime, the learning of capacities such as curiosity, critical thinking, collaboration and communication.

Tom-Chris Emewulu, founder and president of Stars From All Nations, then added the other half of the problem. It is possible to develop all the AI talent imaginable, he argued, but without a route to market and without paying customers its reach remains limited. For a start-up, market access can ultimately matter more than another prize or another small grant.

Training more people does not by itself create a technology industry. Someone has to buy what they build.

AI that can take a phone call

Some of the most interesting interventions in Accra came from solutions designed around what is missing.

Nii Lante Heward-Mills, who leads Viamo in Ghana and Liberia, described AI tools accessible through basic phones without relying on a conventional internet connection. During the session he spoke of deployments in nine countries and twelve languages. A Citi Newsroom report published two weeks earlier attributed a different reach to Viamo — offline AI in thirteen countries and ten languages — suggesting that the figures probably describe different sets of services or deployments. For the summit’s argument, the crucial point was the interface: a voice call from the phone the person already owns.

The development has a public record. On 15 September, Google explained that it was supporting Viamo in bringing Gemini to basic phones through Ask Viamo Anything, whose pilot in Rwanda had already surpassed two million questions. On 23 September, Viamo also published its own formal announcement of a partnership with Google to expand access to AI through phone calls.

In Ghana, Viamo had launched the free 231 platform with MTN days earlier. The service provides voice access from basic phones and initially operates in Twi, Ewe, Hausa, Dagbani, Ga and English.

Farmerline started from a similar constraint. During the session, its representative described the evolution from an agronomic information service delivered through IVR to Darli AI and said the system could understand ‘about 41 local languages’.

Public sources do not offer a single figure. A GSMA article originally published in May 2025 described Darli AI as supporting 27 languages and said Farmerline aimed to expand that coverage substantially. By autumn 2025, the company itself was already advertising 42 languages, disease diagnosis from photographs, and operation on both smartphones and feature phones. Darli’s current website again shows ‘27+ languages’. Rather than arranging those figures as if they were consecutive stages of the same metric, it is more useful to read them as different definitions of language coverage — available models, languages in production or fully deployed services — that public documentation does not fully clarify.

The Viamo and Farmerline cases suggest a different measure of innovation from the usual one. The larger model is not necessarily the more useful one. In some contexts, what matters more is the connectivity required, the device the user already owns, the language in which they can speak, and the cost of reaching them.

When the prototype meets an institution

The Hack-AI-Thon made the same tension visible in an almost experimental form.

HemoDetect’s final presentation closed, according to the announcement from the stage, a selection drawn from ‘200 plus solutions’. The jury did not limit itself to asking whether the models worked. Again and again, it asked what would happen when they tried to move beyond the prototype.

VitaPulse presented a cardiac monitoring device with local inference. The team explained that it had compared its model against open datasets such as PhysioNet. The jury then asked about the next stage: whether the device had been taken into a clinic, whether data had been collected from real patients, and whether its results had been compared with those reference datasets. The answer was not yet.

The distance between those two stages looks small from the perspective of code and enormous from the perspective of medicine.

Another team proposed detecting and tracing fraudulent financial transactions. When the jury asked whether any financial institution was already using the product, the answer was again no. Its base model had been trained on public mobile-money transaction data and the team was looking for banks or fintech companies with which to continue.

The subsequent comment was especially revealing. The concept was good, but it did not yet constitute a viable production solution. The recommendation was to partner with a bank or operator, work within its data-protection rules and use real information to build, test and validate the model.

The algorithm was not the only missing product. The institution was missing too.

HemoDetect, designed for the automated analysis of laboratory samples, opened another version of the same problem. The system included professional review of results, but the jury asked what happened to the promised productivity gain if a specialist had to review every case. The recommendation was to train on real laboratory data and establish confidence thresholds determining when human intervention is required.

Real data, institutions able to provide it, and rules governing when a person must intervene thus appeared as parts of the system rather than as administrative steps added after development.

Husker AI ultimately won the competition, according to the announcement made during the closing session. KNUST had documented the project on 12 September after it became a finalist in the Ghana AI Innovation Challenge. The system can scan cocoa pod waste, count and classify it, estimate its value, and connect producers with potential buyers for uses such as animal feed, organic fertiliser or biochar.

Cocoa waste ended up beating medical and financial proposals with a more obviously technological appearance. There was a signal there too. Model sophistication mattered less than the relationship between problem, product, users and real conditions of use.

Two hours before launch

The panel on privacy, bias and accountability carried that logic into the development process itself.

Kwame Anda introduced himself as CEO of Creative Waves Consulting and explained privacy through the full product lifecycle. Privacy by design, he said, means not calling the privacy team ‘two hours before launch’. It should be involved from problem discovery onwards and continue through design, development, testing, launch and the system’s later life.

The observation sounds administrative until it is connected with what happened during the Hack-AI-Thon. Problems of access to banking data, clinical validation or human oversight were not outside the models. They determined whether those models could exist beyond a presentation.

Lobna Jeribi, a former Tunisian minister and public-policy specialist, shifted the reasoning to climate resilience. Better information does not automatically produce better outcomes, she argued. Institutions have to interpret it, co-ordinate action, mobilise resources and change their interventions.

Her formulation was especially useful for thinking about AI beyond its technical metrics.

‘What happens after prediction?’

A system may improve a weather forecast or anticipate an environmental risk. The outcome then depends on whether a local authority can turn that information into investment decisions, whether a community can respond, or whether people working in agriculture can change their practices.

Model accuracy becomes only one part of the system that produces the outcome.

What leaves the room

Felix Donkor closed the summit’s second edition by shifting responsibility back towards those who had taken part.

‘Build something, sign the agreement, hire a young person’, he said before inviting them to return in 2027. He added one condition: come back with evidence.

The phrase brought together much of what had happened over the two days.

The first day had discussed how to govern AI from Africa, how to control data crossing borders, how to respond to the power accumulated by large platforms and how to build domestic infrastructure. The second showed what happens when those ambitions reach product scale.

A strategy needs compute. A continental alliance needs systems that can interoperate. Training needs jobs and buyers. A language needs enough data to enter a system. A medical tool needs clinical validation. A fraud-detection model needs a bank willing to test it. Automation needs a decision about when a person should intervene. A prediction needs an institution capable of acting afterwards.

Accra did not resolve those dependencies in two days. It did something more modest and perhaps more useful: it made them visible together.

The balance of this second edition can now be measured by the criterion its own organisers chose at the close. In September 2027, it will not be enough to remember what was said in that room. The question will be what managed to leave it.

References