# The universal hammer

Author: Florent Martineau
Language: en
Published: 2026-09-30T00:00:00.000Z
Updated: 2026-09-30T00:00:00.000Z
Canonical: https://florentmartineau.com/en/blog/the-universal-hammer/

Does your problem need AI, or just a good old Excel formula?

![Oprah gives everyone an AI.](https://florentmartineau.com/images/blog/oprah-three-panels-blank.svg)

“You get an AI!”  
“You get an AI!”  
“Everyone gets an AI!”

**You have an Excel spreadsheet with 10,000 rows. You need to multiply two columns, row by row.**

**Calculate everything by hand.**

Good luck!

[Bruce Almighty typing at full speed](https://florentmartineau.com/images/blog/bruce-typing.mp4)

**Write =A1\*B1 and drag the formula down.**

You are done before the coffee is ready.

[A child running at full speed: I am speed](https://florentmartineau.com/images/blog/i-am-speed-kid.mp4)

**Ask your favourite AI to calculate every result.**

Sorry, you have exhausted your quota. It will reset in 6 days and 9 hours.

**Answer D.**

Is that your final answer?

[Gad Elmaleh and Olivier — Who wants to win loads of money?](https://www.youtube-nocookie.com/embed/Ll9kvTQ842E)

The formula: cheaper, more reliable, faster. A win all round.

**You have a list of 1 million messages to sort into five categories. The first 5,000 have already been classified.**

**Classify everything by hand.**

I didn’t come here to suffer, okay?

[I didn’t come here to suffer, okay?](https://www.youtube-nocookie.com/embed/dUzYKDgPzjg)

**Train a small model on the 5,000 examples.**

A classifier. What if that were enough?

![A classifier. An agent with twelve tools and three subagents! A classifier.](https://florentmartineau.com/images/blog/bell-curve-template.png)

A classifier. — An agent with twelve tools and three subagents! — A classifier.

**Ask your favourite AI to classify each message.**

I’m financially comfortable.

[I’m financially comfortable](https://www.youtube-nocookie.com/embed/aIVsz5Pj0eE)

**Ask Jev to do it.**

Fast, cheap and easy to set up.

![A man turns to look at the woman in red, labelled Jev, while his girlfriend, labelled Classifier, watches.](https://florentmartineau.com/images/blog/distracted-boyfriend-template.jpg)

Jev — Classifier

**Is an LLM always the right tool for your problem?**

## A bazooka to kill a fly

An LLM can discuss the role of calves in German literature, give you a strawberry tart recipe, [hack Australian health insurance](https://www.pm.gov.au/media/press-conference-new-york) or [count the r’s in “strawberry”](https://arxiv.org/abs/2412.18626).

Do you really need to know how to translate Ancient Greek into Armenian to program in Rust?

**Your mission, should you choose to accept it: classify emails using a single tool of your choice. The LLM is one of the options. With the same budget, the best tool for this task will make, compared with the LLM:**

**More errors.**

You can choose the LLM as your tool.

It is one of the options, so the best available tool cannot perform worse than it.

**The same number of errors or fewer.**

The LLM is one of the options.

The LLM could be the best possible tool. If so, use it! Otherwise, use the right tool.

**Definitely fewer errors.**

The specialised tool does not necessarily do better.

An LLM may be the best solution. If so, use it. Otherwise, use a specialised tool.

Stockfish plays chess, AlphaGo plays Go. We do not ask them to write a cover letter.

**If you have a specific task in mind, does your tool need to know how to do the endless other possible tasks?**

## A prototype is not a product

**Twenty clients ask you to send follow-up emails. What do you do?**

**Create 20 agents and add the follow-up to their prompts.**

Twenty prompts to write. Twenty behaviours to make reliable.

If the follow-up rule changes, you will need to update and verify it in each of the twenty agents.

**Write a different automation for each client.**

Twenty clients, twenty chances to fix the same bug.

This may address real differences, but it duplicates the shared mechanisms.

**Build a shared follow-up feature, configurable for each client.**

The need repeats. The code can be reused.

Separate the shared mechanism from the settings: delay, recipients and content.

**With an LLM, the time from a blank page to a sexy demo is unbeatable. The bill will come later, when you need to make it reliable and keep developing it.**

The prototype helps you discover the recurring pattern: wait, check whether a reply has arrived, then follow up if needed. This mechanism can become a product feature. Differences between clients become configuration.

**LLMs take you from 0 to 1. Software engineering takes you from 1 to n.**