AI usage policy: what to define before opening AI to teams

An AI usage policy that works fits in a few pages and answers the questions your team is already settling on its own: which tool they can use, which data they can paste into a chat, who checks the text before it reaches the customer and who answers when the agent gets it wrong. People in your company already use AI, with or without permission. Banning it pushes that use to personal accounts on their phones, where the company loses sight of what happens, and allowing it with no rules puts proposals, contracts and customers’ personal data into tools nobody has assessed. A useful policy lists the approved tools, classifies data into four levels, requires human review before anything goes out, keeps a registry of agents and has an owner who reviews it every quarter. You can roll it out in 30 days, with legal reviewing the text and HR building the training into the routine.

At Iasbeck&Co we implement CRM and AI agents in B2B operations, and the question about usage rules comes up early in almost every project. I wrote this for HR, legal and IT, who usually get the job of writing the policy, and for sales, marketing and post-sale leaders, who will live with it. This is management guidance and does not replace a legal opinion: the final version of your policy needs to go through your company’s legal team.

Why doesn’t banning ChatGPT at work solve the problem?

In a Gartner survey released in November 2025, carried out with 302 cybersecurity leaders, 69% of organizations said they suspect or have evidence that employees use public generative AI tools the company prohibits. In the same release, Gartner predicts that by 2030 more than 40% of enterprises will have security or compliance incidents linked to this unauthorized use, known as shadow AI, and recommends that CIOs define clear enterprise-wide policies for AI tool usage. The 2024 Work Trend Index from Microsoft and LinkedIn, with 31,000 people in 31 countries, already showed that 78% of those using AI at work brought their own tools.

In revenue operations, the salesperson pastes the whole proposal, with price and discount, into a chat on a personal account to make the text more persuasive, and the Customer Success analyst exports the month’s tickets, with each customer’s name and phone number, to ask for a summary. Both wanted to deliver faster and have no idea whether the tool stores the content or whether the vendor uses it to train the model. When leadership bans it, that use carries on from the home computer and the company loses the chance to guide it. When leadership allows it without rules, the exposure stays the same and gets the company’s blessing. The policy brings that use into leadership’s view, with tools someone has assessed and limits the team knows.

What does an AI usage policy need to define?

I organize the policy around eight decisions, and each one needs a named owner. A policy made only of prohibitions turns into a document people sign when they join and forget, and one made only of principles leaves the salesperson unsure whether they can paste the proposal into the chat.

Approved tools and the path to request a new one

The policy lists the tools the company has assessed and contracted, with login through the corporate email and with the setting that stops the content from being used to train the model, when the vendor offers that option. The path to request a new tool weighs as much as the list: the marketing analyst who finds a useful tool and has nobody to ask goes back to a personal card. A simple form for IT, with a response deadline and known criteria, such as where the data is stored and what the contract says about use of the content, reduces shadow AI more than blocking websites.

In the workshop I ran in October, the market update we discussed most was the option to bring into the CRM the agents people had built in standalone chat tools, with a single bill and a panel where the manager sees all of them. The agent starts working in the system that already holds customer data and has per-user permissions, instead of running on an account only its creator knows about.

Which data can go into AI: four-level classification

The classification is the section the team checks most, so it needs to fit in a table, with examples from the operation and the type of tool each level can go into.

Level	Examples in the operation	Public tool, no contract	Approved corporate tool	Agent inside the CRM
Public	Website, published posts, gated content, company presentation	Allowed	Allowed	Allowed
Internal	Playbook, processes, training scripts, minutes without customer data	Not allowed	Allowed	Allowed
Confidential	Proposal with pricing, margin, contract, pipeline, targets, strategy	Not allowed	Allowed, with authorization from the information owner	Allowed, within the user’s permissions
Customer personal data	Name, email, phone, tax ID, support conversation, call recording and transcript	Not allowed	Only for the purposes the policy lists and legal has validated	Allowed, for the defined purposes and permissions

When in doubt about the level, people treat the data as the level above, and sensitive personal data, such as health information or biometrics, stays out of any AI tool until legal writes a specific rule. Replacing the customer’s name with “Customer A” solves part of the cases involving running text and does not solve an exported spreadsheet with forty columns, where job title, city and company size identify the person.

Human review before anything goes to the customer

Everything AI produces that leaves the company first goes through a person who has read the whole content, whether email, proposal, ticket reply or report. The policy says what that person checks, because reviewing without criteria turns into skimming. I ask the review to cover numbers and dates, names of people and companies, any promise of deadline, price or scope and any statement about the customer that is not recorded in the CRM.

Low-risk internal uses, such as summarizing a call for the salesperson who ran it, can run without item-by-item approval, with the manager checking a sample. In AI in revenue operations I detail the three lists I use with leaders (what the agent does alone, what it does and waits for approval, and what it never does), and the policy is the place to record them, area by area.

Who answers for what they approved

“The AI wrote it” does not work as an explanation for a customer, a board or an auditor, and the policy needs to say so in writing. Whoever sends or approves content answers for it as if they had written it. Whoever configures an agent answers for the instructions and permissions they gave it, and the area’s leadership answers for what agents do without item-by-item approval, because they decided to allow it.

Registry of agents and automations

Every agent and every AI automation that reads or writes company data goes into a registry with name, owner, goal, the tool it runs on, the data it reads, the actions it takes, who approves and credit consumption. An agent with no registered owner gets switched off. The registry looks like red tape until the day a customer complains about a strange message and nobody knows which automation sent it. The architecture and limits of each agent are in AI agents, and the registry is the part of that design that goes into the policy.

Transparency with customers where it applies

The policy defines when the company tells the customer that AI is part of the conversation. The most common cases are the virtual assistant in customer service, which should introduce itself as an assistant and offer a path to a person, and the recording and transcription of calls, which the meeting invite needs to mention. Legal says where transparency is a legal or contractual obligation, including large customers’ contracts that already carry an AI clause, and marketing and post-sale leaders decide where it is a brand choice.

Mandatory training

Nobody gets access to the approved tools without the basic training, which covers the policy, the data classification, how to ask AI for work and how to check what it returns. A document read during onboarding changes little in the team’s behavior, so the training works with cases from each person’s own area. I cover formats, audiences and measurement in AI training for companies.

Policy owner and periodic review

The policy has a named owner, in IT, compliance or legal, supported by a small committee with HR, IT, the data protection officer and a business leader. In the first twelve months, I recommend a quarterly review, because tools change features and contracts at that pace, plus an extra review after an incident or a regulatory change. The policy also needs a channel for reporting incidents, with the rule that whoever quickly flags data pasted in the wrong place is not punished, otherwise the company only finds out when the customer complains.

AI usage policy template: the structure

The structure below is my starting point. It fits in four to six pages, plus an appendix of examples by area, which is the part the team reads most.

Section	What it contains	Who writes it
1. Purpose and scope	Who it applies to (employees, contractors, third parties with access) and which tools it covers	HR and legal
2. Approved tools	List, corporate accounts, settings and how to request a new one	IT
3. Data classification	The four levels and what can go into each type of tool	IT, legal and data protection officer
4. Permitted and prohibited uses	Examples by area and the lists of what the agent does alone and what it never does	Sales, marketing and post-sale leaders
5. Human review and accountability	What requires approval, what the review checks and who answers	Leaders and legal
6. Registry of agents and automations	Registry fields, who maintains it and what happens to an agent with no owner	RevOps, CRM admin or IT
7. Transparency with customers	When to inform, with what text and on which channel	Legal, marketing and post-sale
8. Training	Mandatory track by audience and refreshers	HR and L&D
9. Incidents	What to do when data goes where it should not and how fast to report it	IT and data protection officer
10. Governance	Owner, committee, review calendar and version control	Policy owner

How does the policy change daily work in sales, marketing and post-sale?

In sales, proposals with pricing are worked on in the approved tool or inside the CRM, and the call transcript stays attached to the deal. The follow-up the agent drafts goes back to the salesperson before sending, and the manager gains an objective criterion: the message went out with the salesperson’s signature, so the salesperson answers for it. For people who already used AI on their own, the policy swaps hidden use for a tool with access to the customer’s history.

In marketing, the lead database stops leaving the CRM as a spreadsheet to feed outside tools, and AI-generated content goes through the review flow the area already has for numbers, quotes and image rights. In post-sale, the virtual assistant identifies itself, a critical ticket only gets a reply after an analyst checks it and the customer conversation stays in the support tool. Leaders in both areas gain, through the agent registry, a view of everything that talks to customers on the company’s behalf.

AI policy and data protection law: where do they meet?

Data protection law still applies when personal data goes into a prompt, a transcript or the base an agent consults. In Brazil that law is the LGPD. Pasting the customer list into a public chat is processing of personal data, with a purpose, a legal basis and a vendor involved, and the company answers for it. That is why the data protection officer needs a seat on the committee, and the four-level classification needs to match the data mapping the company has already done. The points that come up most at this intersection are the contract with the AI vendor and the purpose of each use involving personal data. I cover the topic in AI and data protection, also as management guidance, and the legal reading of your operation belongs to your company’s legal team.

What should HR and L&D consider in the AI policy?

HR usually recognizes the need before the executive team does. People ask in the hallway whether they can use a tool, reimbursement requests show up for subscriptions paid on personal cards and one manager bans AI in their area while the manager next door encourages it. When these signs show up together, the company already has an informal policy, different in each area, and it is worth writing the official one.

Everyone who will have access to the tools takes the basic module on policy and classification, which can be recorded. Leaders take a decision module, where they design the approval lists for their area, and whoever administers the CRM and builds agents takes a technical module on the registry, permissions and credits. HR coordinates the track and adds the policy to onboarding for employees and contractors.

A signed acknowledgment says little about behavior. Over the next three months, track use of corporate accounts, requests for new tools through the official channel, incidents reported by the people themselves and reimbursements for personal subscriptions. Requests and incidents rising at the start are a good sign, because they show hidden use coming to the surface.

How to roll out an AI usage policy in 30 days

This is the plan I use with the committee, from mapping real use to a published and trained version 1.

Week 1: map real use, without a witch hunt. Run an anonymous survey on the tools each person uses, with which account and for which tasks, and cross it with finance’s reimbursements. Set up the committee, name the policy owner and agree with leadership that use declared in this phase carries no punishment.

Week 2: write version 1. Use the ten sections from the structure table, start with data classification and send the text to legal with a review deadline. Contract corporate accounts for the tools the company will keep, so the policy launches with an alternative to personal use.

Week 3: validate with leaders and build the registry. Each leader fills in the permitted and prohibited uses and the approval lists for their team. The CRM admin or the RevOps team surveys the agents and automations already running and registers each one with an owner, or switches off those left without one.

Week 4: train, publish and schedule the review. Train leaders first and the team next, publish the policy where everyone can reach it, collect acknowledgments through HR and open the questions channel. Schedule the first review for 90 days later, with HR’s indicators on the agenda.

The policy covers the rules, and the rest depends on teams knowing how to use what the company has opened up: asking with context, checking what comes back and deciding what to approve. That is the work of the in-company training on applied AI, with cohorts for leaders, sales, marketing and post-sale built on the company’s process and rules, which delivers a prompt library for each area and the agent architecture, showing what each agent reads, what it executes and who approves.

Frequently asked questions

What is an AI usage policy?

It is the internal document that defines which AI tools the team can use, which data can go into each one, what goes through human review and who answers for the result. It also includes the agent registry, mandatory training and an owner who reviews it.

Is banning ChatGPT and other tools at work an option?

A ban tends to move use to personal accounts, out of the company’s sight. It works better to approve corporate tools, state which data goes into each one and offer a fast path to request new tools.

Can employees paste customer data into an AI tool?

It depends on the data level and the tool. Under the classification I propose, customer personal data stays out of public tools without a contract and only goes into a corporate tool or a CRM agent for the purposes the policy lists and legal has validated.

Who should own the AI policy in the company?

A named person, usually in IT, compliance or legal, supported by a committee with HR, IT, the data protection officer and a business leader. Each area’s leadership answers for the uses in its own team.

Does the AI usage policy need to go through legal?

Yes. This article offers management guidance, and legal needs to validate the data classification, the transparency rules and the fit with data protection law and the contracts the company has already signed.

How does HR measure whether the AI policy is working?

A signed acknowledgment measures little. Track use of corporate accounts, requests for new tools through the official channel, incidents reported by the people themselves and the drop in reimbursements for personal subscriptions over the following three months.