What AI Can’t Reach at the End of the Road

In Guatemala, when a new government comes in, important data often disappears. Files get lost, ministers leave, and what the last administration learned about the economy becomes just rumor. To help, a private university in the capital quietly took on a new role: it keeps the country’s information safe. Fernando Escalante, who manages partnerships at Universidad del Valle de Guatemala, explained this simply. “In some cases,” he told a group of development professionals in Washington in April, “the university effectively becomes a keeper of information for ministries.” The university does this work for free. A few people there decided that losing this knowledge was not acceptable and took it upon themselves to protect it.

Keep that example in mind. An institution is preserving something important because a few people would not let it be lost. This level of commitment is uncommon, and no machine can take its place.

All of this took place at the annual Leadership Convening of ANDE, the Aspen Institute’s global network that supports small businesses in developing countries. This year, the main focus was artificial intelligence, with two sessions for different groups. One session was led by academics, the other by founders already working with AI. Both sessions were practical and avoided exaggeration. Despite their different backgrounds, both groups faced the same difficult truth.

Let’s look at what these tools can already do, because their speed is impressive. Mapping an entrepreneurship ecosystem—the network of accelerators, lenders, trainers, and investors around small businesses—used to take months of fieldwork. Now, an analyst with a good AI model can scan the market, identify key players, find service gaps, and spot trends in just a few hours. The impact on financing is just as significant. Matt Wallace, who started the AI-based lender ONow in Southeast Asia, focuses on the “thin file” problem: strong businesses that lack the paperwork banks require. Tasks that once took weeks to review now take days or even hours, especially when local groups help provide accurate information.

For years, people working in development thought the main problem for entrepreneurs and those who support them was a lack of knowledge. The usual belief was that if you mapped the landscape, measured business performance, and allocated resources wisely, you would succeed. But artificial intelligence has changed things by making advanced analysis relatively cheap and easy to access. Now, we have to admit the revealed truth: the real barrier was never just a lack of knowledge.

The Fair Game

As we have seen throughout the history of computing, the toughest problems are the ones that cannot be automated. Most people in the room now agree that technical prowess is no longer a big advantage. Christopher Neu, whose firm Exchange Design helps organizations use these models, pointed out that the industry often focuses on selling tools instead of creating real value. The models are quickly becoming common and essential, but the real strength lies in the hard work of adapting these tools to an organization’s unique skills. In development, the real advantage comes from being close to the community. Years spent building trust with smallholders or founders cannot be replaced. As these tools become cheaper and more widespread, human relationships will matter even more.

The second insight is more concerning because it questions the heart of development work. Cheap intelligence does not guarantee fair results. Julia Kaufman of J-PAL, known for careful evaluations, shared a clear example. In a Kenyan trial, entrepreneurs used an AI advisory tool. The strongest businesses improved their margins, but the most vulnerable ones actually failed. “It worsened revenue and profits for lower-performing entrepreneurs,” Kaufman said. The concern is not only that the technology did not help them, but that it may have made things worse.

Prateek Shrivastava, from the Alliance for Inclusive AI at BFA Global, highlighted a similar problem in Indian microfinance. There, an easy-to-use emergency loan product on WhatsApp led to a debt spiral because it valued speed over caution. Even the “thin-file” advances that Wallace supports have drawbacks. Yes, making invisible borrowers visible can help them get credit, but it can also lead to more defaults. 

A tool that is only efficient and neutral will always benefit those already ahead and, unintentionally, may harm those on the margins. If the purpose of the development sector is to close gaps, then any technology that accentuates disparities must be fixed rapidly and deliberately, no matter the cost. Fairness is a choice, not a default setting.

You could see the sector trying to find a workaround in real time, though some people were clearly uneasy. While some advocate for a tiered service model, for most, it’s not only the optics that aren’t good: the top 10% get personal coaching, the middle group gets a mix of machine advice and human support, and those at the bottom are left with a WhatsApp bot. If you have ever been stuck with your bank’s “smart assistant” when you wanted to talk to a real person, you know the feeling. This setup might seem fine on paper, but as a social strategy, it is risky. It suggests a future where the poorest are served by algorithms, while the most promising get human judgment. Once this two-tier system is in place, it is hard to reverse.

The third lesson moves from software to structure, looking at how the sector uses its resources. Again, the experts agreed: while picking the right model is important, the real challenge is helping organizations learn and adapt. Pilar Martinez-Gohring, who leads Cosecha—which connects Central American farmers to global markets—explained how building an AI app led to big changes inside the organization. They did not just add a new product. Cosecha rebuilt its data systems and hired four times more engineers. She believes you cannot use these systems well without turning your organization into one that focuses on AI. The tool ends up shaping how you work and who you are. Wallace agreed, warning that we often mistake doing things quickly for actually learning. In the end, only learning leads to lasting progress.

If, in our current historical cultural shift, learning is the actual work, then current funding models are not set up to support it. At this point, both academics and practitioners shared the same frustration about funding.

​The Grant

The Development work is usually set up as projects with fixed goals, timelines, indicators, and a final report. When the project ends, so does the funding. Academics said this is why important knowledge keeps getting lost. A donor funds a study, which leads to a report, then a dashboard, and with the final deliverable the grant ends. What comes next is a story on repeat: the staff leaves, and the knowledge goes with them. This project-based approach not only fails to keep knowledge alive but also makes it hard for organizations to adapt to new technology.

The way development is funded has become part of the problem. Wallace believes an alternative approach to grants that begins by funding experiments and learning cycles, instead of just projects, is an absolute necessity today. For Martinez-Gohring, the best path forward is to support existing systems rather than starting programs from scratch. No need to reinvent the wheel if groups like Cosecha already have the networks, data, and digital tools in place. Shrivastava, in turn, recommended using blended finance to reduce early risks when building from the ground up. 

​This brings up a question the academics could not answer. If the data that describes a public good is itself a public good, Mark Marino of VentureWell asks, “who owns the data? Who maintains it?.” Someone has to be responsible for keeping it accurate and funded over time. Jeff Reid, who leads Georgetown University’s entrepreneurship program and jokingly calls universities “a terrible place for entrepreneurship,” made a strong comparison. Public-health data is treated as a public good. Societies have built institutions and spent public money to keep that data up to date because they decided that tracking disease is worth the cost, even without a specific project. Economic development data does not receive the same treatment. It is created in bursts by whoever has a grant, then forgotten when the grant ends. The difference comes down to what we choose to remember.

Whoever controls the data controls how success is defined. So what is at stake here is not just a database. Deloris Thomas, who leads the Joseph Business School and its campuses worldwide, argued that standard assessments typically measure the wrong things. They focus on the most successful firms and overlook smaller ones, or count revenue without considering whether success helps a poor neighborhood or takes wealth away from it. Most current metrics were designed for a different type of entrepreneur than those her school serves. On a global level, this becomes a question of sovereignty, as Shrivastava pointed out: outside a few wealthy countries, who can even build these systems, given the resources they require? When the models that analyze a poor country’s economy are trained and owned elsewhere, being a “keeper of information” is no longer only a local solution. More crucially, it becomes a question of who gets to define a nation’s identity.

Talent and Trust

There is a human bottleneck at the center of all this, and it is more important than many realize. People who can see beyond their own tasks—what we call talent—are rare, which is why Wallace says organizations should hoard them and pay them well. But the bigger risk is not using too little AI, but using too much. Practitioners warned about the danger of over-adoption: a junior staff member might produce a polished, confident AI-generated report without knowing if it is accurate. The best way to think about AI is as an intern: fast, capable, sometimes impressive, but always needing supervision. This ties back to the first lesson. If expertise is what really matters, then a tool that lets people skip the hard work and just get results is not real help and won’t lead to more productivity in the long run. By making things too easy, creating the ilusion of brilliance, it slowly weakens the judgment that made the organization trustworthy in the first place.

Near the end of the academics’ session, there was an exercise that felt like an admission. Participants got sticky notes and were asked to design the ideal institution to manage ecosystem intelligence from scratch. Every group faced the same challenges. A university could take on the role, but as Reid pointed out, there are problems with incentives and funding. A private fund might try, but it answers to investors instead of the truth. Multilateral organizations already create reports that few people read.

When the groups shared their answers, two-thirds named ANDE, the network hosting the event. Pedro Martinez, the moderator, was honestly surprised. “I swear I was not planning for ANDE to be the answer,” he said. The group could not agree on a coalition, a fund, or a group of universities. Instead, they did something very human. Since they could not agree on who should be trusted with the knowledge, they did the most human thing: they chose the organization that brought them together.

That is the main point, and it is not really about AI. The technology has made analysis almost free to produce, but the main obstacles remain. As is the case with society at large, small business ecosystems must consider how to maintain accurate knowledge after the excitement has passed. Who has to take responsibility for it, besides focusing on the money? Someone has to decide it is worth funding, even when no specific project depends on it. Someone has to make sure this affordable new intelligence reaches the farmers who need it most, not just the businesses that were already likely to succeed. None of this can be automated. All of this is the kind of work the development sector has always said it does, and now there are fewer excuses not to do it.

One of the most important lessons from both sessions, as Martinez-Gohring said, was that learning to support each other through these changes will matter more than any single discussion. Escalante’s university understood this too, even if they did not say it out loud. They keep the ministries’ knowledge safe because losing it was not an option, and because someone saw it as a duty, not just a job. This kind of stewardship is unpaid, sometimes seems unreasonable, and is based on refusing to let important things disappear. It cannot be scaled up or automated. Still, it may be the key to turning cheap new intelligence into real knowledge that actually helps the people development is meant to serve.