What is an AI-native firm? Sidu Ponnappa's argument, step by step

Sidu says there are two definitions. The popular one is short: an agent can pursue a goal on its own, using tools, over many steps. He doesn't think it's wrong, but he thinks it leads people astray, because it invites you to picture the agent as another person you hand work to.

The second definition is the one he actually works with. When a skilled person uses an agent well, they aren't delegating to someone else. They're using it the way you use your own fast, automatic mind. You decide what you want, and the agent carries it out, the way your hand reaches for a glass without you planning the movement.

His example is a website. Since GPT-4o, anyone can make one in minutes. But the moment you say "these are my three brand pillars, this is my design language, these are my talking points", you can no longer just let the AI build it. You spend your time pulling it back to the anchors you set.

He borrows a rule from a senior colleague at ThoughtWorks: when you design a system, you get to hammer down about three pegs, things that are hard to change, and you build everything else around them. The problem with today's agents is that they don't respect the pegs. That's true whether the system is software, sales, recruiting or communications.

When you work with an agent you make dozens or hundreds of decisions a day around your pillars. Even if you're careful, a few will slip. The slip isn't noticed, the next day's work builds on it, and the day after builds on that. People think of slop as bad copywriting. Sidu thinks that's the trivial case. The serious case is a system whose structure slowly stops matching what it was designed to do.

He puts a number on it. Today the hard question isn't whether an agent can one-shot something. It's whether you can keep building a system with it, and his answer is that you can do that for about two weeks before it degrades at the architecture level.

Maybe one day, he says. When that happens the economics of everything change, and that's one of the points you could call the singularity. He doesn't see any use in planning past it unless you're a frontier lab. What he plans for is the stretch before it.

Since almost everything will change, the useful question is which things won't. He calls these invariants. Right now a strategy lasts about three months before a model jump changes the ground under it, so he looks for things that will stay true for longer than that. Realfast is built on two of them:

• A human will be required in the loop, until the day one isn't.
• In any non-trivial system there are architectural pegs, and today's models don't respect them.

He rejects the scale. If the agent is an extension of your subconscious, then the question isn't how much you trust the tool. It's how much your own subconscious fools you. You can see your obvious biases. The agent, trying to please you, amplifies the ones you can't see.

At the mild end this means worse work. At the far end, for someone in a fragile state, he says it can mean a real psychological breakdown, and he isn't joking about that. His advice is to assume you already have some AI psychosis and to keep defending against it.

Sidu points out that we've had cyborgs for decades. Millions of intraocular lenses are implanted every year, and knee replacements give people abilities that would have been superhuman for their age and condition. What agents add is intelligence. And because the agent is you, it isn't something separate that is superintelligent. When you work with it, you are.

He's been a self-described transhumanist since ninth standard, after reading Asimov's The Bicentennial Man, about a robot that wants to become human. He realised he wanted to walk the other way, and that humanity as a whole is already doing so through hygiene, medicine and computers.

When Pranav brings up Westworld's tagline, "play God, pay the price", Sidu calls it a Western idea, and a brittle one. In his reading of Indian tradition you are born already paying the price of being alive. He thinks that makes the culture unusually well suited to what's coming.

This refers to his essay "When the Models Get Better, Your TAM Should Go Up". His answer is that a startup should pick a problem that is hard for economic reasons, not just operational ones. From here on, a business's edge comes from how well it captures what each model improvement makes possible. Businesses have always done this with new regulations or a new special economic zone. What's new is doing it every three months.

That changes how he sees pivoting. Startups had to normalise pivots because pivots were seen as failure. Even in 2022 a company that pivoted every quarter looked sketchy. Realfast spent its first two years building to a fixed thesis, and he now thinks that was an obsolete habit carried over from the previous era.

The deals they'd been turning down as "off thesis" became the business. One was a 35-to-40-year-old system built in Delphi that handles regulated corporate-secretarial work like ESOPs, with a .NET rewrite that had been in progress for 15 years. Both stacks were live, and an integration with a stock exchange had to go live in about three months. He says it would have crashed and burned on GPT-4o, would have been risky on Sonnet 4, and only became really workable with Opus 4.5.

He describes a large regulated fintech that keeps Claude Code or Codex for its engineers and has moved PR review and fixes to cheaper open-source models. He thinks that's correct, and it follows from his first point. If the model is an extension of you, then running a weaker one makes you weaker than the person next to you. For routine operations, "good enough" is the right test.

He splits companies into two kinds. Some make acceleration their core value, and he says that's what "tech company" means: Uber against a taxi company, Amazon against a retailer. Others treat speed as useful in some places but not central. Both kinds carry legacy and debt. The first kind will work through it on their own. The second kind needs partners, and that's where Realfast fits.

Then he states the thesis plainly. If the human in the loop is the one invariant, the durable business is the best organisation of humans in the loop. He knows that's the least fundable thing a startup can say.

He's careful to say this doesn't mean a body shop. It means a company that will last because of its people. He names one other invariant he can't use: frontier AI needs huge amounts of capital, and he doesn't have the ability to raise that. He also explains why India matters here. US frontier labs and big tech are absorbing the best American talent, and China won't serve the world, so India has the other large pool.

One of Realfast's core strategies is to hire from the pipeline that trained people for LeetCode interviews and big-tech jobs, which he says is "dead in the water". Those engineers have the intellectual horsepower. What a decade of ZIRP-era culture didn't give them is the habit of engaging with messy business problems directly.

The type he wants is the 1990s and 2000s software consultant: a serious hacker who sits with a gnarly business problem and solves it. At ThoughtWorks he moved from retail to insurance to an early cloud company, and between Java, .NET and Rails, inside about 18 months. Clients expected a consultant to speak their domain and know their legacy stack within four weeks. Today that role has a new name, the forward-deployed engineer (FDE).

No, he says, but the line was never where people think. Every product company has someone between it and the customer who understands the business better than the product team does.

Yes, he says, but nobody has worked out how to deliver it, and the reason is another invariant. Consumer companies know that every extra decision you ask a user to make lowers conversion. Adding AI to a workflow increases the number of decisions the human has to make.

A small group of people, spread across every function, find that energising. Everyone else feels the same anxiety they get from a bad enterprise app. People are fine with chat for summaries and light drafting. The moment the tool becomes something they have to learn to think with, engagement collapses. Board pressure can close the first sale, but retention will be poor.

Realfast deliberately sells to companies with $50M to $500M in revenue, where it can reach the board and only takes projects the board cares about. The reason is that the hard problem isn't the software. It's change management. "If I crack that I have a trillion dollar company, not a billion dollar company."

Pranav suggests AI budgets are moving to managers who each own a metric. Sidu says that's a myth. In consumer companies every manager owns a number. In B2B, outside support and parts of sales, almost nobody does, apart from blunt ones like deadlines or the month-end close.

So enterprises copy the consumer playbook without the consumer machinery. The CIO runs a small rollout, the CFO promises more budget once impact is proven, and then nobody can prove it.

What changed is the benchmark. A senior non-technical leader goes out for drinks with five peers, one of them has actually built something with AI, and now the CFO doesn't want a two-year roadmap. The CFO wants to know what the 1% rollout changed. "This was never a reasonable expectation before."

Big tech will go through a predictable sequence: a token leaderboard appears, people max their usage, quality control arrives, it breaks down, best practices spread, and it gets tied to performance reviews. Enterprises haven't yet accepted that proving impact on short cycles means changing their performance reviews.

On productivity he goes to operations research. Toyota beat everyone because it understood that speeding up one station on an assembly line often lowers the throughput of the whole line. You get more friction, more inventory and fewer finished products.

So is AI accelerating anything? Very clearly, he says, but you can only see it in the frontier labs and in pockets of big tech. The depth Claude Code has gained in a year would be "humanly impossible" for a team starting cold, and OpenAI goes from decision to shipped product in about nine months. He isn't arguing about whose harness design is best. His point is how much they ship.

The real reason is the innovator's dilemma. A 30-year-old company with $5B in revenue has fitted itself closely to its market over decades, and rebuilding it around acceleration would damage that fit. So the realistic move is to bolt something on. "And I am that bolt on. Bolt me on."

He isn't sure about the timeline, but the trend over the last 30 years is clear. More and more of the Fortune 500 are companies whose strategy rests on acceleration. Whether that applies to an oil and gas company was always debatable, because software can't make the drill bit faster.

What's different now is where the speed comes from. AI makes each leader much more capable, not only the operations. Better dashboards and data pipelines have always helped leaders decide. What's new is making the leader personally superhuman.

Pranav compares this to earlier management ideas that became core pillars, like Toyota's kaizen and Ford's Taylorism. Sidu then describes his own dilemma. His deepest market is BFSI, where regulation rightly limits how fast anything can change. He could chase that deep revenue pool with acceleration capped, or go after shallower pools where he can prove that a customer beats its competitors using AI.

He asks the hosts to name one enterprise that is beating its competition because of AI. They can't. Where AI has spread most widely, in support and inside sales, it was used to cut costs, and customers got a worse experience. Nobody's support got better.

OpenAI and Anthropic show that it can be done: they ship products people enjoy, quickly, and everyone from the board down believes in it. Enterprises buy on logos and case studies, so they need to see one of their own do it first. He's confident that in each vertical one company will break the pattern, and it won't win by a small margin.

Pranav brings up Sidu's tweet about a salesperson building a working demo during a sales call. Sidu explains why this matters. In traditional IT services, software projects usually fail, the customer pushes the risk onto the vendor, and the vendor's only lever is staffing. So the vendor needs a 2-to-5-year contract to make money, and success is measured by a sign-off that "the CRM was adopted", not by hard numbers.

Measuring real usage, like where a field salesperson gets stuck and whether that workflow turns into a closed deal, was never worth the cost. Now it is, and it needs a different kind of talent.

The enterprise doesn't yet believe outcomes like this are possible, so he sells only into problems with a clear line to revenue or cost.

On intellectual property he is blunt. AI was trained on IP that nobody paid for, and the result is that most IP has lost its value except at the very top. His example is the art director in Delhi Belly who asks for the banana to smile 7% more. The craftsperson who had to work out what "7% more" meant now has a job with no economic value. The same is true of building simple websites.

Tacit knowledge in knowledge work is also becoming explicit, because digital work is easy to observe. The conventions a new consultant used to pick up over weeks now live in the repository and the harness: how commits are written, how tickets and requirement documents are structured. They also live in the dozens of small scripts that run a modern project's workflow.

He says this is an open question that he now spends real time on. Fifteen years of his onboarding playbooks no longer work. You can't do "wax on, wax off" drills if the agent does the drill for you, and working with agents needs a discipline he doesn't have mental models for yet. He's sure about one thing, though: the AI course is the wrong idea.

I think this is the most important idea in the episode. Most of the economy runs on an unspoken agreement: for this team and this workflow, average talent and average results are fine. India's education system reasonably produces large numbers of people who are good enough.

Give a good-enough person an agent and they suddenly face weeks' worth of decisions in a single day. Is the requirements document right? Are the acceptance criteria there? Are the tests written? Is the log clean? Decision fatigue arrives early, they start giving up, slop gets in, and slop compounds into debt.

He doesn't blame anyone for this. Producing a million IIT-grade graduates a year was never possible, and aiming for good enough was a sensible strategy. It has just stopped working.

He sharpens his position. He isn't pessimistic about sovereign AI. He's pessimistic about trying to copy the US or China, which are $20–30 trillion economies with social systems nobody else has. India is in the $4–5 trillion group with Germany, Japan and France. The sensible goal is to beat them and be a clear third.

India's Fable access was blocked, which shows how a foreign lab can move a country to a lower tier overnight. So India has to build the muscle, starting with properly funding Sarvam and more companies like it. It also has to fix research. He says 60% of notable research papers now come from China, while India's best researchers leave because their breakthroughs are valued elsewhere.

Why India struggles to innovate at scale: an ecosystem gets started either by technocrats at the top, as in China, or by stubborn, rebellious individuals at the bottom, as in the US. India has neither. Asked about academia-industry links, he's just as direct. Outside a few exceptions like IIT Madras and IISc, "there is no academia here, there are teachers."

He's honest that he doesn't know. Businesses that deal in physical things should keep getting more efficient. Bangalore, though, depends on one industry. All the startups together won't add more than 2–3% of the headcount sitting in IT services. Bombay has six industries to fall back on, and Bangalore doesn't.

He's also clear about how hard the change is for a large IT services firm. Its whole talent supply chain, from campus hiring through its training centres, would have to be disrupted. Making one team truly AI-native means doing all of the following:

• fork the compensation bands
• cut review cycles to one quarter
• move the team into a new management structure that reports to the board, not necessarily the CEO
• scale it while it eats into the existing business without spooking investors, and "Indian investors are easily spooked"

Source: "What is an AI-native firm? ft. Sidu Ponnappa", Subtext by Zerodha, published 2 October 2026, hosted by Pranav Manie with Bhuvanesh R. Screenshots are frames from that video.

The quotes come from YouTube's auto-generated captions. I've corrected obvious caption errors (for example "real fast" → Realfast, "Opus 45" → Opus 4.5, "Lindy" spelled out, "Serbam" → Sarvam) and used "…" where I cut words. I haven't otherwise changed wording, so you'll see Sidu's spoken grammar. Timestamps are approximate to within a few seconds.

The diagrams are my sketches of his arguments, not slides from the show.