How Do I Work With My Dream Cos?

A terrible, terrible cover letter

If you had asked me a month ago what OpenAI or Anthropic should do in India, I would have had a list ready. (I mean, who didn’t?)

(As a professional haver of ideas, I would like to think mine were cooler but I know they weren’t.)

More regional language support. More voice-first tools. My mom needs a reason to move away from Google search. Better partnerships with universities. Founder programs. Government engagement. India-specific pricing. Better distribution. Maybe an accelerator. Maybe a fellowship. Definitely something for creators.

Unfortunately, this whole list is almost entirely obsolete.

Over the past week, I read every OpenAI and Anthropic announcement about India I could find. Product launches. Global Affairs updates. Hiring pages. Partnership announcements. Education initiatives. Benchmarks. Interviews. Press coverage. Turns out, the suits at the helm had some idea about what they’re doing. Nearly every obvious idea had already been built.

Access, a question we all were asking when we started thinking of AI for India, is at least partially answered.

So, what after access?

Adoption is not diffusion

Adoption is always framed as a problem of distribution. I don’t know about this product, I learn about it, I use it. A beautiful, linear process that suits people like you and me wonderfully, and has very little to do with how technology gets diffused into a society.

Here’s the difference that I’ve settled on in recent months. Adoption is the decision I make to use ChatGPT or Claude. Diffusion is the effect of my using it on ten more people. Adoption and diffusion are different, and they get measured in different ways. Adoption creates users, MAUs, licence purchases and so on, the statistics that go in the launch posts. Diffusion creates behaviour and behaviour shows up on dashboards only long after the event.

It can be viewed as a three-phase problem. First phase: access. Can I even reach the tool? Second phase: accessibility. Is the tool accessible to me where I am in my language, on my payment rails, at the prices that I’m willing to pay, rather than making me come to it? And the third phase has very little to do with the existence of AI. Even my grandpa believes in AI. What remains is persuading people that they can reshape their working lives using the technology. This does not come from a press release about some feature. It comes from seeing someone trusted doing it.

Somebody in a learning cohort at GrowthX builds the first internal AI application, and suddenly three more GXers build theirs. Suddenly, GrowthX realises that this is a workflow of workflows they could consolidate and teach. A recruiter figures out how to use AI to screen her candidates differently and replicates herself in her whole team in a month. A teacher redesigns her lesson plans, and it spreads among her peers in the staff room. No playbooks are distributed, no files on skills are shared.

Someone sees someone they trust and two things happen at once: she believes that if he could do this, then so could she, and suddenly the technology doesn’t feel threatening anymore, but cooperative.

Sociology calls this a Diffusion Process, and the fact that it maps to AI as the perfect pun is just great!

I spend most of my working life talking to people. It is quite literally my job. Every discussion that I have started with AI has ended in one of four places: career, ambitions, education, and loneliness. I expected these discussions to change with changing models, and they haven’t, which is telling me where the diffusion battle will actually be fought. (The discussion about loneliness deserves a separate essay and will get one).

A market must expand in every way

AI India is not a market expansion story anymore. It’s an organizational redesign story, and organizational redesign consists mostly of gloriously non-sexy problems. Coordination of hospital services. Preparation of your taxes by chartered accountant. Documenting quality assurance in factories. Lesson planning in schools. Movement of documents in district administration. Filing FIR reports in police station. Supply chain management by the export manager. Logistics of delivering orders by Nykaa.

None of those industries ask for AI. Most of them are currently suffering from the poor implementation of AI precisely because the template of such a huge country is to optimize for reach without measuring the quality. Users, institutions, partnerships, licenses are all real metrics, of course, but they are lagging. The leading indicator is how many people reshaped their working life based on the demonstration they’ve seen.

So if the obvious list is done, and the new era is diffusion, what would I build?

Build for operators, not just founders

The conversation around AI circles founders as central characters and for good reason. Founders are, well, founders. They lead the charge, create companies, change processes quickly, and own the onus of changing lives on a daily basis. My time in VC has created a distorted view of change because working with founders makes you think every founder is amenable to experimentation.

Investing in companies that build for MSMEs reminded me of how truly small the world of startups and venture capital is. The guy sitting in a rented 2BHK in HSR with a Delaware Incorporated entity is going to say yes to anything AI is going to say yes only. If you want to see how change emerges in a small business in in India, you really have to go talk to a CA who has done the books for a whole bunch of startups and inherently distrusts ‘all these new things’ because she has seen founders monumentally muck up their books trying to ‘optimise’. It might be the accounts person who has rolled their eyes at an agent workflow that could have been an email. It might be the admin guy who somehow knows how everything works, right from the AC machine to the domain of the Claude-generated website. You’ll meet variations of this person constantly. On paper, they have limited power and authority, but their practical authority makes them arbiters of outsized trust, and hence, owners of outsized context.

India, I’d wager, is a land of operators. We are very good at making terrible systems, and then we are excellent at making people jump hoops to learn how to navigate those systems. Once the hoops are jumped, we consolidate a cadre of people whose entire job is to translate systems to the rest of us chumps. That makes the most influential distiburion layer for AI because their behaviours have multipliers attached to them. If I can convince one founder to use AI well, I change one company. If I help a CA who works with 200 cos find a way AI can become genuinel useful to her practice, I potentially touch 200 cos without having to convince 200 founders separately. It will not look like the glittering ‘we are now AI’ announcements on LinkedIn, and that’s fine. Diffusion is not a moment!

There is a psychological cost to the adoption of AI we tend to gloss over. We have spent the last three years telling people two things simultaneously. One, THIS IS THE MOST IMPOTANT TECHNOLOGY OF YOUR LIFE AND IF YOU IGNORE IT YOU ARE IN TROUBLE. Two, DON’T WORRY, THE TECHNOLOGY THAT CAN TAKE YOUR JOB WON’T…OR WILL IT?

Of course there is fear attached to AI! Of course there is inertia! A ChatGPT or Claude workshop might teach me how to prompt (badly, might I add), but it cannot give me the permission I need emotionally to change how I do my job, something I have been good at for the last 15 years. That permission only emerges from me seeing someone I respect incorporate new technology into their work without becoming meaningfully worse for it. A surprisingly large amount of tech adoption is simply borrowed courage.

This is why the first thing I would build is an Operator Network (that is not a Whatsapp group that everyone mutes in 3 weeks or less). I would curate a deliberate network of people who have disproportionate influence on how work happens inside orgs across finance, legal, hiring, sales, admin, customer support, procurement. Basically, people who keep the wheels turning but rarely get treated as the protagonists of the story of technological innovation. While easier to optimize for industry first, I’d instead focus on the function. A finance operator at a logistics co has a lot more in common with a finance operator at a luggage company than either do with their respective marketing teams.

The measured unit of participation will be workflows. The price of admission will be something that is expensive, repetitive, frustrating, slow, or unnecessarily dependent on one person’s institutional memory. The engagement will involve them rebuilding it with Ai tools, establish and document the changes and failures, and teaching it to someone else. To borrow from teaching hospitals, see one, do one, teach one. The metric of obsession is not an NPS score or number of participants. It is the number of ways and industries a workflow can travel into. Simply, if a workflow built by one finance team can get picked up by another without needing someone to being paid to create adoption for it, success!

Why should an OpenAI or Anthropic do this? Because this network will be an extraordinary listening tool. They already know what millions of Indians prompt every day. I would personally do a lot of marginally illegal things for one day with their data dashboards (hehe). What is ,uch, much harder to figure is what someone does after they close the tab or the desktop app. Did the answer actually answer anything? Did the finance team take any inputs? Did the workflow fall apart when persona A tried to teach it to persona B? Did it work for a 400-person logistics co but completely fall apart for a 12-person distributor dhandho in Surat?

This infprmation is going to become increasingly more valuable as models get better. Once ‘can the model do this?’ becomes redundant as a question, the more interesting one to ask will be ‘why didn’t the org adopt this even though the model could do this?’ An Operator Network answers that in realtime because you get to observe and support people taking frontier models to non-frontier conditions.

More importantly, you start building flywheels. Not ‘use cases’ (ugh) but demonstrated and sustainable changes in workflows that are infinitely replicable. Over time, the goal would be to remove dependency on the network as operators become evangelists and workflows mutate and adapt. We have enough AI influencers. I want to build AI behaviour leaders.

Apprenticeships, not another AI course

I recently attended a ‘Free AI course’ I won on CRED (because who actually wins a jackpot? NO ONE I THINK). The instructor asked the group attending what they wanted out of learning about AI. Everyone said some variation of ‘I am curious about AI because I am an xyz and I can’t figure how it will improve my work.’

No great ambition starts and stops with ‘learning AI’. The goal is always to become better accountants, lawyers, recruiters, designers, teachers, salespeople, architects, researchers, doctors. They want to be promoted. They want to make more money. They want to be better at their own jobs they’re likely pretty damn good at. Obviously, some people want to ‘learn AI’ to learn AI, but they’re already doing that all over my X feed. Serious…all over…

Supply is not a constraint in India. We have spent an enormous amount of energy and resources on AI education. Microsoft says it has trained millions of Indians in AI skills. Google has enormous developer and skilling programmes. OpenAI has Academy, the Learning Accelerator, certifications coming, university partnerships, and education products. Anthropic is doing its own education work. Add the hundreds of courses, bootcamps, workshops, YouTube channels and the guy on Instagram promising to teach you 47 AI tools before breakfast, and I am not convinced another general AI course creates much marginal value.

I think what is fundamentally missing is education that revolves around roles and functions, not the technology itself. If I tell you to teach an CA AI, you’ll start with ChatGPT or Claude, you will explain the model, what prompting means, maybe projects, contexts, agents. Tell her what a skill.md file is, show her a bunch of features, give her examples of how you would use AI, and hope that she will take all this knowledge that is in your context and apply it to her own. What if instead we focus on creating an AI-native Chartered Accountant and start with what is on her desk at that moment? What does she do? She looks at tax notices, reconciliations, client communication, compliance research, documents, reviews. She has a hundred WhatsApp texts running with clients who forget to send her invoices on time. She’s chasing after TDS documents, she is chasing after TAN numbers, and she has spent way too much time typing out “please send Form 16A.” I know all of that only because I called up my CA and asked her, “Hello, what are the things you do on an everyday basis?”

What if AI only entered after we understood the job to be done and then worked backwards from there?

This is why I would build these apprenticeships around professions or skills, and I would use the word “apprentice” very, very deliberately. A course asks me whether I understood what you taught me. An apprenticeship asks whether I can now do what you have taught me in the context of my work. The latter is a lot more difficult. I don’t think it’s useful for a CA to complete like 23 modules and answer a 30-minute quiz about prompt engineering. I want her to bring a tax notice, establish the workflow through which she deals with it today, help her rebuild that workflow, test it against multiple examples to discover failure points, and finally, decide what she is comfortable allowing AI ownership over

The curriculum would have to be modular in nature and would have to be built backward from work rather than forwards from the models. You would have to find a couple of recurring artifacts that define the skills in the job. Then find practitioners, ideally people who have already spent time on social media or YouTube teaching people how to do those things, and use them as leverage for creating teaching flows. You will work with those people, and you will essentially create teaching flows that can be diffused across networks very, very quickly and answer really pinpoint questions. The apprenticeship should ideally teach people what to delegate and what not to delegate, because one of the things that AI needs to work in tandem with is contextual judgment, and that is where the curriculum design becomes incredibly important.

This would also solve one of the weirder things about the current conversation about AI and jobs. We keep asking which professions will AI replace, as though professions are static bundles of skills that have been anointed by some divine power. They aren’t. Jobs have always adopted technologies and rearranged workflows around them. What is unusual about AI is the number of professions being asked to sort of recalibrate what “being good at my job” means. The fact is that that recalibration is a deeply uncomfortable negotiation with what you see as a part of your identity. That is why a lot of people shirk away from it.

An apprenticeship and the community that it can build gives that negotiation a collective space. Instead of telling a 25-year-old who has just done a master’s in organizational psychology, “Hey, your job is going to be replaced by AI. Good luck. You are going to be putting her in the same room as 50 excellent recruiters. You will let them figure out what parts of their work become dramatically better with AI. They publish what they learn so that the profession can learn from their experiences and do a collective negotiation of what ‘good’ looks like.”

India is also a degree- and certification-obsessed country. The apprenticeship framing allows people to get value from this process. Think of our national skills architecture. Think of how employers are still hung up on degrees. Knowing a skill and being appointed as a knower of skill are two fundamentally different things, and I don’t want to change that reality. I just want to piggyback on it. We can just provide models, technical expertise, evaluation, practitioner networks, and institutions that are already doing what they do with training and employment are better equipped to continue their work. The ambition is not that we are India’s largest AI school. The ambition is that every space that makes better, more employable, more educated people is better equipped to support their members through this mass transition into AI systems.

And the commercial reason for doing it is obviously very simple: if I learn ChatGPT or Claude through a generic AI course, Open AI, or Anthropic, I just own one of the tools I use. If I learn how to negotiate my place in my profession and world through one of these organizations, I have a much deeper relationship with them because they have participated in the construction of my new idea of good work and what it looks like. I will have simply built my way around working these goals and that is ownership over a user that new gimmicky image generation social media trends can’t buy.

So if the operator network is about finding people who have influence over work and helping their AI-enabled behaviors become fungible, apprenticeships are a layer earlier in that they help us change what a profession thinks a good practitioner looks like as the norms shift with the emergence of AI systems. We are essentially going to stop producing people who have “AI skills” and start producing people for whom using AI well is simply a part of being really good at their jobs.

Take AI off the laptop

Look, as a person who is an information worker, obviously everything I’ve described so far works primarily with information. The work I’ve described, whether it’s CAs, recruiters, lawyers, teachers, officers, etc., etc., etc., may be very different, but enough of it eventually boils down to words on a screen. There is a very clear path from AI to AiWorks that makes my life better. India, unfortunately, is not an economy made entirely of people sitting in WeWorks with MacBooks.

A bulk of the Indian economy is still deeply physical. We make things, move things, inspect things, repair things, grow things, pack things, load things, count things, and occasionally, often lose things between Bhivandi and Bangalore. Manufacturing is physical, logistics is physical, construction is physical, agriculture is physical, hospitals are physical, and even retail, which looks increasingly digital from the consumer end, has an enormous physical machine sitting behind the little “out for delivery” notification. If AI diffusion in India concentrates primarily on people whose work already happens inside the non-physical, we end up with a very sophisticated transformation for a very narrow slice of the country.

I don’t mean that OpenAI or Anthropic should start creating robots. No, that’s not what I’m saying. What I’m saying is that physical work is surrounded by information work. A nurse spending time with a patient documents what happens, and that documentation is key to her ensuring that the patient gets the care they need. A factory worker inspects a component, and somebody has to record whether it passed QA. A machine fails, and a maintenance engineer goes to diagnose it, find the manual, document the failure, figure out what has happened, and tell the next shift what was done to fix the failure so that they don’t replicate it. Everything has invoices, inventory records, exception reports, render coms, compliance requirements, WhatsApp messages going up and down asking where something is and when it will reach

The cool thing about the current generation of multimodal models is that they do not need a person standing on a factory floor to become a sophisticated prompt engineer. She can simply photograph a defect. She can speak in the language that she’s comfortable speaking in. A field technician can ask a question, looking at what they need to fix, instead of having to return to a desktop, find the right PDF, and figure out the correct instructions in the 200-page manual. The problem is that, for all the awesome demos seen in this space, the deployment has been fundamentally terrible.

That is why I would love to build something I would think of as Physical Economy Labs rather than an AI-for-manufacturing program. I would pick a small number of environments where physical work and information friction constantly leap over each other, and then people who understand models besides people who understand the physical work. I would want the AI people to discover that beautifully designed workflows simply do not work in real-life conditions, and for the operations people to discover that paperwork can transform into something that doesn’t take up half their day.

The metric of observation here is a physical event and all the information that surrounds it when:

• a machine breaks

• a shipment arrives

• a patient is just transferred between departments

• a batch fails inspection

• a field worker encounters something she hasn’t seen before

I’d like to consolidate what happens next:

• Who gets called?

• What gets photographed?

• Which form gets filled?

• Where does the information go?

• How many times is the same information entered into different systems?

• What requires human judgment?

• What can only be done because this is how the process is?

Then you introduce a multi-modal AI into that chain and see what changes. I don’t want to begin with “where can we put AI” and end up with an agent everybody’s pissed off with. The correct question to ask would be: why does this take 4 hours?

We spend so much time talking about AI making the most digitally empowered workers even more productive that it’s very easy to imagine a future in which a lawyer with a MacBook gets 10 times better at what she does (while the person maintaining the machine or making the machine remains exactly where she is). I don’t think that’s inevitable. I think a lot of AI doomers think that’s inevitable, but I am an AI optimism maxer. As voice, vision, and reasoning improve, that does not need to be an inevitability. In fact, one of the most interesting things about India as a market for AI is precisely that so much expertise in AI in India sits outside digitized systems in people’s heads. The interesting question here is whether AI can augment that expertise without flattening it into what is easiest for a chat interface.

If we’re serious about building AI for India rather than AI for the sliver of India whose working day looks exactly like ours, it eventually has to get off the laptop.

Why am I writing this?

I started writing this essay because I wanted to take a punt at working with one of the companies building the models I spend so much of my life thinking about. So why didn’t you send your CV in, Harnidh? You ask. Because AI models can’t really make sense of mine! It’s very irritating!

On paper, my career has been a basket of chaos! I have been a VC, I have championed founders, I built and ran India’s coolest program for young founders, sat up through nights making sure those founders won, stuck around when shit hit the fan, built products, wrote a national bestseller, became an influencer on the way, worked with state governments making sure people had access to sanitation facilities they were building, co-created one of the first SMB platforms for a unicorn, sold NFTs to football teams in Indonesia, created one of the biggest communities for thirtysomethings in Bangalore…phew. I am professionally incapable of staying in one industry, which has historically made it impossible to fill the “what do you do?” box on forms.

Most of my working life has involved discovering something interesting before it is obvious, figuring out who needs to care about it, and then building the machinery that helps it travel. Venture capital is obvious. You find a founder and spend a frankly ridiculous amount of time convincing other people that the future she sees is real. Building WTFund was another version of it. A huge part of the job was creating an environment in which founders borrowed context, ambition, networks and, yes, courage from each other. Writing is definitely a version of it. I take an idea that has been rattling around in my head, translate it into something another person can relate to, and then watch it escape into conversations I am no longer part of.

Turns out, I might just be professionally obsessed with diffusion.

I also think that is why AI has eaten such a disproportionate amount of my brain over the last couple of years. The models are fascinating, but I am not a researcher and I have no desire to cosplay as one. What I find much more interesting is the giant mess that begins once the model leaves the lab. What happens when it reaches the founder in HSR and the CA who thinks the founder is an idiot? What happens when a twenty-five-year-old recruiter has to decide whether using AI makes her better at her job or makes her job disappear? What happens when multimodal intelligence meets a port in Kakinada where the internet is patchy, the ERP is ancient and the documentation has to be done in hand because government records need a triplicate? One could argue that these are tech questions, but I think they are fundamentally questions about people and their incentives. Those are the questions I know how to work on.

The other thing I am very good at is being a translator of contexts. Product teams and design teams. Large institutions and twenty-two-year-olds building companies out of apartments. Brands and creators. Technology people and people who absolutely do not think of themselves as technology people. I know what it feels like to be in a room where everyone understands the jargon, and I know how quickly that jargon becomes useless when you leave the room. If the next phase of AI in India is really about diffusion, I think a lot of the work will happen in that layer of translation.

I am also very comfortable building things that do not yet have an obvious template. WTFund did not come with a manual for how to find exceptional founders under 25 across India, convince them to trust a new programme, design an experience that was actually useful to them, bring institutions around it and then figure out what needed to change after the first cohort. I like zero-to-one.I like the slightly grubby bit between a clever observation and grounded building.

I have no idea whether the three things in this essay are exactly what OpenAI or Anthropic should build. I wrote this from the outside, with public information, conversations with people around me, the companies I have worked with and whatever pattern recognition I have accumulated from spending years poking around the Indian startup ecosystem. Give me access to the people actually building these products, the organisations already using them and the data I joked about doing marginally illegal things to see earlier, and I fully expect parts of this memo to look stupid within a month. That would be useful! The point is not that I have arrived with The Four Ideas™ (three, now. We killed one). The point is that this is how I would go looking for the right ones.

I would go to the CA before designing AI for CAs. I would find the operator everyone trusts before creating an “AI champions” programme. I would spend time on the factory floor before announcing AI for manufacturing. And I would keep asking the question that underscores this entire essay: what happens after somebody discovers that the technology can do something?

I would really, really like to do that work.

If you can make an intro for me, PLEASE do!