LinkedIn X GitHub
Reading

The AI organisation

But first, I'm going to talk about how I built the AI manager.

It was not a very difficult task. First, I had to think very deeply about all the new managerial and employee traps that we live in, how the current system is flawed, and how the current tools that are used in organizations are not good. The business model relies on you spending more time on the apps like Slack and Microsoft Teams, for example.

Their KPI to measure is how much time you spend in their apps, whereas that shouldn't be the case. My AI manager app basically sees how much time you have saved rather than how much time you spend on the apps, because it is counterintuitive to productivity. The biggest plus point is that Microsoft Teams or Slack can't copy our product, because then it just becomes counterintuitive to their entire business model. Since we track our KPI, like how much time we spent on developer screens and on actual focus work. How this works is that there is an agent monitoring all your intent signals and your developer interaction touchpoints, which are:

We use this information to make out an intent map or a digital twin of the developer's brain in office hours or in work mode. How this helps is that we can predict how the developer will react to certain stuff. For example, we can also use this to further drive the claim of reducing the team meetings, which is a very useless task because

Employees waste about 30 hours a week sitting in useless meetings, and then their focus time is lost. They take a longer time to recover from that, so that time is also wasted. That is basically the time that they take to get back into their flow state.

These are some KPIs that we love tracking, like how many developer hours we saved, etc. I'll tell you how we work. We use a context ingestion pipeline. We attach webhooks to your Git commits for your accounts, and we also do authorize ACL frameworks. We limit the authorities who can access certain materials. If anything was inaccessible to an individual before the summarization or implementation of the company-wide knowledge base, then that individual is still not subject to that information. We have each user's ID, and we always cross-reference it with the user IDs that are allowed to consume that information before passing information of any sort to anybody. This means we are not prone to context poisoning and prompt injection, which might leak sensitive data even to internal employees.

So that is how we mitigate those risks. There are also other features, like direct ingestion from different channels, and we are also channel-agnostic, so we can just take information from everywhere. For the parts that we can't attach webhooks to or connect MCP servers with, we use an agent that is always shadowing your screen, which then captures everything you even consume, like documentation and any new knowledge you acquire. We use this in a very employee-protected manner. We don't want to expose what you do on your personal computers. We just want it to help us shape your digital twin better for going into these negotiation and daily sync-up meetings for you. What employees need to realize is that we design systems to work with them, not against them.

I know that some employees, just to keep their brains sharp and fueled, like to use chess and all that. That is, in fact, very nice because it tells us a lot about who you are as a person, inside and outside of work. Maybe we can later expand this product to not just be your digital twin inside of work but also outside of work, while maintaining them in a very sandbox environment that can only talk to each other through extremely secure channels. It's basically like a differentiation of your brain, but everything that is still connected, like your ground truth attitudes towards some stuff and how some stuff spills over from your personal life and your professional life. Those are actually sort of talking through a very secure HTTP or webhook sort of a system that is for future development. Maybe it's just a very rough thought. I haven't thought of it very well.

Coming to the point again, this is about the ingestion. We are actually developing an in-house brain that captures and learns about your intent. I think a very big challenge is going to be that you are essentially going to have to train a model of some sort that is damn good at capturing intent, because there are not a lot of modes in this line of business. I think anyone can copy if they are better distributed. I think a very big differentiator or mode for these sorts of businesses is the price point. For example, our direct competitors, like Glean, if they charge $50 to $60 per seat per month with a minimum quota of 100 employees, it really overspills the budget issues, which is like a $1 million ticket a year, right? That is not feasible for many engineering teams.

Also, our ICP is very different. Our core thesis is that there is an invisible wall that exists for engineering teams who are trying to scale, that is, the 50-employee wall. Before that, they act like a very close, tight group of a startup, where knowledge is communicated seamlessly. Everyone knows what the others are up to, and it works pretty well like that. After you scale past that threshold, I think there's a lot of aggressive hiring that occurs, as well as employees starting to feel friction among themselves, among their managers, etc.

How do you think, from first principles, to try to reduce these friction points? The end result is that it hampers organizational efficiency, right? You want it to be such that it doesn't hamper your organizational efficiency. How do you restore organizational efficiency by making the employees more efficient? How do you make them more efficient by probably helping them waste less time? There are certain non-negotiables in a corporate environment, right? Employees are, by human nature, going to want to take breaks, so you can't strip away their breaks because that is a very self-actualization part. If you take away their breaks, they're going to sort of just resign.

What do you take away from them? look: a manager's image in front of your employees also matters. If you take away stuff that they absolutely hate, then that is going to be a huge efficiency booster for them, thinking very objectively.

What do employees hate? Actually, I spoke to a lot of these employees who work at the exact specification of team size as I mentioned, organizations that are rapidly scaling past this 50-employee threshold start to suddenly see cracks in their company culture, or the lack of it, rather. The consistent feedback I got from these guys was that there are a lot of useless meetings that start to happen after you scale, because you have to bring all the people onto a similar battleground. They can all be on the same boat and share experiences and share expertise, which sometimes gets too tiring because an individual has to repeat themselves over a series of calls. This outcome can easily be achieved through a sort of information give-and-take, which is so easy to do in this era of agentic infrastructure, because it helps you to do negotiations at scale. This similar concept can also be applied to e-commerce systems, which I will get back to in a while.

You see, in a pre-commerce world, I don't think there was a fixed construct of a maximum retail price or a fixed price for anything. I think everything was up for a bargain, so what changed? The advent of e-commerce basically instilled a belief that everyone should buy stuff at the same price, regardless of where they are standing. This gave rise to this brand culture where the same product could be sold at multiple different prices, with an illusion of one being of better quality. This is obviously a non-confirmation bias because you can't confirm it, but you always have a bias for something that's more expensive because you associate it with better quality stuff, better quality goods, better quality materials, and, in the end, a better quality product.

I'm not here to question if that's really true. You see, pre-commerce, there was not a lot of brand culture, right? Even if there was, everything was just up for bargains, and the construct of dynamic pricing very slowly but vastly changed its meaning. Dynamic pricing basically means that you can pay different prices for the same product on paper, but it will vary with quality.

With the advent of agentic technology, the idea was to carry out negotiations on your behalf by your agent and a seller agent. The idea was to get rid of slow-moving inventory when the seller has a better proposition to sell his inventory at a cheaper price rather than keeping it for an elongated time and paying the price in opportunity cost.

I designed a system for that, but the problem was that it was really hard to break into these e-commerce sites. If I tried to do it illegally, I would always run into bot detection measures. There was no partnering with these guys because they would always try to take a cut of whatever I had.

If these e-commerce guys ever try to get smarter and they try to build something like this, just know that I already did it first. I'm pretty sure that they already have an internal system for determining whether the sellers benefit from selling their inventory, which leads to them giving discounts.

This could be spun up in a much more consumer-friendly way, where you are pleasing your consumers rather than your sellers, because in the end you are eluding your sellers. Whereas now you can elude both your sellers and your buyers, I think there is more value proposition in terms of the business model if you do it the way I am thinking.

Anyways, coming back to the point, I think that the sort of information barter that we can do with these agentic systems at a scale gives us a lot of benefits in a real-world situation. It helps us to design systems like the ones I just mentioned above for the betterment of both the parties, the buying side as well as the selling side, because essentially your work output is a direct product of their efficiency, which is a direct product of how satisfied they are, is really leveraging these agentic systems to drive consumer satisfaction as well as employee satisfaction. It is going to be a key business metric. Going forward, a lot of companies are essentially going to be built around this.

This is not to say that, since I have been describing an AI manager, it's very trivial to not assume that these are going to be the jobs taken from actual human managers. I think that it's quite a plausible future that we are very quickly moving towards. Although there may be less acceptance of the fact that we are moving towards this future, I think that the tailwinds are definitely there for something like this to absolutely become a reality.