AI Skills: What You Can Learn Yourself, What Is Better Handed Off
By Anis Hammouche·July 6, 2026
To use AI in your business, you do not need an expert. Three skills can be learned on your own: framing a task, measuring what the tool earns you, and rereading what it produces. It is better to hand off anything that touches your clients' data, tools wired together, or work that has to last.
Take Camille, a freelance consultant (an invented case, like the other examples). She wants to automate her meeting minutes and her invoicing, and she hesitates: pay an expert, or get into it herself. A three-person web agency (invented too) asks the question differently: "who here takes care of AI?" In both cases the right answer is not a job title. It is knowing, task by task, what you can settle with a few tries and what needs someone who does this for a living.
What you can learn yourself
These three skills need no code and no rare profile. They need you to know your trade, and you do.
Framing a task
Framing a task means describing it precisely enough that you can tell whether it is done. Pick a tedious, repetitive task, describe it with its trigger and its result, decide what the tool should and should not do. Camille can write: "at the end of each meeting, I turn my notes into a one-page report". The habit of briefing the tool like an intern is covered in AI Is the Intern You Cannot Afford to Hire.
Measuring a gain
Measuring a gain means comparing the time before and the time after, review included, on the same scope. Say Camille takes 30 minutes to write a report by hand and 10 minutes with the tool, review included (invented figures): the gain is 20 minutes per meeting. One week is enough for a first idea. The method is in AI ROI: How to Measure What a Tool Really Earns You.
Rereading what the tool produces
AI can produce a result that is wrong but plausible. You have what it takes to spot it: you know your clients and your numbers. Read AI Invents Figures, Sources and Terms for the most common traps.
These three skills rest on your knowledge of your trade, which an outside expert would first have to acquire. They do not replace an expert on every subject, as what follows shows.
What is better handed off
There are cases where learning alone costs more than paying someone. Three criteria help you spot them: the data, how autonomous the tool is, and how long it has to last.
Your clients' data is involved
The CNIL, the French data protection authority, says that when a use involves supplying personal data (of clients or staff) or sensitive documentation, it generally seems more appropriate and safer to favour an on-premises deployment, which limits the risk of data being extracted by a third party. The CNIL talks about hosting, not about hiring: the conclusion I draw is that the choice of hosting, access and contracts is not something you improvise. For the simple case of a consumer tool, start with What You Paste into Free ChatGPT or Claude.
The tool acts without you
An automation that reads your inbox, writes into your accounting or sends messages to your clients does not just suggest a text. A mistake goes out on its own. As long as a human rereads before sending, you can handle it. When the tool acts alone on sensitive matters, it is safer to bring in someone who does this for a living. For simple cases, automating without code with n8n, Make or Zapier stays within your reach.
It has to last
A tool built on one specific model breaks the day that model is retired: Anthropic says on its deprecations page that it regularly retires older models, with at least 60 days' notice for a publicly released one. I go through the risk in Your AI Tool Can Be Cut Off Overnight. If the result serves a client, someone has to maintain it. That is development, not tinkering.
These three cases are a developer's work: a complex automation, a custom AI agent, a development project. For the rest, you can start on your own, and that is a good idea.
On your own or handed off: how to decide
| Task (invented examples) | On your own | Hand off | Why |
|---|---|---|---|
| Draft reply to a client, reread before sending | Yes | No | You read every message |
| Meeting minutes from your notes | Yes | No | The notes are yours and the review is quick |
| Invoice reminder sent automatically | With a no-code tool, if you review the first sends | Yes, if it writes into your accounting | A mistake goes to the client |
| Agent that answers clients with access to their file | No | Yes | Personal data and replies nobody rereads |
| Internal tool delivered to an agency client | No | Yes | Maintenance and liability |
The "On your own" column can be tested in a week, and the "Hand off" column is prepared with a good needs sheet. Either way, you gain by knowing exactly what to ask for.
A sheet to decide
Copy this sheet and fill it in for the task you would like to delegate to AI:
Task in one sentence: [what the tool must do]
Trigger: [what starts the task]
Expected result: [what I must get]
Data involved: [my notes / client data / payment data]
Does the tool act without me rereading? : [yes / no]
How long it takes me today: [minutes per week]
Who maintains it if the tool changes? : [me / nobody / someone to find]
Verdict: [I test on my own / I hand off]
Frequently asked questions
Do I need an AI expert to get started? No. The first uses call for framing, measuring and rereading, three things you learn by trying. An expert becomes useful on uses that touch sensitive data or have to last.
I have neither the time nor the technical skills. What do I do? Start with one task, with a consumer tool, no client data. A week of tests already tells you whether the idea is worth it. If it is, but goes beyond what you can do alone, you will be able to explain your need precisely.
What if my need is technical? Then handing it off is the right decision, and you gain by arriving with your sheet filled in. You know which task, which data, which result. The provider's work is shorter and the quote is fairer.
Can I paste my clients' data into a consumer tool? Avoid it until you have read what the provider does with what you send. The CNIL recommends favouring an on-premises deployment when a use involves clients' personal data. The details are in What You Paste into Free ChatGPT or Claude.
Sources
- CNIL, Les questions-réponses de la CNIL sur l'utilisation d'un système d'IA générative (in French, read on November 1, 2026)
- Anthropic, Model deprecations (at least 60 days' notice before a publicly released model is retired, read on November 1, 2026)
The examples of Camille, the web agency and the table are fictional. This article is a decision guide: it rests on no personal measurement.
The next step is From Solo to Small Team: What AI Replaces, and What It Will Not, where the same question comes up on the hiring side. I share this kind of landmark in my newsletter, you can sign up right below.
