Short answer. Your employees already use an assistant: the question is on which account, and with what data. Five rules. A company account, not a personal one. Three bins: what is pasted as it is, what is pasted once cleaned, what is never pasted, passwords, data about people, documents under confidentiality. A reread by whoever signs, because the assistant can state something false. No action without human approval. Two hours of training per team. Five uses can be done from tomorrow without risk; the three most useful, those that touch your customers and your documents, need this frame, then an assistant connected to your data.
This guide is for the manager who discovers that their teams use ChatGPT or another assistant, and for the founder based in France who wants the same time saving in their company in Algeria. The definitions quoted come from public sources: the chatbot, the hallucination, prompt injection, retrieval-augmented generation, “shadow AI”. The data-use conditions are those the publisher of ChatGPT states in November 2026; they change, and each assistant has its own. The company and its uses are a constructed example. This guide does not cover the legal rules on personal data or professional secrecy: see your legal adviser.
What can be pasted into an assistant: three bins
The figure sorts twelve examples into three bins.
Why three bins. The undeclared use of AI tools by employees has a name, “shadow AI”: a recent form of those IT tools adopted without going through the IT department. It is not theoretical: in May 2023, according to press reports, Samsung restricted the use of assistants after engineers had sent internal code to ChatGPT. The answer is not a ten-page policy: it is a rule applied in ten seconds, at the moment of pasting.
The middle bin. It is the most useful and the most delicate: minutes, a sales table, a procedure. Two ways in. Clean it: names become Customer A, amounts become X, and the result is pasted back into the real document. Or work on the company account, whose conditions are known. The third bin. Passwords, access keys, data about identifiable people, documents covered by a confidentiality clause: never, whatever the tool and the plan. A password pasted into a conversation is a password to change, as after a hack.
Setting a frame, in this order
The first step is accented: take stock.
Take stock. An anonymous questionnaire of five questions: do you use an assistant, which one, on which account, for which tasks, have you ever pasted a company document into it? The answers often surprise management. Do not ban. A ban does not remove the use: it moves it to the personal phone, where no rule applies any more.
Choose, write, train. One plan, one account per person, the rule on one page, two hours per team on the real tasks found in the stock-take. Equip. Ten standard prompts per job, reread once by someone who knows the work: following up a quote, the minutes of a meeting, the reply to a complaint. They are kept where everyone can find them, for example in the intranet.
«A ban does not remove the use: it moves it to the personal phone.»
Personal account or company account
The figure compares the two on six points.
What the publisher says. ChatGPT is a generative AI chatbot developed by OpenAI. According to what the publisher states on its public pages in November 2026, sharing conversations to improve the models is on by default on personal accounts, free or paid, and can be turned off in the settings; for the plans aimed at companies and for programmatic access, content is not used for training by default. These conditions change, and each assistant has its own: reread those of the plan you take out.
What the company account changes. The history stays in the company's space and is closed when the employee leaves; an administrator adds and removes members; sign-in is with the work address. What it does not change. The assistant does not become more accurate, and the three bins still apply: data about a person or a document under confidentiality is no more to be pasted there. If nothing must leave your walls, there is another route, described in the guide to running a model on your own servers.
Five rules that fit on one page
The accented layer is data.
Reread. A hallucination is a response generated by AI that contains false or misleading information presented as fact. The assistant invents a reference, a figure or a clause with the same assurance as everything else. Hence the rule: whoever signs rereads, in particular figures, names, dates and quotations. A text nobody has reread is not sent.
No action without approval. As soon as an assistant can read your emails or web pages, another weakness appears: prompt injection, an attack in which innocuous-looking inputs are designed to cause unintended behavior in the model. A booby-trapped document can “ask” it to pass data on. The defence is a simple rule: it proposes, someone decides. It is also the rule for the agents described in the guide to the AI agent in a company. Train. Two hours are enough for use to become a skill.
«It proposes, someone decides.»
Ten uses, sorted by time saved and by sensitivity
The figure places ten uses by time saved and by data sensitivity.
From tomorrow. Rephrasing and correcting an email, translating between French, Arabic and English, summarising a long public document, writing a spreadsheet formula, building the outline of a proposal: no internal data goes in, and the gain is immediate. That is where a team learns to give an instruction and to reread.
With a frame. Answering customers from your documents, searching your contracts and procedures, analysing sales per customer: these are the three uses that would save the most, and they touch your data. The technique is called retrieval-augmented generation: it enables a model to retrieve and incorporate new information from your own data sources. The assistant then answers from your documents, with rights per person. It is something to be built, described in the guide to putting AI into an application, and it is checked before being opened to everyone: see the guide to measuring that an AI works. No. Applying legal or medical advice as it is, or entrusting it with passwords.
Four weeks, from taking stock to the first review, in 2027
Four weeks in the example, the first one to take stock.
- What the manager receives: the result of the stock-take, the three-bin rule on one page, the company accounts opened, ten standard prompts per job, and a review after four weeks.
- What the company no longer does: find out by chance that a contract was pasted into a personal account, ban to no effect, send a text nobody has reread, let an assistant act on its own.
- What is measured: share of employees on a company account, tasks given to the assistant, incidents reported, time people say they saved per job.
- What Symloop does: the stock-take, the rule and the training, opening the accounts, the standard prompts, then, for the uses that deserve it, the assistant connected to your documents, with its rights, its measurement and its human approval.
Week 1, taking stock, anonymously and without judgment. Week 2, the plan and the accounts. Weeks 2 and 3, the rule and the training: two hours per team, on the real tasks.
Weeks 3 and 4, the standard prompts. Week 4, the first review: real uses, incidents, the time each person says they saved, and which of the three sensitive uses now deserve an assistant connected to your documents.
- How do you use ChatGPT in a company without risk?
- With five rules: a company account, the three-bin rule for what is pasted, a reread by whoever signs, no action without human approval, and two hours of training per team. Start with the uses where no internal data goes in.
- Can company data be put into ChatGPT?
- It depends on the bin. An already public text, yes. An internal document, once cleaned of its names and amounts, or on a company account. Passwords, data about people or a document under a confidentiality clause: never, whatever the tool and the plan.
- Does ChatGPT keep what you write to it?
- Conversations stay in the account's history. According to the publisher, in November 2026, they may be used to improve the models on a personal account unless that sharing is turned off in the settings, and are not used for that by default on the plans for companies. Check the conditions of your plan.
- Do we need a company subscription?
- As soon as use goes beyond rephrasing and translating texts with no internal data, yes: the history, the administration and the terms of use move to the company's side, and the account is closed when the employee leaves. It does not remove the need for the three bins or the reread.
- How does a ChatGPT plan for companies work?
- It is a workspace in the company's name: each employee has an account linked to their work address, an administrator manages members and rights, and the data conditions are those of a business contract. The exact features change: read the plan's page when you take it out.
- Should we ban ChatGPT for our employees?
- No. A ban does not remove the use: it moves it to the personal phone, where no rule applies and where you can no longer see anything. A simple frame, applied in ten seconds, protects better than a ban nobody respects.
- What should be done if an employee has already pasted a confidential document?
- Delete the conversation, turn off sharing in the account settings, change any password or key that was in it, and check with your legal adviser what the document's confidentiality clause provides. Then make it the first case in the training, without looking for someone to blame.
- Can ChatGPT's answers be trusted?
- Not without rereading. An assistant can produce false information presented as fact: a reference, a figure, a clause. It is reliable for rephrasing, translating, structuring; for a fact, an amount or a text that commits you, whoever signs checks.
- What is the difference between ChatGPT and an assistant connected to our documents?
- ChatGPT answers from what it has learnt and from what you paste. An assistant connected to your documents looks for the answer in your contracts, procedures and catalogue, cites its source, and respects each person's rights. It is something to be built, not a subscription.
- Where do we start?
- With an anonymous stock-take: who uses which assistant, on which account, for what. Then the five risk-free uses, on a company account, with the three-bin rule. The uses that touch your customers come afterwards.