I use AI almost every day now. ChatGPT is probably the most obvious one, but it doesn’t stop there. I use RAG when I need to work with a large amount of information and want the AI to work against the actual documents rather than simply give me something it thinks is correct. I experiment with AI video generation using tools such as Seedance 2.5, use AI for research, analyse documents, structure presentations, challenge business ideas and sometimes simply throw a half-formed thought at it to see whether there is something I am missing. The productivity gain is real. Things that earlier took hours can now take minutes. But there is one rule I have started following very consciously: I don’t let AI do my thinking for me.

I think this is becoming an important distinction. We are entering a phase where AI can give us an answer to almost anything. Ask it to write a strategy and it will write one. Ask it to analyse a problem and it will give you an analysis. Ask it for ten ideas and you will get ten ideas in seconds. The problem is that we can very easily stop thinking ourselves. And that, for me, is a much more immediate danger than some of the more extreme AI scenarios we keep reading about. If I let AI do the thinking, I may become faster, but I am not necessarily becoming better.

There is another problem I worry about even more: AI can slowly kill originality. Give the same problem to a hundred people and let AI write the answer, and you will probably get a hundred beautifully written answers which sound surprisingly similar. The language will be polished, the headings will be perfect, the recommendations will be logical and every paragraph will talk about “unlocking value”, “driving transformation” and “leveraging synergies”. The problem is that after a while everything starts sounding the same. I don’t want that. I would rather have an imperfect thought that is genuinely mine than a perfectly written thought that could have been produced by anyone.

I come from operations, and operations has taught me something which I find extremely useful when working with AI. The problem is rarely that we don’t have enough information. The problem is knowing what actually matters. I have worked in Africa, North America and India, and the operating environments have been very different, but this principle has remained the same. You can have an ERP, WMS, TMS, Power BI dashboards and now AI sitting on top of all of them, but if the basic operating logic is wrong, you are simply using technology to produce a more sophisticated version of the wrong answer.

Take something as simple as operating expenses in a warehouse. When OPEX goes up, the immediate temptation is to start looking for technology, automation or manpower optimisation. Sometimes that is absolutely required. But many times I have found that the first question should be much simpler: what exactly are we spending money on, and why? Are we using the manpower because it is actually required or because the roster has always been like that? Are we moving material twice when it only needs to move once? Are we paying for capacity that nobody is using? Are people performing an activity because it adds value or simply because an old SOP says so? Sometimes the biggest OPEX reduction doesn’t require AI at all. It requires someone to stand in the operation and ask, “Why are we doing this?”

This is probably where my approach to AI has evolved the most. Before I ask AI to solve a problem, I try to solve it myself. I look at the numbers, understand the process, speak to the people involved and form my own hypothesis. Then I take that thinking to AI and ask it to challenge me. I may ask it to identify something I have missed, analyse a larger dataset, compare scenarios or find a pattern across hundreds of documents. In other words, I don’t use AI as the person sitting in the driver’s seat. I use it as the very smart person sitting beside me saying, “Have you thought about this?”

That is also why I like the RAG approach. If I am working on a real business problem, I don’t want an AI model giving me generic management-consulting answers based on whatever information it has seen during training. I would rather give it the actual contracts, SOPs, warehouse reports, historical correspondence, financial numbers, tender documents and operating data and ask it to work within that context. Suddenly the conversation becomes very different. AI is no longer giving me generic advice about “supply-chain excellence”. It is looking at my actual operating environment and helping me find relationships, gaps and anomalies that may otherwise take hours to uncover.

Working in Africa taught me another side of operations. You learn very quickly that you cannot build a business assuming everything will work exactly as planned. Infrastructure can be unpredictable, information can be incomplete, supply chains can get disrupted and sometimes the solution has to be created on the ground rather than pulled out of a textbook. You learn to think in terms of contingencies, redundancy, practical execution and what I would call operational jugaad — not cutting corners, but finding a workable solution when the textbook solution isn’t available.

6

North America gave me a very different perspective. At scale, small inefficiencies become big numbers. One extra minute in a process doesn’t sound like much until it is multiplied across thousands of packages. One unnecessary touch doesn’t look expensive until you multiply it across millions of movements. A few points of productivity improvement can translate into significant financial impact. This is where AI becomes extremely powerful because it can process huge amounts of operational information and identify patterns much faster than a human team can. But even there, the experienced operator still has to ask the uncomfortable question: is this actually the problem worth solving?

The KPI itself can also become a trap. Operations people love numbers — PPH, UPH, OTIF, turnaround time, inventory accuracy, utilisation, cost per MT, cost per order, shrinkage and so on. AI can analyse all of these beautifully. But the KPI is not the operation. It is only a representation of the operation. If I improve labour productivity but create a safety issue, I haven’t improved the operation. If I increase warehouse occupancy but trucks start waiting outside, I haven’t necessarily improved the business. If I reduce manpower but create bottlenecks somewhere else, the saving may simply have moved from one line of the P&L to another. AI can optimise the metric. The operator has to understand the system.

I also enjoy using AI on the creative side. I have experimented with tools such as Seedance 2.5 to take an idea in my head and turn it into something visual. Earlier, if I had a concept for a video, there were several layers between the idea and the finished output. Today I can think through the story, create references, experiment with scenes and get something on screen very quickly. That is where I see AI at its best. It removes friction between an idea and execution. But the idea still needs to come from somewhere. If I simply tell AI to “make me a great video”, it can probably make something visually impressive. But if there is no thought behind it, it is just another nice-looking video.

And then there is the bigger question — what happens when AI itself becomes much more capable than the systems we are using today? This is where I think we should take the warnings from people working at the frontier seriously, without getting carried away by the science fiction. There are genuine debates around autonomous AI agents, cyber capabilities, biological risks, recursive self-improvement and whether safety mechanisms can keep pace with rapidly increasing capabilities. People such as Dario Amodei and other researchers have publicly argued that frontier AI development needs stronger evaluation and coordination mechanisms. I don’t know where all of this eventually goes, and I don’t think anyone honestly does. But I do know one thing from operations: when the consequences of failure are potentially enormous, you don’t wait for the failure to happen before building controls.

The same principle applies inside a company. I wouldn’t allow an AI system to make an important operating decision simply because the algorithm says so. I would want controls, validation, exception management and human accountability. The same applies to procurement, inventory, finance, manpower planning or customer commitments. AI can recommend. AI can flag. AI can analyse. AI can even automate a decision when the risk is understood and the boundaries are clearly defined. But somebody still needs to own the outcome.

So I am not anti-AI. Quite the opposite. I probably use it more today than I ever expected I would. But I have become more conscious of where I want it involved and where I don’t. I want AI to make me faster, not lazier. I want it to challenge my assumptions, not create my assumptions. I want it to analyse more information than I can physically process, but I still want to understand what the information means. I want it to make execution seamless, but I don’t want it deciding what is worth executing in the first place.

Perhaps this is where the next generation of operating leaders will have to differentiate themselves. It won’t simply be about who knows the latest AI tool or who can write the cleverest prompt. It will be about who can still look at a messy operation, a pile of conflicting data, an unhappy customer, an expensive process or a difficult commercial decision and figure out what actually matters. The ability to think clearly may become more valuable precisely because machines are becoming so good at producing answers.

And that is why my approach is quite simple.

I put more pressure on my own head to think first. Then I let AI do the heavy lifting.Because AI can make my work seamless.

But I still want the thinking to be mine.

– Karan Bajaj