Apollo's chief economist looked at the 493 companies in the S&P 500 that aren't big tech: revenue per worker flat, profit margins flat, no sign of AI in productivity yet. My take from building AI every day: it pays off when you rebuild one boring workflow around it, not when you add a chatbot on top. This kit walks you through finding that one workflow and building the first version.
#1. Why "one workflow" and not "more AI"
Adding AI on top means the same steps as before, plus a chat window. Rebuilding a workflow means asking: if a fast, cheap assistant did the first draft of every step, which steps would still need me?
Good first workflows have four things in common:
- They repeat. Every day or every week, not once a year.
- They follow rules. You could explain them to a new colleague in ten minutes.
- The output can be checked. A wrong result is easy to spot before it causes damage.
- They are boring. Nobody will miss doing them by hand.
#2. Find it: the 15-minute task audit
Write down everything you did last week that took more than 10 minutes. Then paste the list into Claude or ChatGPT with this prompt:
Here is a list of tasks I did last week, with roughly how long each took and how often it repeats. For every task, score it from 1 to 5 on: how often it repeats, how much time it takes, how clear the rules are, and how easy it is to check the result. Put everything in one table, sorted by total score. Then pick the one task I should hand to AI first and explain why in three sentences. Do not pick anything where a mistake would be expensive and hard to notice.Score it yourself if you prefer:
| Question | 1 point | 5 points |
|---|---|---|
| How often? | a few times a year | every day |
| How long? | under 5 minutes | over an hour |
| Clear rules? | depends on gut feeling | could be a checklist |
| Easy to check? | errors show up months later | errors are obvious at a glance |
Start with the highest total. Skip anything with a 1 in the last row.
#3. Map it before you automate it
Most AI projects fail because nobody wrote down what the workflow actually is. Do that first:
I want to hand this workflow to AI: [describe it in a few sentences]. Ask me questions one at a time until you understand every step: what comes in, what I do with it, which decisions I make, what goes out and who gets it. Then write the workflow as numbered steps and mark each step as AI (draft it), ME (decide or check) or TOOL (a script or app does it).The goal is a list where most steps say AI or TOOL and only the decisions say ME.
#4. Three boring workflows that pay off
These are the three from my Reel, with a prompt to start each one.
Checking documents. Contracts, invoices, applications, forms: anything you read against a checklist.
Check this document against my checklist below. For every point, answer PASS, FAIL or UNCLEAR and quote the exact sentence you based it on. If something on the checklist is missing from the document, say so. Do not summarize the document, only check it.
Checklist:
[paste your checklist]Sorting emails. Turn a full inbox into a short list of what needs you.
Here are my unread emails. Sort each one into: REPLY TODAY, REPLY THIS WEEK, READ LATER, IGNORE. For REPLY TODAY, write a short draft answer in my tone (direct, friendly, no filler). Give me the result as a table with sender, subject, category and the draft. Never invent facts, appointments or promises in a draft; write [ASK ME] where you need information.The first version of code, with tests. The AI writes the boring first draft, and tests prove it works before you trust it.
Write a function that does the following: [describe it]. First write the tests: normal cases, edge cases and one case that should fail. Then write the code until all tests pass. Show me the test results. If a requirement is unclear, ask before you write code.#5. Build it in three levels
Start at level 1. Only move up when the level below works every time.
- Saved prompt. Keep the prompt in a note and paste your input each time. Free, works today.
- Project with instructions. Put the prompt, your checklist and two good examples into a Claude Project or a custom GPT. Now you only paste the input.
- Automation. A script or an automation tool runs the workflow on its own (for example every new email or every new file in a folder) and sends you only what needs a decision.
#6. Prove it pays off
Before you start, time the workflow by hand three times. After a week with AI, time it again, including the minutes you spend checking.
| By hand | With AI | |
|---|---|---|
| Minutes per run | ||
| Runs per week | ||
| Mistakes found |
If the time with checking isn't clearly lower, the workflow is the wrong one or the checklist is too vague. Pick the next one on your list.
#7. Where not to start
- Anything with legal, tax or medical consequences that you can't check yourself.
- Decisions about people (hiring, grading, firing).
- Workflows where data must not leave your company. Ask first, or use a local model (see sangeeths.ai/local).
#Sources
- Apollo Academy, The Daily Spark (Torsten Slok), 24 Oct 2026: the S&P 493 numbers quoted in the Reel.
Made by @sangeeths.ai: real AI engineering from Switzerland. Share this kit: sangeeths.ai/workflow · All free kits: sangeeths.ai