Teddy Barbu
Sheet 1 of 3 · The map
There is no best AI tool, only a best fit for one job. Find the row that matches the work you keep repeating, then use the tool in that row before you buy anything else.
| The job | Reach for | Because |
|---|---|---|
| Drafting and rewritingEmails, posts, product copy, first drafts | ChatGPTor Claude | Both draft well. Claude holds a long document together; ChatGPT is quicker for short, varied pieces. |
| Reading long documentsContracts, reports, transcripts, research | Claudeor NotebookLM | Long-context models summarise without losing the middle. NotebookLM answers only from the files you give it. |
| Finding current factsMarket checks, competitor scans, prices | Perplexityor Gemini | They search the live web and cite sources, so you can check the claim. A plain chatbot cannot see last week. |
| Moving data between appsForm to CRM, order to invoice, alerts | Zapieror Make | Most "AI problems" are plumbing. Connect the tools first; add a model only where judgement is needed. |
| Answering the same questionsSupport, FAQs, booking, intake | A scoped assistanton your own content | Answers drawn from your documents, with a clear handover to a person when it is unsure. |
| Notes and follow-upsMeetings, calls, actions | Your meeting tool'sbuilt-in summary | A separate note-taker is one more app to run. Use what is already in the call before adding another. |
| Images and layoutSocial, simple graphics, mock-ups | Canva AIor a design pass | Fine for volume and speed. Anything that carries your identity still wants a human eye on it. |
| Code and technical workScripts, fixes, small tools | Claude Codeor Copilot | Strongest when the task is specific and testable. Weakest when nobody can tell whether the output is right. |
Two rules that save money. Pick the simplest tool that fits the systems you already run — a pilot that adds steps gets abandoned. And never automate a task nobody has written down; you will automate the confusion too.
Before anything sensitive: check where the data goes, who at the vendor can see it, and whether it trains their model. Get that in writing.
Teddy Barbu
Sheet 2 of 3 · The plan
Fill this in before you buy a subscription. If you cannot answer 2 and 4, you are not ready to start — and that is a useful thing to find out for free.
One repeated task, named the way the person who does it would name it.
Your baseline. Times per week × minutes each, and who does it. Guess if you must, but write the guess down.
One name. It should be the person doing the task, not only the person paying for the tool.
A number and a date. "Two hours a week back by the 30th" works. "Use AI more" does not.
Name the step that always gets read by a human, and what happens when the tool is unsure.
Wrong answer to a customer, a number out by a decimal, something private in the wrong place.
Come back to this row and tick one. Stopping is a real answer, and a cheap one.
Teddy Barbu
Sheet 3 of 3 · The loop
Five steps and one decision. The arrow back is the part most people skip, and it is the part that stops a trial quietly becoming a subscription nobody uses.
If the first version feels embarrassingly limited, it is the right size. You are testing whether the task suits a machine, not building a product.
The baseline in step 02 is the only thing that can tell you later whether this worked. Written down beats remembered, every time.
A trial you end after thirty days cost you thirty days. A tool nobody admits is useless costs you every month, quietly.
Make it work. Then make it yours.