𝐖𝐢𝐥𝐥 𝐀𝐈 𝐞𝐚𝐭 𝐚𝐥𝐥 𝐧𝐨𝐧-𝐀𝐈 𝐬𝐨𝐟𝐭𝐰𝐚𝐫𝐞? The conventional AI narrative — "AI will eat all software" — is not wrong, but it is too blunt. 𝘛𝘩𝘦 𝘴𝘵𝘢𝘳𝘵𝘪𝘯𝘨 𝘢𝘴𝘴𝘶𝘮𝘱𝘵𝘪𝘰𝘯: all software, existing and new, will be AI-updated or AI-generated in the coming years. That part is inevitable. But not all software is affected the same way. I can think of four buckets to categorize software companies: 𝐈𝐦𝐦𝐮𝐧𝐞. Software that physically controls the real world — power grids, payment rails, building management systems, manufacturing plants, avionics, hospital infrastructure, hardware devices, etc. AI improves things around it and will enhance the features and functionality faster. Incumbents will continue to thrive. 𝐄𝐧𝐡𝐚𝐧𝐜𝐞𝐝. Software with deep data gravity, complex workflows, and wide integration surface area, spread across the entire organization. SAP, Oracle, CrowdStrike, Cadence, and many more. AI makes these platforms stronger, not weaker. They are deeply embedded in companies and the majority of their revenue comes from non-software enterprises. Companies in this category have strong incumbent advantages as they absorb AI. 𝐂𝐨𝐧𝐭𝐞𝐬𝐭𝐞𝐝. Internal-facing software that helps humans do things inside the company — project management tools, legacy dev tools, data analysis platforms, low-code builders, SMB CRMs. The pressure here comes from two directions. From above, agentic AI platforms such as Claude Cowork that can absorb these product features directly. From below, engineering teams that can now build custom alternatives in weeks — a tailored ticketing system, a lightweight CRM, shaped to exactly how they work. At enterprise scale, incumbents still win on compliance and coordination. But for startups and mid-size teams, the "just build it" option is real now. Some will evolve. Others will be absorbed. 𝐑𝐞𝐩𝐥𝐚𝐜𝐞𝐝. Software whose core value was applying intelligence to existing content — it helped humans understand or process things. Survey analytics, contract review, support ticket routing, standalone BI dashboards, simple website builders, marketing analytics tools, landing-page builders. LLMs can do this natively now. The most interesting bucket is not the one being replaced. It is the contested one — because that is where the fight for the future of software tooling is happening right now. How do you see the categories? As founders, how do you see it?
Neeraj K.
1mo
This framework maps very cleanly to medical imaging when you think in layers. The “immune” layer is the hardware + acquisition stack (scanners, core control systems), still dominated by vendors and strengthened—not replaced—by AI. The “replaced” layer is where imaging software was mostly about interpreting images—standalone analytics, basic AI tools. LLMs and foundation models compress this quickly. The real opportunity is in the “contested” middle layers: converting signals into quantitative physiology translating that into clinical decisions embedding those decisions into workflows This is where incumbents are constrained by legacy systems, and LLM players lack context and integration. So the wedge for new companies in medical imaging is not “better AI models,” but: controlling how measurements are derived owning the decision layer and learning from real clinical workflows over time AI doesn’t eat all imaging software—it reshapes the stack. And the biggest companies will be built in the layers where measurement, decision, and workflow come together.
Adoption isn’t about capability. It’s about intent vs constraints. Bucket 1: Can, but won’t Enterprises sitting on legacy systems. AI could help. But risk, accountability, and “what if it breaks?” win every time. Bucket 2: Want, but can’t Teams excited about AI. But blocked by cost, clarity, and more often — human behaviour. Unclear thinking. Misalignment. Conversations that should happen… don’t. So we blame the tech. But the real gap isn’t AI. It’s how humans decide, collaborate, and act under uncertainty. That’s the problem we’re trying to solve with CoPrompt. Not better answers — but better thinking, before the answer.
Ashwin Ramasamy
1mo
Shekhar, I am curious why the industry is still limiting its thinking to ‘serving’ customer industries instead of building AI-native companies in every vertical. The terminal value of software companies has changed. It will become easier to get a PE outcome in a traditional industry with an AI premium than to get a pure-play software co. to an exit.
Chandra Kumar
1mo
The now famous saying that AI will eat all software is more about AI becoming the first port of call rather than replacing software or generating new ones. All the 4 categories are collapsing into one category of becoming an API for AI to consume. The only moat legacy apps have is control over internal data and I see pressure arising from clients to open it to AI, not waiting for compliance and other areas to get sorted. Will incumbents be ok with this new way of working is a Hobsons choice at best.
Anuj Kathuria
1mo
This is a useful framing, especially the contested layer. The shift, however, is not just about which software survives. It is about what layer of value survives. AI is collapsing the distinction between: • application logic • workflow orchestration • user interface Which means many products will not be replaced by competitors. They will be absorbed into decision and execution layers. I believe the real question is not: Will this software be replaced? It is: Is this product a system of record, or just a system of interaction? Systems of record persist. Systems of interaction are being unbundled rapidly. The winners will not be the most AI-enabled products. They will be the ones that own critical workflows, data gravity, and decision rights inside the enterprise. Everything else becomes optional.
Mid to Large Enterprises will always be slow to adopt AI across their tech stack as the risk of going south is higher due to complexity. Functional roles such as Marketing/HR/Sales are seeing more adoption. Its more universal in SMBs. Outcome based Services is leading to a big boom for Service as a Software business model as a result of AI becoming more prominent.
Pratik Sharda
1mo
Strongly agree with this - especially the “immune” bucket. From what we’re building at CraftifAI, software for hardware systems (devices, robotics, industrial infra) isn’t something coding tools/IDEs can just replace. The real-world constraints like latency, reliability, heterogeneous compute - make this a fundamentally different game. What AI is doing instead is amplifying this layer. We’re seeing this firsthand - enabling teams to build and deploy AI-native pipelines across GPUs, MCUs, and FPGAs without getting lost in hardware complexity.
Vijay Bathula
1mo
Thanks for the breakdown. I see these categories as a shift in the "Biggest Function" of software—moving from managing a process to simply defining an outcome. To your point, the "Contested" and "Replaced" buckets are driven by a collapse in friction. Take Consumer B2C (Food Delivery) for example: The "Old Way": Hunting for paper menus and calling over a noisy phone line. The "Software Way": Scrolling through marketplaces like DoorDash or Uber Eats. The logistics are invisible, but the human still does the "work" of searching. The "AI Way": Intent-based fulfillment. "Get me something healthy for two under $40 by 7 PM." The function moves from Browsing to Decision Fulfillment. We can map this to almost every use case, but with frameworks like OpenClaw showing how easily agents can be spawned, the cost of entry to market is plummeting. It’s exactly what Jensen Huang highlighted at GTC: we are moving from asking "how" to asking the machine to "create, do, and build." In the B2C space, the standalone "utility app" is being replaced by the outcome itself.
Sagar Das
1mo
Shekhar, this is one of the most grounded frameworks I've read on where AI actually lands — not where we want it to. The Contested bucket gets all the anxiety. It deserves more of the optimism too. Pressure at both ends — agentic platforms from above, 'just build it' from below — is the best forcing function this category has seen in years. It's pushing every product to answer a harder question than 'does this work?' The new question is: 'Why does this need to exist at all?' That is not a threat. That is a gift. The software that survives this moment won't survive because it added an AI layer. It will survive because it finally got honest about the one thing it does that nothing else can replicate — deep workflow logic, institutional memory, coordination at scale. The contested bucket won't be won by the biggest or the oldest. It will be won by whoever makes the answer to that question the most obvious. That clarity is what this moment is producing. And that's worth being genuinely excited about.
In services first companies, there is another shift underneath this. As business models move from FTE- and seat-based delivery toward outcome-based delivery, the software stack has to change as well. A lot of existing software was built around standardized human workflows. Outcome-based models need software that is much closer to the operation itself. AI changes the make-vs-buy equation here. More of that software would be built much closer to where the work happens.