Why you need a policy
Employees already use AI tools, with or without a document. Without rules, each person decides what can go into an open chat, and a contract with a counterparty’s prices ends up on someone else’s server.
A policy turns this from a personal decision into a company rule. It gives three things: a list of allowed tools, a limit on information, and an owner for each question.
The document is also needed for the law. Under 152-FZ, Russia’s personal data law, the company as the data operator answers for customers’ personal data even if an employee sent it to an outside service. The policy records which rules the company set and how it trained its people.
Write it short: an employee should read the document in ten minutes and remember the main points. The structure has six sections:
1. Purpose and scope: whom the document covers and which tasks it includes.
2. Allowed services and how new ones are added.
3. Information that must not be sent to AI, and the anonymization rule.
4. Checking the result: what can be used at once and what a person checks.
5. Responsibility: who approves, who controls, and whom to tell about a breach.
6. Training and review: how new employees learn the rules and when the document is updated.
Allowed services
This section lists the services an employee may use at work and the tasks they are allowed for. For each one, state where requests are processed, whether there is a contract with the vendor, and which information may be sent there.
It is handy to split services into three groups:
1. Allowed for work tasks: corporate tools with a data processing agreement, or models on servers in Russia.
2. Allowed only for general tasks without information on customers and money: open services for drafts, translations, and search.
3. Banned: services without a contract, personal accounts for work tasks, and browser extensions that read CRM pages.
Spell out separately how to add a new service: who checks the vendor’s terms, the place of processing, and the contract, and how soon the employee gets an answer.
If employees work with customers’ personal data, an allowed tool must process it in Russia. In our deployments, customers’ personal data is processed on servers in Russia under 152-FZ, and the data and access stay with the client.
Which data is off-limits
This section sets the limit with examples, not general words. The phrase “confidential information” tells an employee nothing, while the list “passport, customer phone number, price from a contract” is clear at once.
What goes on the banned list:
1. Personal data of customers and employees.
2. Contract terms, prices for individual customers, margins, and sales plans.
3. Financial reports before publication, salaries, bank statements.
4. Passwords, access keys, addresses of internal systems.
5. Source code and documentation of internal systems.
Next to it, put the anonymization rule: which fields an employee removes or replaces before sending, if the task cannot be solved another way. A separate page covers in detail what employees send to AI tools and what it means for the company.
Checking the result
AI makes plausible mistakes: confident text with an invented number looks the same as a correct one. So this section says which result can be used at once and which one a person checks.
It is handy to split tasks by the cost of a mistake:
1. Drafts for yourself: no check is needed, since the employee edits the text anyway.
2. Texts for customers and partners: before sending, the author reads the text and checks numbers and promises against the source.
3. Contracts, calculations, legal and financial conclusions: a specialist in the field checks them, and AI gives only a draft.
For AI agents that answer customers by themselves, the policy states where their rules are kept, who approves them, and what the agent passes to a person.
On a separate line: a person signs off on the AI result. If an email went to a customer, the employee is responsible for it, not the AI tool.
Who is responsible
This section names roles and specific people, not whole departments. Nobody carries out a document without names.
1. The policy owner approves the document and settles disputed cases. Often this is the head of AI or the CEO.
2. Legal and security check services, vendor contracts, and compliance with 152-FZ.
3. The IT department connects allowed services and grants and revokes access.
4. Department heads watch the rules in their processes and collect employees’ questions.
5. Each employee is responsible for what they sent to AI and what they gave to a customer.
Spell out whom to tell if information has already gone to a banned service, and how fast. A fast report matters more than punishment: then the leak can still be stopped.
How to assign these roles in an AI department is described on the page about structure and roles.
How to roll it out without banning everything
A full ban does not stop employees: they keep using AI on their personal phones, and the company loses control. The order is the reverse: first give an allowed tool, then limit the rest.
1. Collect a list of the services employees already use and the tasks they use them for.
2. Choose allowed tools for these tasks and connect them before any ban.
3. Write a short policy along the six sections above and give it to a lawyer to adapt.
4. Run training on examples from the departments’ work: what can be sent, what cannot, and how to anonymize text.
5. After a month, collect employees’ questions and fix the document.
Subscriptions for employees give no effect by themselves if processes stay the same. Why this happens is covered on the page about subscriptions instead of processes.
In a 30-minute review, we will look at which services your employees already use and tell you which section of the document to start with.