There's a narrative genre that in July repeats in every Italian company: the strategic meeting where it's decided "in September we'll do something serious on AI". A manager is nominated, consultants are mentioned, a budget is estimated, it gets postponed.
September arrives, priorities change, "something serious on AI" becomes "something serious on AI in Q1 2027". And so on.
This piece is to sabotage this cycle. Five real experiments, launchable in 5-7 days of effective work. No consultants needed. No enterprise budget needed. What's needed: one person from the company with time, an account on a commercial LLM (ChatGPT Team, Claude Pro Team, or Gemini Workspace), and the discipline to close the pilot within the week — not open it to close it at Christmas.
Three rules of engagement.
Rule 1: one experiment = one concrete problem. No "let's see what AI is good for". Yes "let's see if AI solves this specific thing". Rule 2: declared effort budget. 5-7 days of one person. If more is needed, the pilot failed (and that's fine — you know that road doesn't work). Rule 3: binary success criterion. Yes or no. No "interesting but worth investigating".
The five experiments, in order of increasing ease.
Experiment 1 — Analyzing 100 customer emails
The problem: in your commercial inbox request emails arrive from potential customers. Every email has signals inside — sector, company size, urgency, specific problem — but nobody classifies them. They get read one by one.
What you do in a week: take the 100 most recent emails, anonymize personal data, and feed them to the model in 5 batches of 20 with your classification template (sector, size, urgency, request type). The model returns an Excel to you.
What you discover: probably, two or three dominant segments you hadn't noticed. A concentration of requests in a specific sector, or a recurring problem you didn't know was so widespread. It's the raw material to reposition September's communication.
Success if: at least two actionable insights emerge that you didn't have before.
Experiment 2 — Site content audit
The problem: the site has 50-200 pages. Some are excellent, others outdated, some overlap. Nobody has ever done a real audit because it would take a senior person's week.
What you do in a week: download page content (even just from a sitemap), give the AI precise instructions — "evaluate each page for: clarity of message, freshness of data, overlap with other pages, alignment with the attached brand voice" — and receive a report.
What you discover: usually 10-15 pages to revise immediately, 3-5 to eliminate, 2-3 obvious overlaps generating SEO confusion.
Success if: the report identifies at least 5 concrete actions to take before September.
Experiment 3 — Dynamic FAQ for help desk
The problem: you receive repetitive questions in pre-sales or support. When an email arrives, the answer is practically always the same. Your customer care team spends half their time copy-pasting.
What you do in a week: take the last 200 customer support conversations, have the model categorize recurring questions, generate a dynamic FAQ of 20-30 entries. Then configure an AI assistant on the site that answers these 30 questions, with escalation to human if the question doesn't fit.
What you discover: between 60% and 80% of requests fit into these 30 questions. The chatbot doesn't replace customer care, but frees time for real requests.
Success if: after a week of testing on the site, at least 50% of users who used the chatbot didn't need to talk to a person.
Warning: before publishing the chatbot, read our piece on AI Act phase 2. There's a transparency obligation.
Experiment 4 — Automatic briefing for calls
The problem: before a call with a prospect, you should do 30 minutes of research on them: site, latest LinkedIn posts, sector news. Nobody ever does it.
What you do in a week: create a prompt that, given the customer's name and site URL, returns: what they do, positioning, latest public announcements, typical problems of their sector, a draft of 3 smart questions to ask. Keep it in a favorite. Before every call, launch it.
What you discover: the quality of commercial calls rises 30% without investing an hour more per call.
Success if: after a week, people using it say it's "uncomfortable not having it".
Experiment 5 — Generating variants for social ads
The problem: the creative team produces 3-4 variants of a social campaign. They're few. Meta's bandit algorithms need 8-15 variants to learn well.
What you do in a week: take the variant that worked best in recent campaigns, use it as a seed to generate 10 variants with the AI model — text + image description (images are made by generative tools, but even as a brief for the internal designer it works).
What you discover: 2-3 of the new variants perform better than the original. It's not automatic, it's statistical. More test volumes = more probability of finding a winner.
Success if: the average CPA of the campaign drops at least 10% after two weeks.
The principle behind all five
Did you notice the common structure?
- Specific problem, not abstract
- Material you already have (emails, pages, conversations, briefings, ads)
- Weekly effort, not quarterly
- Binary success
This is the AI pilot discipline for 2026. Companies doing AI seriously don't spend two quarters "evaluating the vendor market". They do five experiments in five weeks. Three fail, two work. On the two that work, they scale.
Everything else is theater.
What do you really need to start? An hour of a decision-maker — you — to choose which of the five to begin with. And a person willing to do the 5 days of effective work. Total cost of the operation, on the books: a week. In licenses: 30-50€ (a team account on an LLM). Risk: zero, or close to it.
Now you have no more excuses. Now, maybe, you also have time — it's July, the company runs at half speed, it's the perfect moment.
Thursday, July 9 we return to the glossary: #HardWords: RAG (Retrieval-Augmented Generation). Why all AI vendors are proposing "a RAG on your documentation" — and when you really need one.
Want one of the five experiments done well, four hands? Under artificial intelligence for business we do AI pilots with fixed time, fixed price, fixed output. Let's talk.