The right question isn’t “is AI useful?” but “at what, exactly, and how much”. On that, unlike most of the debate, there are controlled studies with numbers. And the numbers are interesting partly because they show where AI doesn’t help.
This article puts three things together: what the research shows, seven concrete uses ordered by risk, and the rules (legal ones included) about what you can put into these tools.
How much time it really saves: the two studies to know
Professional writing tasks. Shakked Noy and Whitney Zhang of MIT published a controlled experiment in Science in 2023 on professionals doing realistic writing tasks (press releases, short reports, work emails). The group with access to ChatGPT completed the tasks in roughly 40% less time, with quality rated higher by independent reviewers. The more significant detail: the improvement was much larger for those who started from lower writing quality, compressing the gap between strong and weak writers.
Customer support. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the rollout of a generative assistant across more than five thousand customer support agents. The average productivity gain, measured in issues resolved per hour, was around 14%. Here too the distribution matters more than the average: the benefits were concentrated among less experienced workers, while for the most experienced they were close to zero.
From these two pieces of work you get an honest reading that doubles as an operating rule:
Generative AI compresses the distance to average. It gets you quickly to “good enough” at things you can’t do well, and helps you little or not at all at things you’re already strong at.
Which is why it makes sense to point it at the tasks that weigh on you and aren’t your actual craft, not at the ones that are your distinctive competence.
First things first: what you can paste in
This needs settling before the examples, because it’s the most expensive thing to get wrong.
The legal framework. EU Regulation 2024/1689 (the AI Act) introduces, among other things, an AI literacy obligation: organisations deploying AI systems must ensure a sufficient level of competence among the staff who use them. On personal data, the GDPR applies: pasting personal data about clients or colleagues into an external service is processing, and has to be treated as such.
That this isn’t theoretical is a matter of record: the Italian data protection authority, the Garante, imposed a temporary limitation on processing against ChatGPT in March 2023 and later adopted a sanction decision against OpenAI. These are public, consultable precedents.
The practical rule I use, in three tiers:
| Tier | Content | Rule |
|---|---|---|
| Green | Public text, already-published material, generic content | No issue |
| Amber | Internal non-confidential documents, drafts, notes | Only on company-approved tools, anonymised data |
| Red | Personal data, health data, contracts, proprietary code, credentials, client data | Never, unless it’s a company tool with a contract covering it |
Check the settings too: the main providers let you turn off the use of conversations for training, and business plans generally exclude it by default. The terms are described in OpenAI’s documentation and its help centre.
The seven uses, ordered by value-to-risk ratio
1. Summarising long documents (low risk, high value)
The best possible starting point. Minutes of a meeting, a tender, an internal policy, a contract you need to grasp before reading it properly. The risk is contained because an imperfect summary costs you a few minutes; it doesn’t make you sign the wrong thing.
The prompt that works best isn’t “summarise”, but: “summarise in five points; then list separately what the text leaves ambiguous or doesn’t specify”. The second part is what tells you where you need to read the original.
2. A first draft of something you’ll rewrite anyway
The value isn’t the text produced: it’s the removal of the blank page. It matters most for formats you know poorly, such as minutes, a status note or a formal reply, where your block is structural rather than about content.
A caveat from the MIT study: the gain is greatest where you’re weak. If you write well, a generated draft will slow you down, because fixing it costs more than writing it.
3. Finding the tone in a difficult message
Saying no, chasing someone, replying to an angry counterpart. The right use here is to ask for three versions in different registers and choose, not to have the final version written for you. Generated text tends toward a uniform, recognisable register, and in a delicate message recognisability is exactly what you want to avoid.
4. Rehearsing for an interview or a hard meeting
One of the best and most underrated uses: give the context, ask it to play the other party and to ask follow-up questions about your answers. It doesn’t get tired and it doesn’t judge you. It works very well on interview questions and for rehearsing a pay rise request out loud.
5. Getting something explained without the embarrassment
A technical term, an Excel formula that won’t behave, a concept everyone takes for granted in the meeting. The real advantage isn’t pedagogical, it’s psychological: it removes the social cost of asking.
With one important qualification: on verifiable facts (rules, figures, deadlines, references) the output must be checked against the source, always. Fine for understanding a concept; not for establishing what a law says.
6. Generating alternatives when you’re stuck
Ten titles, five possible approaches, three ways to structure a presentation. AI is mediocre at choosing and excellent at producing variety fast. Use the second and keep the first for yourself.
7. Standardising text you rewrite every week
Response templates, recurring descriptions, communications with a fixed structure. The return here compounds: build the template once, reuse it for months. It’s the use case with the best return over time, and also the most boring, which is why almost nobody does it.
The structural limit: why it’s wrong with confidence
“Hallucinations”, plausible false statements produced in the same tone as correct ones, aren’t a temporary defect to be fixed in the next version: they follow from how these systems work. The technical literature treats them as a category in their own right; a standard reference is Survey of Hallucination in Natural Language Generation (ACM Computing Surveys, 2023).
The practical point that matters: there is no visible signal distinguishing a correct answer from an invented one. The confidence of the tone is identical. Which gives a simple rule:
The more verifiable an answer is (a date, a figure, a rule, a citation, a name), the more likely it is that it needs verifying.
The minimum verification method, which costs little:
- Ask for the source, then open the source. If the URL doesn’t exist or doesn’t say what the AI claims, you have your answer.
- On a numerical figure, look for the number directly on the site of the institution that produces it.
- If you can’t verify it, don’t use it. Not “use it with caution”: don’t use it.
The NIST AI Risk Management Framework formalises this approach at organisational level, if you need a reference for proposing rules at your company.
The accountability rule
One sentence worth keeping: the signature stays yours. If you hand in a document with a wrong number, “ChatGPT wrote it” doesn’t cover you, not professionally and, in the relevant cases, not legally. The tool speeds up production; it doesn’t transfer responsibility.
That also draws the line of honest use. Using it to write something faster that you’re able to check: legitimate. Using it to produce something you can’t evaluate: you’re signing a blank cheque.
Where to start
One task, the most repetitive one in your week, for two weeks. Time it before and after; it’s the only way to know whether you’re actually gaining something or just doing the same thing in a more entertaining way.
The factor that determines output quality more than any other is how you frame the request: the method is in how to write effective prompts.
Frequently asked questions
Can my employer forbid it? Yes, and many organisations have specific policies, especially on data and confidentiality. Using an unapproved tool with company data can be a disciplinary breach. Ask first: in most cases there’s already a written rule nobody has read.
Are my conversations used to train the models? It depends on the service and the settings. Business plans generally exclude it by default, and on consumer plans it can usually be turned off. Check the provider’s documentation rather than assuming.
Do I have to declare that I used AI? In academic contexts and in many newsrooms yes, and it’s the rule. At a company it depends on the policy. As a general principle: if the value of what you hand over lies in the fact that you produced it (an assessment, a professional recommendation), declaring it is the right call. We declare ours on the how we write page.
Which paid version is worth it? It depends on the use, and it changes fast. The stable criterion is different: before paying, check you’ve saturated the free features. People who don’t get good results for free rarely get them by paying.
Will AI replace me? The available data so far shows transformation of tasks more than replacement of roles, with stronger effects on standardisable work. The most useful reading is Brynjolfsson’s: less experienced people gain more, which means the competitive advantage shifts to the things AI doesn’t do, namely judgement, accountability and context. That’s the subject of what skills employers actually ask for.
Sources
- S. Noy and W. Zhang, Experimental evidence on the productivity effects of generative artificial intelligence, Science, 2023. Roughly 40% reduction in completion time on professional writing tasks, with the larger benefit for those starting from a lower level.
- E. Brynjolfsson, D. Li and L. Raymond, Generative AI at Work. Study of over 5,000 customer support agents, average productivity gain around 14%, concentrated among less experienced workers.
- EU Regulation 2024/1689 (AI Act). Transparency and AI-literacy obligations.
- EU Regulation 2016/679 (GDPR). Processing of personal data entered into third-party services.
- Garante per la protezione dei dati personali. Decisions concerning ChatGPT (temporary limitation on processing, March 2023, and subsequent sanction decision).
- OpenAI documentation and help centre. Settings on data retention and use of conversations for training.
- NIST AI Risk Management Framework. Reference framework for managing AI system risks.
- Z. Ji et al., Survey of Hallucination in Natural Language Generation, ACM Computing Surveys, 2023. Review of hallucination in generative models.



