Every AI assistant getting funded in San Francisco this year solves a version of the same problem: a person has too many emails, too many tabs, too many bookings, and a calendar that fights back.

The product these teams are building makes sense for the market they sit in. An American professional runs their life through an inbox, a browser, and a card on file. Digitise the assistant and you have covered most of the friction in that person's week.

Ask what actually consumed most hours for someone sitting in India last month and the answers look very different. The RO filter guy who said he’d be there on Tuesday and came on Saturday. Two visits to the RTO. A PF withdrawal that needed a physical signature attestation. The school admission form that had to be submitted in person. Chasing a courier that was marked delivered and was not. A hospital appointment booked over three phone calls. The cook who ditched you on WhatsApp at 6am.

None of that is an email problem. Very little of it has a working API. For most of India, the inbox is where OTPs and offers go to die, and browser checkout covers a narrow slice of consumption. An assistant that masters tasks for 10% of Indians in Tier 1 cities has addressed 0.2% of the market and the bottom 0.2% of their frustration.

In this dynamic, my sense is that products like Instinct and Muse will be a party trick that will stay largely unavailable for mass India. It will summarise, remind, book a flight, and then run out of things to do. Retention will be worse in India, because the tasks a digital-only agent can complete are the tasks Indians spend the least time on.

Agents for India will be a mix of AI and on-ground fulfillment

The version of an AI assistant that works here will look like intelligence plus fulfillment. An agent that understands the request, and a network that executes it in the physical world. Agent plus Dunzo, to put it in shorthand.

Two economics make this possible in India and close to impossible in the US.

The first is the cost of a physical task. Sending a human to do something in an Indian city costs $1-2 dollars. In San Francisco, the same errand costs $25. A Western assistant that tried to close the loop on atoms would price itself out of the category on the first order.

The second is that the fulfillment layer here has been built and paid for by someone else. Urban Company has 48,000-plus service professionals and did 13.2 million fulfilled orders last quarter. Quick commerce runs riders through every pin code that matters. Aggregators sit on top of technicians, drivers, agents and small vendors in every city. An agent company does not need to build this from zero. It needs to orchestrate it.

We’re already seeing the seeds of this with multiple India-first businesses that have popped up such as M, Hulp, Aviha, Faff to name a few.

There’s another interesting thread to pull in the India context. An agent that works in India should be able to make phone calls. The supply side has no API, so the call becomes the API. Booking a table, confirming a lab report, pushing a vendor on price: these run through a conversation with someone who will never integrate with anything. India needs an outbound version for managing calls, built for consumers.

Language makes this an India-first engineering problem. A single call can move through two or three languages before it ends. A restaurant answers in English and switches to Hindi mid-sentence. A technician is most comfortable in a third regional language entirely. Layer on a patchy line, a two-second delay, and someone talking over you, and you're a long way from the conditions voice stacks built in the West were tuned for. Whoever gets this working accumulates millions of minutes of Indian vendor conversation, a corpus that stays out of reach for anyone building the same product from San Francisco.

Reaching the physical layer in India starts with a phone call, and the phone call is where the hard engineering lives.

A second interesting aspect is about network effects between agents. Zuckerberg has argued in a chat with Alex Health that agents are a "multiplayer game" and soon it will have switching costs as shared context between you and your relationships will compound. Instinct has already shipped a version of this with "connection," letting your agent work directly with a trusted friend's agent for things like trip planning. I've used it, and it's a step up from the alternative. But the agent still has to check back with its owner on most calls, so it often ends up smoother for the two of us to just talk to each other. It's super interesting, but the loop isn't fully closed yet.

Put the two threads together, and the sequencing gets clearer. Whatever coordination advantage agents build with each other has to survive contact with India's actual bottleneck first: getting a human on the phone who was never going to integrate with anything to begin with.

Complexity is a historic moat in India

Here’s the part that matters for defensibility. OpenAI can ship a better model to Bangalore tonight, at a lower price than any Indian startup can match. But OpenAI will not build a technician graph in Nagpur. It will not learn which of four plumbers in a pin code turns up, which vendor accepts a UPI mandate, or how to escalate at a sub-registrar office. That knowledge compounds, sits outside the model, and gets harder to copy each quarter. The moat lives in the atoms.

The obvious objection: this was tried and it died. Dunzo started in 2014 as a WhatsApp errand service, became a verb in Bengaluru and then shut down. The West ran the same experiment through GoButler,Magic, etc and every one of them eventually wound down.

But the old model ran on humans at every layer. A person read each request, worked out what it involved, priced it, found someone to do it, chased them, and checked the result. Only one of those steps needed on-ground operations. The rest was coordination, and coordination was the expensive part. Rough arithmetic: a dispatcher earning ₹25,000 a month and clearing 40 requests a day loads about ₹25 of coordination cost onto an order that earns a ₹60 fee. Scale made it worse, since each new order needed another slice of a person.

AI takes over the coordination layer and leaves the on-the-ground operations. Comprehension, pricing, routing, vendor calls, follow-ups, and quality checks can all be done by AI, and that cost falls each quarter toward a rupee or two per task. Economics and scalability move together:margin that coordination used to eat stays in the business, and one operator supervises thousands of live tasks, so growth stops demanding proportional hiring.

India-specific nuanced distribution should be key

Distribution weighs on this category as heavily as product quality does, and India makes the problem harder because there is no default channel for an agent to ride None of the obvious channels is open to a builder here: Apple and Google keep the operating system to themselves, WhatsApp admits no third-party agents, a standalone app narrows the funnel before the product gets a chance, and an email-first wedge recruits only the salaried professionals who live in their inboxes.

So the India winner has to invent a distribution mechanic. That has been true of every generational Indian consumer company. Jio priced data at zero. Meesho ran on WhatsApp resellers. PhysicsWallah built an audience on YouTube for years before it sold anything.

The test I would apply to any agent company pitching me: does completing a task put the product in front of someone who does not have it?

The first loop runs through supply. Every physical task ends at a vendor: a plumber, a diagnostic lab, an electrician, a service centre. The agent hands them paid work they did not have to chase, and they come to depend on the volume. The transaction is the pitch, so supply-side acquisition costs are close to nothing.

The second loop runs through the people a task touches. Physical coordination in India involves a household rather than a person. Consider this. You live in Bangalore, and your parents live in Coimbatore. The agent books a cardiology slot, arranges the cab, refills the prescription, chases the report, and posts it to the family thread your sister reads in Dubai. One task, and four people watched the product work: you, your sister, your father, and the neighbour your mother tells the next morning. The lab has now seen where the patient came from.

Both loops need the right first use case. Here are three filters that can be helpful:

• The pain is large enough that people pay for it today, in broker fees, agent commissions, or a relative's leave from work.
• It recurs on its own schedule, so the product earns a place in the household's routine rather than waiting for the user to come up with assignments.
• It involves three or more people per household, so the second loop fires on every completed task.

Potential pitfalls and open questions

Physical fulfilment caps gross margin, and one could reasonably say this is an ops business wearing an AI costume. Take rate on the vendor side and routing efficiency have to carry the P&L, and neither is proven at scale. Token costs falling could also let a global player partner its way into local fulfilment faster than I expect, though platform businesses have historically been bad at that in India.