Hi Devansh, Great article. I had similar thought process. We are currently delving in the textile industries. They also work on thin margin and I believe lot of optimizations that can happen in that industry. Overall I believe it can be applied similarly in other manufacturing like Steel, Glass.
The edge computing argument here is really compelling, especially that cost breakdown showing cloud models running 20x more expensive for continuous inference. The evolutionary valley concept also helps explain why so many brownfield operators have resisted digitizaton even when the benefits seemed obvious. One thing I'm curious about is how the Hallucination Tax plays out in practice for QSRs and construction firms. That 26% failure rate on 20-step workflows sounds catastrophic, but maybe the human oversight layer catches most of it before any real damge? Either way, the DVI framework is a solid lens for evaluating which sectors actually need this badly enough to tolerate the AI tax.
Devansh, I worked in construction for many years, ultimately as a business owner. While there's some inefficiency in home building, most of it is in management and coordination, not in the hands-on parts. Still, many hands-on tasks have been automated so that most house components are manufactured off the job site. What I'm missing in your essay is the fact that carpenters like me really enjoy our work. To tell you the truth, becoming the manager of my employees was much less enjoyable than doing the work myself.
If AI is to be a benefit then let it take over management so I can keep crafting beautiful, durable homes in collaboration with the people who will live rich lives in them and become my friends in the process.
This article is amazing and correct in many ways, but I would argue with some of the conclusions. At the current stage, I think AI works best for a business that already happens to have some tech savvy on board. For the ones you mentioned that are completely new to it, I don't think the risk calculus is lowered as much as think. A deployment can absolutely still fail with or without AI, and software maintenance is not only required but more difficult given the spaghetti code most AI vibes produce.
I am a small business owner in a field that tends to have a lot of minimum wage folks reviewing forms for data entry and compliance. Until recently I was able to avoid that aspect through a third-party vendor, but when faced with it (and only 3 months notice!) I have indeed been forced to use AI to help write the code and to embed it operationally -- and with a human in the loop at the very end. Luckily for me, I am an experienced software developer with decades of experience. We are already one of the few mid-size companies that completely own our own software stack. 90% of the rest are on platform integrators that collect their fees in various extractive ways. Those integrators are now beginning to offer AI "employees" at multiple dollars per transaction.
I don't think it will be very easy for most businesses to avoid using them and similar services.
I kind of touched on it in passing but in hindsight I should have stressed it more. I'm thinking of a replication of the iqidis model-- you have a domain expert leading the charge, backed by a good development team. The point is to build simpler software (most will likely not require domain specific reasoning systems ground up) that can be built relatively simply + deployed quickly with support engineers. I believe the key is in going beyond the standard valley to find these amazing domain operator founders.
These are some of the most thoughtful arguments Ive encountered on the subject and I've been around the block from a both technical and business perspectives.
Assuming that AI is the rising tide that lift all boats, I still think that your argument that those already exacting rents (ie the incumbent software providers) will either buy (via M&A) or copy their way to the new workflows or features given enough time even with huge tech debt, is the likely outcome.
I also think that one of the main reasons current operators want to wash their hands and not take on brownfield projects is not entirely economics, but cover your ass mentality in that a CTO selecting a software can outsource any issues and liability when inevitably shit hits the fan.
Having followed these legacy operators and also been a one time founder myself operating one of these software businesses, many end market operators want to ensure the best practices of the industry and to ensure that they are included in the workflow/features. And it's hard to know it all just from seeing from within one's own business walls and it spans regulatory but just industry best practices.
Really loved the take on how AI is going to eventually diffuse in brown field industries. But I think with open source models and reliable hardware the companies going after such brown field operations companies could be around 60-70% net margin rather 25%
Hi Devavansh, I greatly appreciated the quality of your work in bringing such valuable and well-founded information; it was enlightening. I take this opportunity to share with you a text I wrote yesterday, I believe which, in addition to complementing yours, can generate new insights and valuable exchanges between us. I hope you enjoy it: https://open.substack.com/pub/felipefrisoni/p/sensitive-empath-ai-cosmovision
Hi Devansh, Great article. I had similar thought process. We are currently delving in the textile industries. They also work on thin margin and I believe lot of optimizations that can happen in that industry. Overall I believe it can be applied similarly in other manufacturing like Steel, Glass.
very true
The edge computing argument here is really compelling, especially that cost breakdown showing cloud models running 20x more expensive for continuous inference. The evolutionary valley concept also helps explain why so many brownfield operators have resisted digitizaton even when the benefits seemed obvious. One thing I'm curious about is how the Hallucination Tax plays out in practice for QSRs and construction firms. That 26% failure rate on 20-step workflows sounds catastrophic, but maybe the human oversight layer catches most of it before any real damge? Either way, the DVI framework is a solid lens for evaluating which sectors actually need this badly enough to tolerate the AI tax.
yes
Devansh, I worked in construction for many years, ultimately as a business owner. While there's some inefficiency in home building, most of it is in management and coordination, not in the hands-on parts. Still, many hands-on tasks have been automated so that most house components are manufactured off the job site. What I'm missing in your essay is the fact that carpenters like me really enjoy our work. To tell you the truth, becoming the manager of my employees was much less enjoyable than doing the work myself.
If AI is to be a benefit then let it take over management so I can keep crafting beautiful, durable homes in collaboration with the people who will live rich lives in them and become my friends in the process.
This is a business focused article. If you enjoy carpentry, then you obviously don't have to.
This article is amazing and correct in many ways, but I would argue with some of the conclusions. At the current stage, I think AI works best for a business that already happens to have some tech savvy on board. For the ones you mentioned that are completely new to it, I don't think the risk calculus is lowered as much as think. A deployment can absolutely still fail with or without AI, and software maintenance is not only required but more difficult given the spaghetti code most AI vibes produce.
I am a small business owner in a field that tends to have a lot of minimum wage folks reviewing forms for data entry and compliance. Until recently I was able to avoid that aspect through a third-party vendor, but when faced with it (and only 3 months notice!) I have indeed been forced to use AI to help write the code and to embed it operationally -- and with a human in the loop at the very end. Luckily for me, I am an experienced software developer with decades of experience. We are already one of the few mid-size companies that completely own our own software stack. 90% of the rest are on platform integrators that collect their fees in various extractive ways. Those integrators are now beginning to offer AI "employees" at multiple dollars per transaction.
I don't think it will be very easy for most businesses to avoid using them and similar services.
I kind of touched on it in passing but in hindsight I should have stressed it more. I'm thinking of a replication of the iqidis model-- you have a domain expert leading the charge, backed by a good development team. The point is to build simpler software (most will likely not require domain specific reasoning systems ground up) that can be built relatively simply + deployed quickly with support engineers. I believe the key is in going beyond the standard valley to find these amazing domain operator founders.
You have extracted my brain and I'm impatient.
These are some of the most thoughtful arguments Ive encountered on the subject and I've been around the block from a both technical and business perspectives.
Assuming that AI is the rising tide that lift all boats, I still think that your argument that those already exacting rents (ie the incumbent software providers) will either buy (via M&A) or copy their way to the new workflows or features given enough time even with huge tech debt, is the likely outcome.
I also think that one of the main reasons current operators want to wash their hands and not take on brownfield projects is not entirely economics, but cover your ass mentality in that a CTO selecting a software can outsource any issues and liability when inevitably shit hits the fan.
Having followed these legacy operators and also been a one time founder myself operating one of these software businesses, many end market operators want to ensure the best practices of the industry and to ensure that they are included in the workflow/features. And it's hard to know it all just from seeing from within one's own business walls and it spans regulatory but just industry best practices.
Really loved the take on how AI is going to eventually diffuse in brown field industries. But I think with open source models and reliable hardware the companies going after such brown field operations companies could be around 60-70% net margin rather 25%
Eventually.
Devansh, thank you for everything you do, and your continuous efforts.
Hi Devavansh, I greatly appreciated the quality of your work in bringing such valuable and well-founded information; it was enlightening. I take this opportunity to share with you a text I wrote yesterday, I believe which, in addition to complementing yours, can generate new insights and valuable exchanges between us. I hope you enjoy it: https://open.substack.com/pub/felipefrisoni/p/sensitive-empath-ai-cosmovision