AI in firms and companies: the prompting techniques that make the difference.

Over the last two years, generative AI has moved from social-media demos to the desks of accountants, lawyers, administrators and entrepreneurs. Some use it to summarise a circular, some for a first draft, and many — more often than admitted — use it covertly, in free versions, with mixed results. The gap between those who gain a real advantage and those left disappointed is not explained by the model: today's tools are all extraordinarily capable. What changes the outcome is the method.

1. Why the model is not enough

A recurring mistake is treating AI as software to buy: activate the licence, problem solved. In reality, language models are general tools: they can do a great deal, but they do not know your organisation, your clients or the way you work. Without precise instructions they produce textbook answers — correct in the abstract, useless in the concrete case.

Two skills make the difference: knowing how to formulate requests (prompting) and knowing how to make AI work inside your processes, not alongside them. The first can be learned in weeks; the second requires a project.

2. Prompting: the input determines the output

Asking an AI to "explain the two-year preventive tax agreement" produces an encyclopaedia entry. Try instead assigning a role ("you are a tax advisor specialised in SMEs"), providing context (sector, figures, constraints of the case), defining the task (a table of benefits, risks and points to verify), and setting controls (current legislation only, flag uncertain points): the result is a working document that saves an hour, not five minutes. The difference is not politeness or length: it is structure.

3. The five components of a professional prompt

• Role — who the AI should "be": it shapes language, depth and perspective.
• Context — the facts of the case: sector, size, figures, constraints. The richer the context, the more concrete the answer.
• Task — what it must produce and to what end: assess, compare, draft, spot issues.
• Format — table, list, one-page memo: the right format makes the output immediately usable.
• Controls — cite sources, use only current legislation, declare uncertainty instead of improvising.

Then there is a multiplier worth more than any formula: the material you attach. A balance sheet, a contract, the text of the relevant provision transform the conversation: the AI stops drawing only on statistical memory and starts reasoning on your documents. An output anchored to verifiable sources is not just more useful: it is more defensible.

4. Hallucinations: why AI invents, and how to reduce the risk

Hallucinations — plausible, well-written but wrong answers — are not an occasional defect: they are a structural feature of generative models, which produce the statistically most likely text, not necessarily the true one. The risk drops sharply by combining several levers: request step-by-step reasoning; anchor answers to attached sources; require online search for data that changes; instruct the model to declare uncertainty.

⚠️ No technique eliminates error on its own. Human verification of every citation, figure and conclusion remains the step that makes the output usable — and it is what the law now demands: under Italian Law 132/2025, AI may only support the professional service, never replace the person's critical contribution.

5. Beyond chat: agentic systems

Prompting is the first half of the journey. The second, more promising half is agentic systems: AI stops answering single questions and pursues a goal across multiple steps — it plans, consults external tools, checks results, corrects course. Connection standards such as MCP (Model Context Protocol) act as connectors between the model and the organisation's systems: databases, management software, document archive, email. The qualifying point: everything happens in a controlled and traceable way.

What can an agent actually do? Monitor legislation and deadlines, flagging only what concerns your clients; review files in due diligence; pre-fill returns from accounting data; run reconciliations; prepare drafts with automatic citation checking. The agent performs the repetitive steps; the person keeps the judgement and the signature. This is where DIY reaches its limits: connecting an agent to a firm's or company's systems requires engineering, security and professional understanding together — exactly the ground BoezioAI was built on.

6. The rules of the game: the AI Act and Law 132/2025

Using AI in a firm or company today is a regulated activity. The European AI Act (Reg. EU 2024/1689) includes an AI literacy obligation (Art. 4, applicable since 2 February 2025): those who use AI systems must ensure their staff receive adequate training. Italian Law 132/2025 (in force since 10 October 2025), at Art. 13, sets the rules for intellectual professions: AI may be used only as a support tool, the professional's intellectual work must remain prevalent, and the client must be informed in clear language. In practice, minimal governance is needed: which tools are allowed (professional versions offer contractual guarantees on confidentiality that free ones do not), who may use them and for what, how output is verified. Banning everything is the worst choice: "shadow AI" brings the risks with none of the control.

7. Where to start: four steps

• 1. Assessment — map activities and find where repetitive time is concentrated: that is where AI pays off most.
• 2. Tool selection — platforms fit for professional use: confidentiality, access controls, traceability.
• 3. Pilot project — one contained, measurable process, to decide with numbers rather than impressions.
• 4. Training and oversight — train people on real documents and maintain rules and checks over time.

The same techniques that make a firm more efficient — structured prompts, agents on repetitive processes, governance rules — apply to the administration, management control and document flows of an SME. The critical factor is that strategy and delivery speak the same language: a roadmap is worth little if those who write it cannot build what it promises.

Frequently asked questions

Will AI replace professionals and staff?

No: it automates the repetitive steps, not the judgement. Italian Law 132/2025 requires the professional's intellectual work to remain prevalent. The job changes, it does not disappear: value shifts from executive to interpretative activities.

Do I have to tell clients I use AI?

Yes: Art. 13 of Law 132/2025 requires informing them in clear, simple and comprehensive language. The tidiest way is to address it in the engagement letter.

Can I upload client documents to a free AI?

A bad idea: consumer versions do not offer the contractual guarantees on confidentiality and data that professional secrecy and the GDPR require. Use professional plans — and, in any case, prudence and data minimisation.

Where do I start if my organisation is small?

With the method, not the spending: one professional platform, the five-component prompt checklist, one measurable pilot process. To move from individual use to process integration, work with people who design and build these systems every day.

Article by Dario Carta (Lawyer and Chartered Accountant), adapted from the full version published in Italian on strategiefiscali.it. Reproduced and translated with the author's consent. References: Reg. (EU) 2024/1689 (AI Act), Art. 4; Italian Law no. 132 of 23 September 2025, Art. 13.

← All news