​AI Can Think. Software Still Has to Be Accountable.

Namaste Dear Reader,

Most software founders I have spoken to this year are asking/ being asked the wrong question: ‘Whether AI will eat their lunch?’

It is a reasonable fear. OpenAI is no longer just an API provider. Anthropic is no longer content to stay at the model layer. The platform companies are moving up the stack, closer to the customer, closer to the workflow, and closer to the value pool.

But “will AI eat us?” is still the wrong question because it is passive. It leads to anxiety, not strategy.

The better question is this: what can software do that a foundation model fundamentally cannot?

My answer is simple: AI can think, but software still has to be accountable.

That distinction matters more than most founders realise. And most software being built in India today is on the wrong side of it.

The model layer is no longer neutral

For a while, the default startup playbook looked obvious: take a foundation model, wrap a workflow around it, and sell the vertical product. Legal AI. HR AI. Compliance AI. Sales AI.

The model does the reasoning. The startup owns the interface. Everyone makes money. That logic worked when the model layer behaved like infrastructure.

It does not anymore.

The large model companies are increasingly shipping product features, connectors, actions, and workflow primitives that move them closer to the user. So if your entire company is built on the assumption that the model will stay below you, you do not really own your product. You are renting it.

That is why so many AI software companies feel fragile today. Their differentiation is not wrong. It is just too shallow. If your product is mainly a better prompt, a nicer UI, or a thin wrapper around model output, the platform can move toward you faster than you can move away from it.

This is why I think the key founder question is no longer, how do I add AI to my software?

It is: where in the workflow do I become hard to remove, even if the models keep getting better?

The distinction most founders miss: cognition versus consequence

Let me put the core argument directly. Most writing on this topic lands on “AI threatens software.” That is true but not useful. The more precise version is: AI threatens one aspect of software firms very clearly: Cognition

A foundation model can review a contract, analyse a ledger, draft an email, or recommend the next action.

It can often do this well.

It will almost certainly do it cheaper over time.

Moreover, these models are launching their own agents to capture even the the execution layer that software prizes today – An agent can file, settle, update, reconcile, verify, or submit information faster than humans and a software company can no longer simply be a postman.

However what the LLMs cannot do, at least not on its own, is be held responsible for the outcome.

That is the real dividing line.

A lot of founders still think defensibility in AI will come from better reasoning, depth of workflows, better prompts, or controlling the execution layer. I do not think that is where the moat settles. The moat settles where failure has a cost.

If your software is used in a workflow where errors are expensive, visible, regulated, or operationally painful, then the customer is not only buying intelligence. They are buying accountability.

That is why I keep coming back to the same line: AI can think, but software still has to be accountable.

Take payments. A model can flag fraud patterns, optimise checkout copy, or help underwrite risk. But it cannot be the counterparty that settles funds, handles reversals, manages disputes, and sits inside a regulated payment flow.

Take tax and compliance. A model can classify invoices, suggest a GST treatment, or summarise a legal provision. But it cannot be the trusted system through which a company actually files, signs off, and closes the loop with confidence.

This is the distinction I think many founders are missing. If your product mainly helps a user think better, you are competing with the model layer and everyone building on top of it. If your product takes responsibility for what happens after the thinking, you are building something much harder to displace.

What defensibility looks like in India

This is also why I think Indian founders have a better opportunity than they sometimes realise.

India’s best software opportunities are often not clean, abstract, software-only markets. They are messy, exception-heavy, regulation-shaped workflows where the work does not end at insight.

The AI can review the contract. It cannot file it with the Registrar of Companies.

The AI can suggest the GST position. A licensed professional or entity still has to stand behind it.

The AI can draft the compliance report. It cannot be the regulated institution that submits it, owns the audit trail, and bears the consequence of getting it wrong.

This messiness is not a bug in the Indian market. It is the source of defensibility.

Global model platforms will go after large, horizontal categories with massive upside. They will not go deep into every India-specific workflow, every edge-case-heavy regulatory process, or every niche operating problem where trust, local context, and execution matter more than raw intelligence.

That creates an opening.

In fact, one of the more useful questions for Indian founders is not just, how big is the market?

It is also, is this category small, messy, and local enough that the global platforms will not bother going deep, but painful enough that customers will still pay well for a real solution?

That is often a better place to build than the obvious category everyone is crowding into.

A manufacturing workflow in India is a good example. A company that sits between shop-floor systems, ERP data, quality checks, compliance requirements, and customer reporting is not just building software. It is building operational trust into a fragmented environment. That is not the kind of market a foundation model company is likely to win with a plugin.

The path that works: from wedge to accountability

The founders I feel most confident about today are not the ones with the most impressive demo. They are the ones who have chosen a workflow where trust matters early, and who have a credible path toward becoming harder to remove over time

In practice, this usually starts with a wedge that looks less glamorous than pure software investors would like.

Sometimes it is product plus service. Sometimes it is a human-in-the-loop workflow. Sometimes it is a narrow use case where the customer first buys reliability, not automation.

That is fine.

In fact, in many Indian categories, that is the right way to start.

A legal workflow is a good example. Customers do not only want faster drafting. They want fewer mistakes, faster turnaround, better consistency, and less operational friction. A product-plus-service wedge can earn trust faster than a pure tool because the buyer is not only evaluating software quality. They are evaluating whether you can actually carry part of the burden with them.

But founders should be honest about where they are.

The founder questions that matter now

When I look at a software company in this environment, these are the questions I care about most:

• What burden are you actually absorbing?
  If something goes wrong, what pain sits with you instead of the customer?

• What compounds with usage?
  Are you building proprietary workflow data, exception handling, approvals, or outcome history that a general model does not see?

• How close are you to execution?
  Are you only advising, or are you increasingly becoming the system through which work gets done?

• Is the human in the loop a feature or a crutch?
  In many Indian workflows, licensed human judgment has legal or operational standing. That can be part of the moat. But only if it helps the product move toward trust and accountability, not if it permanently hides a weak product.

• Are you model-agnostic where it matters?
  If your company is fully hostage to one model provider, your product risk is higher than you think.

These questions matter more to me than whether the demo looks magical.

What this means for software defensibility

I do think the public markets are already telling us something useful, though not in the simplistic way people often frame it.

This is not a story of “software is dead.”

It is a story of bifurcation.

The software categories that still hold value are the ones with deeper embedding, stronger retention, more consequential workflows, better data loops, or genuine system-of-record characteristics.

What gets pressured is the software that mostly sits on the read path, adds shallow intelligence, and can be substituted as the models improve.

So when founders ask whether software is still attractive, my answer is yes, but only if you are building on the right side of the line.

Not all software will be commoditised equally.

But the burden of proof is much higher now.

Where I land

Over the last year, I have become more cautious on undifferentiated software.

Not because software is unimportant. Quite the opposite.

It is because I think a lot of what used to look like software value was actually just packaged cognition. And packaged cognition is getting cheaper very fast.

The companies I am most interested in are the ones building in narrow, messy, high-context workflows where customers still pay for trust, execution, and consequence.

I like founders who understand that the goal is not just to layer AI onto an existing process. The goal is to become part of the process the customer cannot afford to rip out.

That usually means workflow ownership. It often means local context. It sometimes means product plus service in the early days. And it almost always means earning the right to move closer to the write path over time.

The question I now ask every software founder is simple:

If your product disappeared tomorrow, would your customer’s workflow actually break, or would they just replace you with another tool that looks broadly similar?

That is the test.

Most software being built today fails it.

The founders I want to spend time with are the ones building toward consequence rather than cognition, toward accountability rather than convenience, and toward workflow ownership rather than interface polish.

That is where I think the next real software value will be created.

If you are building there, I would love to hear from you.

Dhruv