Ayekart is redefining rural commerce by combining artificial intelligence with deep on-ground engagement to solve some of agriculture's most pressing challenges. From enabling better market access and faster credit to building multilingual voice AI and resilient supply chains, the company is creating technology tailored specifically for the realities of Bharat rather than adapting urban digital models for rural use.
Srajan Agarwal, Founder and Editor-in-Chief, News4Bharat, had an exclusive interaction with Milind Borgikar, CTO, Ayekart. He shares insights into the company's AI-driven vision, responsible data governance, multilingual innovation, engineering challenges in rural India, and how technology can create measurable improvements in farmers' incomes. The discussion also explores the future of AI-powered agriculture, digital inclusion, and what it truly takes to build technology that earns the trust of rural India. Edited excerpts:
Q1. As Ayekart builds AI-powered infrastructure for rural commerce, how do you ensure responsible data governance while maintaining user trust?
Trust is the real currency of rural commerce, and we treat data governance as an extension of that trust, not a compliance checkbox. Our approach rests on three principles.
- Context: We take consent in the user's own language, explained in plain terms, often reinforced by our on-ground teams who sit with FPOs and traditional food businesses and walk them through what data is collected and why.
- Minimalism: We collect only what improves the user's own outcomes, whether through transaction history that unlocks credit or supply patterns that improve price discovery.
- Accountability by Design: Role-based access, encryption at rest and in transit, and audit trails aligned with the DPDP Act framework.
But the crucial point is that AI does not replace the human trust layer; it strengthens it. Our field officers remain the face of Ayekart. AI helps them explain a credit decision or a price recommendation transparently, so the farmer hears the "why" from a person they know, backed by a system they can question.
Also Read | Insurance for All by 2047: The Three Changes India Needs Right Now | Exclusive Interview
Q2. Beyond AI, which operational challenge in rural commerce has demanded the most engineering effort, and why?
Honestly, the hardest engineering problem has been reconciling the physical and digital supply chains in real time. Building a system that reflects this ground reality, rather than forcing a rigid e-commerce template onto it, took far more engineering effort than any model we have trained.
Concretely, this meant building flexible order and inventory systems that tolerate partial fulfilment, weight variance at weighbridges, and multi-leg payments; reconciliation engines that can match a cash advance with a UPI payment; and workflows for our field teams so that a transaction recorded in a low-network mandi syncs cleanly hours later without duplication.
This is the hardest because, in rural commerce, the exception is the rule. Every edge case you would deprioritise in urban e-commerce, whether spoilage, renegotiated prices, or a truck that arrives a day late, is a routine part of operating in the hinterland. Our engineering philosophy became simple: digitise the trade as it happens, don't demand the trade change for the software. AI then sits on top of this honest data layer, helping our teams predict, plan, and act, but the foundation is unglamorous, hard-won operational plumbing.
Q3. What single metric will tell you that Ayekart's AI has delivered meaningful value to farmers and rural enterprises over the next five years?
If I must choose one, it is the sustained increase in net income realised per participant on our platform, whether a farmer, an FPO, or a traditional food business, measured year on year for those who transact with us repeatedly. Not GMV, not downloads, not model accuracy. Income in the participant's hand.
I choose this metric because it is unforgiving. It forces every AI capability we build to prove itself on the ground. Better demand forecasting only counts if it reduces distress selling. Credit underwriting models only count if timely working capital helps a food business fulfil larger orders.
Price intelligence only counts if the farmer actually realised a better rate at the mandi. Repeat participation is embedded in the metric deliberately because rural users vote with their feet, and they return only when value is real, not promised.
Internally, we track supporting indicators: reduction in credit turnaround time, decrease in post-harvest losses for supply chains we orchestrate, and share of transactions completed by our field teams with AI-assisted tooling. But those are means. The end is a measurable, durable rise in earnings for the people we serve. If our AI cannot move that number, it is technology for its own sake.
Also Read | "Loyalty Isn't Cashback—It's Recognition"; Ayush Jhawar on the Future of Brand Loyalty | Exclusive Interview
Q4. How did you adapt Ayekart's AI to work effectively on low-cost smartphones and in low-connectivity rural environments, and what challenges still remain?
We started with a design constraint most AI teams never face. Assume a sub-₹10,000 Android device, patchy 2G/3G coverage, shared phones within a family, and a user who may be more comfortable speaking than typing. That constraint shaped everything.
Architecturally, we went offline-first. The app and our field-team tooling record transactions, orders, and KYC inputs locally and sync opportunistically when a network appears, with conflict resolution built in. Heavy AI inference runs on our servers, but we compress the exchange through lightweight payloads, aggressive caching of price and catalogue data, and small on-device models for tasks like input validation and voice capture, so the experience does not stall when the tower does. We test on the actual devices our users own, in the actual geographies they live in.
Our field force is part of the architecture. Where connectivity or digital literacy is a barrier, an Ayekart team member with an assisted-commerce interface completes the journey with the user. AI accelerates that person by pre-filling forms, flagging anomalies, and suggesting the next best action, rather than assuming everyone will self-serve.
What remains hard? Voice AI accuracy in noisy mandi environments, keeping on-device models current without large downloads on metered data, and fraud detection when transactions sync late. These are active engineering fronts, and candidly, they will remain so for years.
Also Read | Why Responsible Interest Rate Regulation Can Strengthen India's Small-Ticket Lending Ecosystem?
Q5. What has been the biggest challenge in building multilingual voice AI for rural India, and what does it take to truly serve Bharat's diverse languages and dialects?
The biggest challenge is not language; it is dialect, domain, and dignity, together.
Standard speech models handle textbook Hindi reasonably well. But a farmer in Vidarbha speaking Varhadi-inflected Marathi about soybean rates, in a windy mandi with tractors idling nearby, using trade vocabulary that never appears in web-scraped training data. That breaks generic models immediately. Rural commerce has its own lexicon: local units of measurement, crop variety names, and credit terms that shift from district to district. Getting a model to transcribe "quintal" is easy; getting it to understand the ten colloquial ways it is spoken across Maharashtra, Bihar, and Odisha is the real work.
What does it take? Three things.
- Ground-truth data collected respectfully: voice samples from real transactions, with consent, annotated by people who actually speak those dialects, often our own field teams.
- A domain layer: we constrain and fine-tune models around agri-commerce vocabulary, which dramatically improves accuracy versus general-purpose speech recognition.
- Humility in the product design: the AI proposes, the human confirms. Voice input is always confirmed back to the user in their language before a transaction is committed, and a field officer can step in at any point.
Truly serving Bharat means accepting that language is identity. When our system responds in a user's own dialect, comprehension improves, but more importantly, the user feels seen. That is when adoption stops being a metric and becomes a relationship.
