You can't AI your way out of a broken base
Every founder now asks how AI will transform their finance function. The honest answer starts with a warning: pointed at a broken base, AI doesn't fix the problem, it accelerates it. Sriram Chidambaram on the precondition that decides who gets value from AI.
Summary
- AI genuinely compresses the bottom of the finance stack: gathering data, reconciling it, and drafting the first report.
- The number one blocker to any return on AI is data readiness. AI is an amplifier, not a corrector, and on bad data it produces confident wrong answers at machine speed.
- The sequence is not negotiable: fix the base, then automate.
Every founder I talk to now asks some version of the same question: how is AI going to transform our finance function? It's a good question. But it's usually the wrong first one, and the honest answer starts with a warning rather than a promise.
Here it is, as plainly as I can put it: you cannot automate your way out of a broken base. AI is genuinely powerful, and it is genuinely coming for finance. But pointed at bad data, it doesn't fix the problem. It accelerates it. Before you spend a rupee on AI, this is the thing to understand.
Let me give AI its due first, because the opportunity is real.
What AI is genuinely good at
The most time-consuming, least valuable work in any finance function happens at the bottom of the stack: gathering data, reconciling it, and producing the first draft of a report. It's where finance teams lose most of their week, and it's exactly the kind of work AI compresses beautifully.
Management reporting and variance analysis are, by most accounts, among the fastest-paying uses of AI in finance right now. The surveys of PE- and VC-backed finance leaders put the at a matter of months, not years. Deloitte's CFO research shows the overwhelming majority of finance leaders now consider AI important to how the function will run.
So the prize is real: done right, AI hands your finance team back the hours they currently lose to fetching and formatting, and lets them spend that time on the work that actually matters (the meaning, the decisions, the judgment). That's a genuine unlock, and I'm not remotely a sceptic about it.
But notice where that unlock happens. It compresses the bottom three layers of the stack: the data, the measurement, the first-draft analysis. And it only works if those layers are built on something solid.
The precondition nobody wants to hear
The same surveys that are bullish on AI's payback name the same blocker, again and again: data readiness. Not talent, not budget, not the technology. The number one thing standing between finance teams and any real return on AI is the state of their own data.
This is the part founders don't want to hear, because it's not exciting. The bottleneck isn't the AI. It's your Source layer: whether your systems agree on what a number means, whether your definitions are locked, whether your three data streams reconcile. If that foundation is broken, no model on earth will save you, because the model can only work with what you feed it.
And here's the trap. When your data is a mess, the instinct is to reach for AI as the fix, to hope that a clever enough tool will impose order on the chaos. It's precisely backwards.
AI is an amplifier, not a corrector.
Feed it a clean, coherent base and it amplifies clarity. Feed it a broken one and it amplifies the mess: faster, at scale, and far more convincingly than any human ever could.
Wrong answers, faster and more convincingly
That last point is the one I want to sit on, because it's the real danger.
A human analyst working with bad data at least tends to hesitate. They sense when a number looks off; they add a caveat; they say "let me check that." AI does no such thing.
Point it at data where "revenue" means three different things and it will produce a confident, beautifully formatted, entirely wrong answer. It will do it in seconds, and it will do it every time you ask. It removes the very friction that used to catch errors.
So the failure mode of AI on a broken base isn't that it doesn't work. It's that it works persuasively on the wrong inputs, and hands you confidence you haven't earned, at machine speed. That's how a company makes a fast, bold, well-presented decision that happens to be based on a number that was never real.
Bad data was always dangerous. Bad data with AI on top is dangerous at scale.
Autopilot is only as good as the instruments
I think about the whole information stack as a cockpit: the instruments a company flies on. AI, in that picture, is autopilot. And autopilot is a genuinely wonderful thing: it takes the exhausting, repetitive work of flying off the pilot's hands and does it more consistently than a human could.
But autopilot is only ever as safe as the instruments feeding it. Connect it to a faulty altimeter and it will fly you, smoothly and confidently, straight into the mountain, precisely because it trusts the reading and doesn't hesitate.
No pilot would engage autopilot on instruments they didn't trust. Yet that is exactly what companies do when they layer AI onto a Source layer they haven't fixed.
The sequence is the whole point
Which gives you the sequence, and it's not negotiable: fix the base, then automate. Get your Source layer clean: definitions locked, streams reconciled, one version of the truth held by a neutral owner. Then point AI at it, and watch it compress the grunt work and give your team its judgment back.
Do it in that order and AI is transformative. Do it in the other order (automate first, in the hope of skipping the unglamorous foundational work) and you've built a faster way to be wrong.
There's a broader shift worth naming here too. As AI makes real-time reporting normal, boards and investors are starting to expect it, and the days-late monthly pack is quietly becoming unacceptable. But real-time doesn't help you if the underlying numbers are wrong; it just means you're wrong faster and more often. The rising bar makes the foundation more important, not less.
The unglamorous truth
So the honest answer to "how will AI transform our finance function?" is this: enormously, but only after you've done the boring work first. AI raises the ceiling on what a great finance function can do, and it raises the penalty for a weak base at exactly the same time.
The gap between the companies that get value from it and the companies that get expensive, confident nonsense won't be the quality of their AI. It'll be the quality of the data underneath it.
The winners in the AI era of finance won't be the ones with the cleverest tools. They'll be the ones who did the unglamorous work of building a base worth automating. Fix the foundation first. Then let the machines fly.
How relevant and useful is this article for you?
The next one
Get what we publish next, by email.
Working notes on raising, borrowing, protecting, growing and structuring capital in India. One email a week at most, and you can leave any time.
We use your address only to send this. See our privacy policy.
We store your address to send you these emails and nothing else. See our privacy policy.
Related reading
