Last week, I was in a discussion about how AI could improve an established business process. We soon had a long list of possibilities: automate parts of the workflow, personalise the output, reduce turnaround time, improve the reporting. All sensible, and some of them were genuinely useful.

At some point, the discussion moved to a more basic question. If we were designing this process today, knowing what AI can now do, would we build it this way at all? We had been trying to make the existing process better without spending enough time on why it existed in its current form. Some of its design reflected genuine requirements. Other parts reflected constraints that had been around when the process was created.

I have been thinking about this because we are going to run into it everywhere. Getting a decent answer to a business question has suddenly become very easy. AI models can give you a market analysis, lay out the options, challenge them and recommend a course of action before you have finished explaining the problem to a colleague. And, of course, your competitors can do exactly the same.

So, this week, I want to look at first principles thinking. It is an old idea and the phrase gets thrown around rather loosely, but I think AI has made the underlying approach much more useful. When answers are so readily available, we probably need to spend more time examining what went into the question.

Going back to what we actually know

In business, we naturally lean on what has worked before. We use experience. We look at benchmarks. We learn from competitors. If I were opening my twentieth store, I would be worried if the team insisted on forgetting what we had learnt from the first nineteen.

Where it gets interesting is when an assumption outlives the circumstances that created it. Something begins as a sensible response to a constraint and, a few years later, has become part of the furniture.

You hear this all the time in business. Enterprise sales takes six months. Consumers won’t pay beyond this price point. You need a physical distribution network in this category. This position requires ten years of experience. The Monday review needs a 60-page deck. There may be good reasons for each of these. But something does not become true simply because we have repeated it for years.

You don’t have to go deep before this becomes useful. Think about a customer-service operation. AI makes it possible to automate a large proportion of incoming queries, which is clearly valuable. But if you discover that a third of the queries are customers asking where their order is, the more interesting question may be why they need to ask. Better information at the right moment could remove the interaction altogether. Automating the query and eliminating the need for it are quite different ways of looking at the same problem.

What changes in an AI world

There is something dangerous about how good AI is at making an argument.

Imagine someone saying in a meeting, “Our customers are very price-sensitive.” Depending on who says it, somebody is likely to ask for the evidence. Put the same assumption into an AI tool and ask it to develop the strategy, and a few seconds later you can have a beautifully organised argument explaining why the company should protect the entry price point. The original assumption hasn’t become any more true. It has simply acquired a very competent advocate.

There is some evidence that we need to watch for this. A 2025 study of knowledge workers using GenAI found that people with greater confidence in AI tended to engage less critically with its output. Other research on AI and creativity has found something else worth thinking about: AI can help individuals produce better work while making the ideas produced by different people more similar.

That matters if five companies in the same industry are asking similar models similar questions based on much the same information. We may all get smarter answers without necessarily getting more distinctive thinking.

But there is another aspect of AI that I find more interesting. Many of the processes, structures and business models we use today were designed around constraints that AI is beginning to loosen. Marketing had to work in broad segments because individual communication was expensive. Customer service needed large teams because every interaction required a person. Reporting happened periodically because collecting and making sense of information continuously took too much effort. Even some organisational layers grew out of the need to gather, interpret and pass information around.

Many of those structures will continue to be useful. But before we use AI to automate an existing process, it is worth checking whether the assumptions that shaped it still hold. In some cases, the bigger opportunity may be to redesign the process rather than make the old one faster.

Why this matters for founders

I see a version of this in some of the founder conversations I am part of. A founder doesn’t necessarily disagree with the conventional wisdom in the category. Often, they know it extremely well. But every now and then they will pull at one assumption that everybody else seems comfortable with.

Take customer acquisition. If CAC in the category is ₹2,000, there is a perfectly good discussion to be had about getting it to ₹1,700. Better creatives, sharper targeting, improved conversion, lower drop-offs.

But you can also pull apart the ₹2,000. What exactly are we paying for? Is the customer unaware of the product? Do they need educating? Do they not trust us? Are we paying to persuade them to switch from something else? Once you understand that, other possibilities appear. Perhaps distribution can come partly from the product. Perhaps trust can be built differently. Perhaps AI changes the cost of education or personalisation. Perhaps nothing fundamental changes and ₹1,700 really is the right problem to solve.

At least the team has first understood which parts of the current acquisition model are genuinely hard to change, and which are not.

I like that last possibility because first principles thinking can otherwise become an excuse for clever people to challenge everything. The objective isn’t to produce an unconventional answer. It is to make sure the conventional answer has earned its place.

A few ways to make this useful

For the decisions that really matter, a few simple habits can make this much more practical.

1. Start with the outcome, not the current solution.

“How do we automate onboarding?” already assumes that onboarding in its current form should survive. “How do we help someone understand the company, meet the people they need and become productive quickly?” leaves more room. The existing process may still turn out to be the right answer, but at least you have not built the conclusion into the question.

2. Separate what you know from what you assume.

I like a very simple exercise: two columns headed What we know and What we assume. The interesting discussion usually starts when someone asks why an item is in the first column. There are often fewer things there than we initially thought.

3. Look for the constraint that created the current answer.

Was the process designed around scarce expertise, expensive labour, slow information, limited computing power or the difficulty of personalising something? Then ask whether that constraint has actually moved enough to matter. AI is changing some of these economics much faster than organisations are changing the systems built around them.

4. Ask what you would do if the current solution did not exist.

This is particularly useful when a team has spent years optimising something. If there were no weekly review, no sales demo, no physical branch network or no legacy workflow, what would you build today to achieve the same outcome? You may end up recreating much of the existing model, but you are much more likely to notice the parts that no longer earn their keep.

5. Use AI to challenge your thinking, not just improve the answer.

I increasingly use AI to test the thinking before asking it to improve the answer. Prompts such as “What am I assuming in the way I have framed this?”, “Which of these assumptions would you challenge?”, “What would have to be true for the opposite strategy to work?” or “Make the best case that my conclusion is wrong” can be very useful. AI is remarkably willing to agree with us when asked. It is also quite capable of being an awkward colleague when given that job.

6. Do not use first principles everywhere.

Most of business benefits enormously from accumulated experience, precedent and routines that work. There is no virtue in rediscovering things other people already know. I would reserve the deeper questioning for places where something important is not working, the economics have shifted, or an industry assumption is starting to feel a little too comfortable.

The way a team is managed matters too. If every review rewards the person who arrives fastest with a polished recommendation, people will naturally spend less time challenging the assumptions underneath it. For the decisions that matter, it helps to make room for someone to ask what evidence the recommendation rests on, what has been taken for granted and what would change the conclusion.

AI can now give us ten plausible strategies in the time it once took to gather the information for one. I think that makes it worth spending a little more time on what happens before we ask for those strategies: what exactly are we trying to solve, what are we taking for granted, and which of those assumptions may no longer hold.

What is one thing everyone “knows” about your business that might be worth checking again?