I personally sold at least $50M of SAP HANA contracted value between 2012 and 2015. The pitch was simple: workloads that took hours could now run in minutes with this new database. Sometimes 100x faster because of beautiful columnar storage rather than row-based. But the business process rarely became 10x faster. The bottleneck moved. I first learned this idea in chemistry through the limiting reactant. A chemical reaction is constrained by the ingredient that runs out first. Adding more of everything else does not increase output. I was speaking to a startup founder recently where the same question came up with AI coding. Their engineers were getting dramatically faster. If engineering becomes 3x faster, engineering may simply stop being the constraint. Now the bottleneck could be product judgment. QA. Distribution. Customer validation. That is why I would be careful with AI metrics such as percentage of code written by AI or hours saved. As a CEO, I would rather track: How fast does an idea reach a customer? or How many experiments can we run? or How quickly do we learn? AI can make a function much faster without making the company much faster. SAP HANA taught me that more than a decade ago. AI is teaching companies the same lesson again.

I sometimes wonder how little I know about the world. This experiment at scale can change the trajectory of the world

If you want to climb the corporate ladder or be a founder, eventually you need to become a beginner in a new discipline. Keep learning.

I started an ecommerce dropshipping business to become a better SaaS CEO. LTM: ₹5Cr revenue, ~10% EBIT. 0 full time employees and plenty of agents. But the real benefit was becoming the customer. When you spend your own money on ads, fulfil orders and live with the results, you start asking very different questions of the software you build or use.

Years before ChatGPT, I built a cricket statistics search product. At its peak, CillyPoint handled ~60K searches a day during IPL activity. That was exciting. But usage alone does not establish a business. Today's AI tools make impressive demos easier than ever. I still want to know who returns, who pays and how economically the product can acquire customers.

Game Theory and prisoner's dilemma. This is bad advice in my opinion. The question to ask: if your agent more powerful within product than agent and mcp/browser access?

Today I learned that Lindt has increased prices by approximately 40% since 2022. Reason: Cocoa cost. Whattay inflation

Mattel just hired a CFO from Amgen. Pharmaceuticals to Barbie and Hot Wheels. Interesting career move. 😄

Want to develop a promising young manager? Give them a small budget and a real business problem.

I wrote The Little Lessons in 2023, connecting ideas from maths, physics, economics and history to everyday decisions. One thing I think experienced operators should do more often is explain the conditions behind their advice. "Never discount" sounds wise until a carefully designed pilot is the best way to enter a market. "Always move fast" sounds wise until a decision is irreversible. Good mentorship should explain what would make the advice wrong. That is how people develop judgment.

AI is quietly changing enterprise sales. A polished 40-page RFP response used to signal considerable effort. Now almost every vendor can produce one.

Imagine your sales team finishes a quarter at 140% of quota. Everyone celebrates. The board is excited. Management begins discussing whether next year's growth plan should be raised. There is nothing wrong with celebrating. The problem begins when the company assumes that the exceptional quarter represents its new normal. Perhaps two large enterprise contracts closed together because procurement approvals arrived earlier than expected. Perhaps a competitor had a temporary problem. Perhaps one account represented an unusually large portion of bookings. Some of the performance may reflect genuine improvement. Some may reflect timing and luck. Statistics has a concept called regression to the mean. Extreme observations are frequently followed by less extreme ones, particularly when outcomes contain meaningful random variation. I think CEOs need to be especially careful about this when converting exceptional revenue performance into permanent operating expenses. Hiring twenty additional salespeople is a recurring cost. The unusually large deal that justified the decision may never repeat. Before raising the baseline, I would want to know which part of the improvement is reproducible. An excellent quarter deserves recognition. A permanently higher cost structure requires stronger evidence.

When your strongest employee suddenly stops disagreeing, do not automatically celebrate alignment. They may have changed their mind. Or they may have stopped believing it is worth arguing. The useful 1:1 question: "What have you stopped raising because nothing changes?"

Imagine 5 competing SaaS companies. One bundles an AI assistant for free. The other 4 follow because nobody wants to look behind. A year later, everyone has the feature, everyone pays the inference bill, and customers treat it as standard. Individually rational decisions can collectively destroy margin. The important question is which AI features create value customers will actually pay for.

A heroic employee can quietly make a company weaker. Imagine one engineer knows the entire billing system. Every incident gets fixed because everyone calls them. Great employee. Increasingly fragile company. Over time, others stop learning the system. Documentation suffers. Everything depends on one person. I would reward the hero. But I would also ask whether the organisation is becoming more capable because of them.

I increasingly think of management hierarchy as a compression system. A frontline team sees 100 things happen. A manager summarises them into 20. A VP turns those into five themes. By the time the CEO sees the problem, it might be one red box on a slide. That compression is necessary. A CEO cannot process every customer ticket, sales call and production incident. But compression is lossy. Nuance disappears. Weak signals disappear. Bad news gets softened. A strange customer comment that looked unimportant three layers down might have been the first sign of a major shift. So as organisations scale, I think leaders need two information systems. The compressed system: dashboards, reviews and management summaries. And a few deliberately uncompressed channels: customer calls, support tickets, skip levels, product usage, frontline conversations.

I wrote in 2014 that cloud software would verticalize because vertical products could reduce churn + use GTM spend more efficiently. I think AI is replaying that movie. The model is horizontal. The work is not. Lawyers, marketers, DBAs and network engineers have different workflows, data, permissions and definitions of “done.”