The last-mile AI gap
How social enterprises are harnessing AI to scale impact — and where the tech falls short.
- Blog
- All impact sectors
- All Regions
While debate rages over AI utopias and dystopias, social enterprises are quietly implementing these tools into their daily work. Two early and dynamic adopters are Acumen investee Haqdarshak, a civtech firm operating out of India, and Atram, a climate risk platform for financial service providers working with lenders in Kenya and Colombia.
We recently connected with the founders of both companies to cut through the noise and take a hard look at how social enterprises are actually using AI on the ground and how it’s helping or hurting their attempts to scale.
Their experiences reveal a striking paradox. On the one hand, AI is making social enterprises more effective, and that is, in turn, creating climate resilience and economic improvements for actual customers. At the same time, the gaps that have prevented poor and marginalized communities from accessing digital resources in the past (infrastructure, finances, language) are excluding them from the benefits of new AI-enabled services, threatening to leave the most vulnerable further behind.
“We’re seeing a concentration of services going to people who already have services, whereas people who do not have services are getting left out even more.”
Aniket Doegar, Founder and CEO of Haqdarshak
AI lets humans do more
At its best, AI makes humans more effective. When deployed strategically in companies with a social mission, it can create outsized impact.
Aniket has seen this first hand with Haqdarshak. His social enterprise operates a network of 7,000 door-to-door agents, called “haqdarshaks,” who help Indian citizens and small businesses connect to state and federal government welfare schemes. For the haqdarshaks, AI-powered walkthroughs and probability engines make it faster and more effective to connect people with the benefits they have earned. Over the last 18 months Haqdarshak has almost halved their workforce while growing their household reach from 1M to 1.8M, almost a 4x increase in households reached per staff.
That may raise eyebrows; it did for us. AI literally reduced the number of jobs. But it’s important to understand which jobs. Haqdarshak has not decreased its reliance on the human network of agents (its haqdarshaks) that do the actual work of connecting with people and then connecting them with benefits. What AI has done is make it much less intensive to manage that network, to monitor what agents are doing and where they are going. Similarly, it reduced the need for call center staff to follow-up on customers.
Aniket told us, “We want to invest in building AI for our physical agents rather than replacing the physical agents.” What cannot be automated is the person who sits with another person, trying to understand them and their needs. That person cannot be replaced, but they can be made more powerful. In fact, haqdarshak clearance rates have doubled, which directly translates to increased income.
On Atram’s side, founder Davíd del Ser showed us how his company had built a team of AI agents to super-charge the climate readiness of lenders and their staff. Climate affects every aspect of a lending business: product design, screening, management, and collections. Having the right products, processes, and risk information in place — with a level of speed, accuracy, and customization unthinkable before AI — can help banks provide more climate-resilient financial services, years before a flood even strikes.
The most forward-looking banks are already starting down this journey, and it is there that Atram is the most effective. “If you haven’t started your climate journey, call me when you do, because there’s no way we from outside can push you in the right direction. But the deeper you are into that journey, the more you love what we bring to the table”. Atram lets those companies design products that are more flexible and accessible, then bring them to bear in the places and moments they are needed.
Concretely, Atram might encourage a bank to advise a shopkeeper in a flood-prone region to move their inventory onto shelving, or a farmer to harvest their berries before they are destroyed. These nudges can reduce the severity of impact; if paired with the right insurance or contingent loan product, they can facilitate rapid recovery.
Interestingly, despite being an AI-native company (Atram started in 2024) humans remain fixed within its loop. The company layers multiple weather forecasts together, but it is a human meteorologist that ultimately makes the call to send out weather shock warnings. False positives are annoying, false negatives could be devastating.
We connected with Karen Little, who leads Strategic Partnerships and Rapid Response at PayPal. PayPal has supported investment in Haqdarshak through Acumen and investment in Atram through Prosper Global Ventures. Karen told us that the pairing of AI and impact is something that PayPal sees across the small businesses that it serves.
“Small businesses are the backbone of local economies, yet they and their customers are often the first to feel a climate or disaster shock and among the last to be reached by new technology,” she says. “At PayPal, we invest in entrepreneurs and in AI skills for small businesses because access reduces vulnerability and speeds recovery and long-term resilience. Both Haqdarshak and Atram show that AI works best when it extends the reach of trusted people and institutions. That is the kind of AI for small businesses we want to help build.”
The infrastructure deficit
The future is here, but it has never been so unevenly distributed.
At its core, AI allows more trains, carrying more weight, to move faster than ever before on the digital rails of our economy. But that’s cold comfort to someone who is still pushing a cart up a hill.
More than 2 billion people have still never been connected to the internet. In India alone, 600 million people (about 45% of the population) are offline. Those users are overwhelmingly rural and harder to reach. Doegar was honest about his frustrations with the gap between whom Haqdarshak wants to reach and whom it can profitably reach.
“If I have to service a factory worker in Pune today, I can do 50 services a day, and I can make a profit. The same service, if I have to deliver it in Uttar Pradesh [a populous but heavily rural state in northern India], my cost is 2.5x higher, my field person can do 10 services a week, and I don’t make any profit.”
The barriers are numerous, and they are compounding. More rural areas are more likely to speak dialects that don’t have native LLMs. One of David’s biggest needs in West Africa is models that speak Yoruba and Hausa. They are less likely to be digitally literate or to own a smartphone. There will be less data on their borrowing and loss patterns. And unfortunately, from Atram’s perspective, they are less likely to be served by a financial institution that is digitally savvy and climate aware (digital literacy matters for B2B players too).
These barriers are, del Ser told me, “losing us the people we most want to get.” They’re unlikely to be solved by a for-profit enterprise. But Doegar was excited about the idea of an ‘anti-AI fund’ that would make long-term bets on rural infrastructure. Gates Foundation and others are funding more work around real local language models. But we need even more funding on things like smartphone access, community support, and digital identity to ensure that people can benefit from AI advances, instead of being steamrolled by them.
What social enterprises actually need
The tools we are talking about enable outreach and customization at a scale that would literally not have been possible even two years ago. del Ser was excited to tell me: “We just talked to a thousand borrowers yesterday. That’s pretty cool.” But the pace of innovation cuts both ways. Aniket talked about how hard it was to stay focused when a new model drops every month; he could literally be retraining his teams at all times.
When I asked both of them what social enterprises need the most when it comes to AI, del Ser was specific. There is loan-loss data within each financial service provider that would be an invaluable resource to predict future climate shocks and the depth of their effect. The tools to mine that data have never been better, but in many cases the doors remain locked. Creating more public datasets around climate-critical information (such as loss levels) would unlock smarter decision making.
Doegar had a big-picture request. Given their customer base and public benefit, he sees a powerful case for social enterprises to access concessionally priced tokens. A similar concession was made years ago by AWS and Microsoft when it came to cloud computing, but the gap has re-opened now with AI. Lower-cost tokens and model access could create a boom of last-mile service models, or companies serving that market could get priced out.
More from our Knowledge Hub
Beyond the lamp count
Could PULSE, a new Paris-aligned carbon credit methodology, unlock additional revenue for off-grid solar enterprises?
