Field notes on what it takes for a whole organisation to become AI-native: the processes it has to redesign, the people it has to bring along, and the engineering it can’t skip. From 14 years of applied statistics, machine learning and agentic AI across retail, manufacturing, banking and healthcare.
Everything I write comes back to these. They are opinions, formed by watching AI programmes succeed and stall across very different industries.
AI earns its money by removing the handoffs between teams. Bolt it onto one step of a chain that is otherwise unchanged, and the change management costs more than the model will ever return.
Then the organisation concludes that AI doesn't work. The model was rarely the problem. The half-measure was.
Read the full argument →Most processes are drawn as a straight line: plan, source, make, move, sell, serve. In reality these decisions happen at once and pull on each other. Price moves demand, demand moves inventory, inventory moves cash.
Run them as one connected system and both revenue and margin rise. Automate the line as drawn and you make the old mistakes faster. Adopting AI means changing the process first.
“You’ve never run a business.”
True. I’ve spent 14 years inside the data of businesses that do, and the pattern holds every time.
Data, pipelines, models and adoption multiply each other. Four layers that are each good enough compound into a result nobody trusts.
AI pays off when every layer is close to right. Anything less and it returns very little, however good the demo looked.
Read the full argument →Research-backed essays that take one conviction apart: the evidence for it, the evidence against it, and what a leadership team should do on Monday.
AI earns its money by removing the handoffs between teams. Bolt it onto one step of an unchanged chain, and the organisation ends up blaming the model for a failure of design.
Read →EssayPrice moves demand, demand moves inventory, inventory moves cash. Automate the process as it is drawn on the slide and you make the old mistakes faster.
Read →EssayData, pipelines, models and adoption multiply each other. Four layers that are each good enough compound into a result nobody trusts.
Read →I trained as a statistician at IIT Bombay and have spent 14 years applying data science and AI inside large organisations. Seven chapters, from cricket at BHU to owning the whole AI cycle.
September moved applied AI from assistant pilots into operating systems for work. The centre of gravity shifted to agents that can use tools, change software, monitor systems, analyse evidence and continue projects over time. The same move exposed the weak layer: authorisation, monitoring, liability and rollback. For adopters, the lesson is practical. Build AI-native workflows with security engineering and process ownership, or keep agents in controlled experiments.
The playbook an AI transformation should follow, from the first conversation to the follow-up. Personal advisory is currently limited.