For quite some time, I have been wondering that if one were to dream big, be audacious and ambitious, what would the AI-powered organisation of the future look like? Can it run by five managers? What about 3 functional execs? Or maybe even one ?
Circa Sept 2025- What is the market feeling?
There are multiple theories floating around
Around job losses
- Companies are laying off people in hopes of enhanced productivity from AI & hiring hasn’t necessarily picked up
- Companies are subsidizing cost of AI by laying off people to maintain a healthy P&L Around enterprise AI adoption
- 25-70% of code generation in enterprises is done by AI. Source
- Companies heavily laid off Customer support -> Companies who heavily reduced Customer support staff for AI support are hiring them back. Source
- AI companies are growing at an insane pace and have the highest Revenue / Employee ratio. Source
- Most of these companies are loss making
I believe the market is right on certain aspects and wrong on a few. It is often a symptom of a technology adoption cycle. But one thing is for sure - Organisations of the future will be small with large Revenue per Employee than traditional companies
My thoughts?
A million dollar company of the future is small -I think it can be as small as even one person.
This is the hypothesis I am here to test - can one person, assisted with AI, build such a company?
Why I think this could work?
There are clear market indications that companies are finding success augmenting human capabilities with AI - in the field of knowledge retrieval, guided / orchestraded tasks and many more. Research into agentic systems also indicates promises of human-level agency, coordination and capability,which a year ago were easily dismissable. This paper , which I highly recommend, explores this novel concept and has fascinating outcomes - in a sand-boxed environment , human observers outside the system were unable to distinguish between agent communication and a human proxy.
Although, external signals point in this direction, my claim is motivated by my personal experience.
For the past 3 months, I have been exploring the space. I have leveraged coding assistants like Claude Code, Windsurf and Cursor to build prototypes. I have built a production-ready IOS app without writing one line of code. Caveat- it does require knowing how to code , understanding systems and technology and properly guiding the tools to course correct when needed.
I have also played around with local LLMs, Memory and MCPs as well as deployed n8n agents to automate low complexity tasks. Here is my takeaway from this experience:
- LLMs greatly augment a skilled human - If you know what you want to get out of the system, with the right guidance, it will do it for you
- LLMs do not have agency - it requires constant human monitoring and guidance
- LLMs hallucinate. a lot.
There are ways around it- through cost controlled & guided instrumentation.
What are the factors in play?
Capability augmentation through instrumentation & guidance - LLMs today are fairly skilled, but handcuffed. With the right tools for integrations, memory, context and user guidance , LLMs can be fairly agentic. For example, take the case of a Product Review Agent - built to scan for reviews on amazon, provide analysis and generate user responses. Such capabilities can be achieved with integrations - ability to read amazon and website review, memory - details about the product , context - agent persona and response guidance, in order to generate analytics, responses and insights. Guidance comes into play with a human-guided reinforcement system to rank responses.
Build multiple such horizontal agents and you have augmented / replaced your product analytics function.
Cost Control & Arbitrage - Ofcourse, all of the above makes sense only if -
- Cost of system < Cost of payroll for the same task
- Productivity improvements > systems in place* This is a little hard to comment upon as there are upward pressures on LLM token pricing. On the other hand - economies of scale might keep the costs stagnant or reduce prices. The productivity improvements will be a bigger factor to look into. With sufficiently good accuracy, 24x7 operations and prudent cost controls - productivity gains are bound to come.
How am I testing it?
A simple playbook-
- Build a product, launch and market it.
- Identify repetitive tasks -code builds, bugs, deployment, copy-writing, marketing videos, routine maintenance etc.
- Build a horizontal agentic system to automate the same.
- Identify hand-off percentage- unguided AI operations vs hand-holding requirement
- Build continuous improvement mechanisms
- Repeat.
Heres what I have built so far -
- Bug resolution agent for my IOS application-
- Maintain a bug list document.
- The testing agent ( in claude code ) reads these bugs, builds context around the application, creates a change manifest and seeks my approval
- Post approval, the agent updates the code and builds the application.
- Any errors during application build are handled and rectified
- Updates the document with steps to test.
- Bug resolution agent for the backend application -
- Goes through server errors and creates issues on the github repository
- Cursor agent identifies the bugs and proposes code fixes.
- Agent opens a pull request to the github repository
- I review the pull request. Upon approval -redeploys the application
- In pipeline-
- Idea exploration agent - For a particular idea- do market research, technical analysis, feasibility and product documentation to create a base
- An agent to build marketing copy and videos
The jury is still out on the impact these have on my productivity but here is what I think would make these more powerful -
- abstractions so that these work for any generic application - IOS & web
- a centralized user approval mechanism - A dashboard that informs me the state of the system and pending approvals.
- a central course-correction mechanism
- a robust memory for org-wide context, control and self-improvement mechanism.
- horizontal agents to scale them beyond a single application.
I will dive deeper into the efficiency, efficacy and challenges soon!
If you found this article interesting, or have questions and thoughts - feel free to reach out. Always open to learn more, hear new ideas and collaborate.