Why Agentic AI Fails to Act Intelligently
AI has taken over many tasks we once thought were uniquely human. It can detect fraud in milliseconds, recommend what to watch next, and automate entire workflows in fintech and e-commerce. But there’s one area where AI still fumbles hard: real-world shopkeeping.
Running a store — whether online or offline — isn’t just about transactions. It involves multitasking, understanding nuance and making judgment calls in real-time. That’s something even the smartest AI agents still can’t do well.
Let’s checkout where things go wrong, what companies are losing because of itand what the path forward might look like.
The Real-World Challenges AI Can’t Yet Solve
1. Lack of Contextual Understanding
AI thrives in clean, rule-based environments. But shopkeeping is messy. Customers can be emotional, store conditions change constantly, and no two days are the same. Most AI agents aren’t trained to handle this dynamic reality.
Razorpay for example, can resolve routine support tickets in chatbot around payments and refunds. But when a transaction fails due to a rare backend issue or bank-side delay, the AI often loops the user through generic responses. It doesn’t know how to escalate smartly or sense urgency — leaving the customer frustrated.
2. Rigid Workflows Break Under Pressure
AI workflows tend to be rule-based and linear. They expect users to follow predictable paths. But real customers don’t behave that way. Shopkeepers need to make flexible decisions — like replacing an out-of-stock product with a similar one or adjusting prices for a loyal buyer.
Dunzo for Business (D4B) faced this when it scaled its vendor assignments using AI. The AI couldn’t adapt when shops shut early or when merchants made on-the-fly substitutions. Orders got cancelled or delayed, hurting merchant relationships.
3. Emotional Intelligence Still Lacking
A seasoned shopkeeper or support agent can read body language, listen between the lines and calm a frustrated customer. AI, for all its data-crunching power, still struggles with emotional cues.
In fintech, this gap is crucial. When money is involved, customers expect empathy. If a loan is delayed or a transaction fails, users want reassurance — not a scripted message. That’s where AI still fails.
4. Data Biases and Gaps in Language/Locale
Many AI models are trained on urban, English-centric data. In India, a majority of digital users are from Tier II and III towns, speaking in Hindi or regional languages, often with mixed transliteration.
Get Snehagrawal’s stories in your inbox
Join Medium for free to get updates from this writer.
Paytm’s AI KYC agents have often flagged perfectly valid documents simply because the photo is taken in poor lighting, or the ID card has creases. In such cases, the AI’s decision feels robotic and unfair — because it lacks localized empathy and flexibility.
The Cost of Poorly Designed AI Systems
These struggles aren’t just user experience issues. They have real business consequences.
1. Customer Churn and Drop-Offs
If users can’t resolve an issue quickly, they’ll leave. In fintech especially, delays in verification or refunds drive users to competitors. Poor AI workflows amplify this churn.
2. Brand Trust Gets Impacted
When users feel ignored or misunderstood, it damages your brand. AI agents that give irrelevant answers — or worse, refuse to connect you with a human — can turn a small issue into a viral complaint.
3. Lack of Conversions and Sales
Rigid AI that can’t handle substitutions or dynamic inventory leads to abandoned carts and failed deliveries. A smart human would have solved it on the spot. AI? Not so much.
4. Increased Support Load from Escalations
When AI fails to resolve basic exceptions, it increases the burden on human agents — nullifying its intended efficiency. Every unresolved case becomes a support ticket, which costs time and money.
What Needs to Change: From Smart to Street-Smart AI
These problems aren’t permanent. AI can do better — but only if we design it to complement, not replace, the human experience.
1. Using Hybrid Models
Don’t aim for 100% automation. Use AI to handle repetitive, structured tasks — but hand over complex or emotional interactions to trained humans. This hybrid setup works best in high-touch industries like payments and commerce.
2. Investing in Context-Aware AI
Train AI to understand soft signals — like urgency in tone, the context of user frustration, or regional transaction patterns. This means moving beyond keywords and towards intent + sentiment understanding.
3. Prioritizing Multilingual and Localized Intelligence
Build AI systems that can converse fluently in Indian languages and understand rural behavior patterns. Use localized datasets and regional voice inputs to reduce errors and biases.
4. Building Smart Escalation Paths
Don’t trap users in chatbot loops. Ensure that there are clear, human fallback mechanisms — especially for payments, identity verification, and order issues.
The Future: Empowering Shopkeepers, Not Replacing Them
AI won’t replace the neighborhood shopkeeper, customer support agent, or delivery boy anytime soon. But it can amplify what they do — by helping them respond faster, manage stock better, or track payments smarter.
Imagine an AI system that reminds a kirana store owner when to reorder popular SKUs or automatically follows up on digital credit payments with polite, customizable nudges in the user’s language.
That’s the kind of AI we need: context-aware, culturally fluent and collaborative.
Until then, the chaiwala with no CRM tool but perfect memory still outsmarts the AI assistant.. 😁😁
