AI projects case

Alberto Boschi: an AI assistant for 80 employees, with the two full answers.

Alberto Boschi

Founder, Dedalix · Enterprise Software & AI Executive · Sales, Consulting and Project Management

BEFORE · the non-expert's prompt

“We want to give all the employees of our company (80 people, in administration, sales and production) a generative artificial intelligence assistant. Prepare a project plan for the rollout, with phases, timing, an indicative budget and the KPIs to measure the return on investment.”

What was missingWhich processes and which data the AI should touch, and under what rules. The answer sounds like a cautious consultant, but it would fit any company, and it calculates CHF 59,300 in costs and an ROI of −11% on assumptions: CHF 30 per month, CHF 35 per hour, 20 minutes a day.

AFTER · the professional's prompt

Act as a senior digital transformation consultant, expert in project and change management in SMEs. The client is an SME in Ticino with 80 employees; some already use AI assistants with personal accounts and there is no policy. Prepare a draft adoption plan for a generative AI assistant for all employees, with a short business case.

Start from the problem, not the tool: measurable goals and 2–3 use cases to start with, with the reason.
Before the pilot: data rules (what can be entered, company accounts, usage policy, Swiss Federal Act on Data Protection (FADP)) and a baseline measurement.
A pilot with written go/no-go criteria, an owner for each phase and a risk register. The return is calculated afterwards, on the measured times.

Do not invent data, prices or studies: write [to be verified] and list your assumptions. Format: a 5-line business case, then tables for phases and risks.

Act as a senior digital transformation consultant, expert in project and change management in SMEs. The client is an SME in Ticino with 80 employees; some already use AI assistants with personal accounts and there is no policy. Prepare a draft adoption plan for a generative AI assistant for all employees, with a short business case.

Requirements:
1. Start from the problem, not the tool: measurable goals and 2–3 use cases to start with, with the reason.
2. Before the pilot: data rules (what can be entered, company accounts, usage policy, Swiss Federal Act on Data Protection (FADP)) and a baseline measurement.
3. A pilot with written go/no-go criteria, an owner for each phase and a risk register. The return is calculated afterwards, on the measured times.

Do not invent data, prices or studies: write [to be verified] and list your assumptions. Format: a 5-line business case, then tables for phases and risks.

Read the two full answers ›

The AI projects case, in full

The two full answers

Same model, same question, two different prompts. Here are the complete answers from ChatGPT, Sol 6.1 (High), with no cuts: switch between them and compare.

Prompt 1 · the non-expertGets a number: CHF 59,300 in costs and an ROI of −11%, calculated on assumptions (CHF 30 per month, CHF 35 per hour, 20 minutes a day) for any company at all, because the prompt does not say where we are or how we work.Read the full answer ›

Prompt 2 · the professionalGets a method for deciding: an immediate rule on personal accounts, data classification, governance, a go/no-go with three outcomes, and no invented figures.Read the full answer ›

Answers obtained with ChatGPT, Sol 6.1 (High). Changes from the original: the answers have been translated from Italian, the amounts are shown in CHF with the same values, without conversion, and the sentence about the reference currency has been removed. The figures in answer 1 are the model's assumptions, not real data.

Same prompt, different answer: that is normal

If you copy one of these prompts into your AI, even with the same model, you will not get the same words, the same numbers or the same structure as the two full answers. This is not a mistake, and it happens with ChatGPT, Claude, Gemini, Perplexity, Grok and the others. These are the reasons.

The AI does not copy, it generatesIt writes one word at a time, choosing among the most likely ones with an element of chance. That is why the same question, asked twice in a row, already gives two different texts.

What the AI already knows about youMemory, the history of previous chats and the custom instructions in your account all feed into the answer, even though they do not appear in the prompt. Two people with the same prompt do not start from the same place.

The project and its filesInside a project, the AI also reads the instructions and documents you uploaded. In tools for agents, it reads files such as CLAUDE.md or AGENTS.md. The context changes, so the answer changes.

The model settingsThe same name can mean versions updated over time, different reasoning levels (the answers in this guide use «High»), and different plans and modes.

The tools that are switched onWeb search, code execution and connectors add information that changes from one day to the next and from one account to another.

The conversation so farWhat you wrote earlier in the same chat, the language, and even the way a sentence is worded all shift the result.

Copy the method, not the answer.What stays is the difference between BEFORE and AFTER: with context, requirements and limits, the answer is more precise, easier to check and better suited to your case, even when the words change. The answers in this guide are an example obtained at a specific moment, not a result to reproduce. And every answer, even the best one, must be checked.