---
title: "The best AI strategy is sometimes no AI"
url: https://sharadbapat.com/writing/the-best-ai-strategy-is-no-ai/
description: "A client wanted AI in their workflow. I told them not to: the problem was deterministic, and intelligence would only have added cost, latency and doubt."
published: 2026-08-19
author: Sharad Bapat
---

# The best AI strategy is sometimes no AI

19 August 2026 · Essay

Sometimes the best AI strategy is not using AI at all.

A client came to me wanting to add AI to a workflow. We sat down and mapped it properly: the inputs, the rules, the exact shape the output had to take.

Halfway through, I already knew what I was going to tell them.

The rules were fixed. Nothing depended on judgment. Same input, same correct output, every single time. That's not an AI problem. That's an `if` statement nobody had written yet.

> Don't use AI here.

Which is the opposite of what I'd been brought in to say. But a few lines of conventional code would be faster, cheaper and more reliable than putting a frontier model in the middle of it: a model that, unlike a plain function, can occasionally decide to get creative when nobody asked.

## Capability is not fit

The models are genuinely extraordinary now, and that's exactly what causes this. When a tool can plausibly do anything, the instinct becomes to run everything through it.

The hammer got very good. Everything started looking like a nail.

But capability isn't the same as fit. Intelligence earns its cost when there's ambiguity to resolve, context to weigh, or a rule that doesn't exist yet, where the right answer genuinely depends on reading the situation.

When the rules are already known and the answer has to be exact, intelligence isn't a feature. It's overhead. It's slower, more expensive, less predictable, and considerably harder to debug than the five lines it replaced. You've taken something that either worked or didn't, and replaced it with something that mostly works and can't tell you why.

## Not every problem needs intelligence

Some just need good engineering.

That's the part that gets lost in the current moment. Determinism isn't a limitation you graduate out of once you have a big enough model. It's a property worth protecting. A function that returns the same answer forever is a feature. Audit trails, exact reproduction, a failure you can actually trace: these are things you give up the moment a model enters the loop, and they should be given up deliberately, in exchange for something.

The real skill in AI strategy isn't finding more places to use AI. It's being able to tell which problems were ever asking for it, and having the nerve to say so when the answer is no.