Somewhere in your sales team right now, a rep has a chat assistant open in one tab and the CRM in the other. They are pasting a call transcript in, asking for a follow-up email, and pasting the answer back out. Nobody approved it and nobody banned it. It saves them twenty minutes a day, so they keep doing it.
That part is manageable. The part most policies miss starts the day someone connects that assistant to the CRM itself, and the rep stops copying and starts asking it to “clean up my deals.”
What is an AI acceptable use policy for a sales team?
An AI acceptable use policy for a sales team is a short, written set of rules covering three things: which AI tools reps may use for work, which customer and CRM data may go into each one, and which AI outputs or actions need a person to review them first. Generic templates from security and HR vendors cover the first two. Sales needs the third, because sales AI writes to buyers and to the system of record.
A company-wide policy usually exists already, written by IT or legal. It tends to say “do not put confidential data into public AI tools” and stop there. That sentence does not tell a rep whether a prospect’s job title counts as confidential, whether a call recording can go into a summarizer, or whether they may let an assistant update fifty deal stages at once.
The sales version translates the company rule into the objects reps touch daily: contacts, companies, deals, notes, call recordings and emails. It should fit on one page. If it runs longer, reps will not read it, and the policy turns into a document that exists for audits rather than for the team.
Shadow AI in sales is mostly copy and paste
Shadow AI in a sales team is any AI tool reps use for work that nobody approved, and most of it looks harmless: transcripts pasted into a free chat tool, account research run in a personal account, emails drafted outside the CRM. The data risk is real. Forrester’s 2026 predictions tie more than $10 billion in lost enterprise value to ungoverned GenAI, and sales handles more customer data than most teams.
Banning the tools rarely works. Reps use them because they save time, and a ban moves the use to personal phones where you see even less. The better move is to approve a short list of tools that meet your data rules, make them easier to reach than the unapproved ones, and say plainly what may go into each.
The copy-and-paste half has one more cost that security reviews skip. Each AI-written email that leaves the building carries your brand. If the assistant drafted it from stale CRM data, the buyer sees the mistake before you do. Our paper on AI-generated sales emails and buyer trust covers what that costs and the review step that prevents it.
The bigger risk: a rep, a chat window and write access to the CRM
The larger risk sits in connectors that let a chat assistant read and update CRM records on a rep’s behalf. A rep who is not a CRM expert can ask for a bulk change in plain English, and the assistant will try: overwriting fields, merging or re-associating records, filling blanks with values it inferred. Nobody reviews the change, and the system of record absorbs it.
Picture the request. “Update all my open deals with a next step and a realistic close date.” The assistant reads each deal, guesses a next step from the last email, picks a date, and writes both. Some guesses are right. Some overwrite a date the manager set on purpose. A few deals get associated with the wrong contact because two people at the account share a surname. The rep sees a tidy pipeline and moves on.
Four questions decide how exposed you are:
- Who approved the connector? If the answer is “the rep installed it,” your CRM’s write permissions now depend on each rep’s judgment.
- What scopes does it have? Read access to contacts is one decision. Write access to deals, owners and associations is a much bigger one, and connectors differ by vendor and change often, so check your CRM’s own documentation for each.
- What did it change? If you cannot pull a list of records the connector edited last week, you cannot undo a bad batch.
- How would you know? This kind of damage tends to surface late: a forecast number that moved for no clear reason, or a customer email meant for someone else.
Our paper on why AI agents should not get write access to your CRM yet covers autonomous agents and field-by-field permissions. The policy question here is narrower and more urgent: the person-driven assistant that any rep can connect this afternoon.
Which CRM data can go into which AI tool?
Sort CRM data into three classes and match each to a tool tier. Public and business-card data (company, title, website) can go into any approved tool. Customer conversations and deal details (notes, recordings, pricing, contract terms) go only into tools under a company agreement. Regulated or sensitive data (personal phone numbers, health or financial details, anything a contract restricts) stays out of AI tools unless legal says otherwise.
Write the classes in the words reps use, not in data-governance terms. A rep should be able to look at a field and know which row it belongs to in under five seconds.
| Data class | Examples from the CRM | Free or personal AI tools | Approved tools under a company agreement | CRM-connected assistants |
|---|---|---|---|---|
| Public and business-card data | Company name, website, industry, a contact’s title and work email | Allowed | Allowed | Read allowed; writes reviewed |
| Customer conversations and deal details | Call recordings and transcripts, notes, pricing, discount, contract terms, competitors named by the buyer | Not allowed | Allowed | Read allowed; writes reviewed |
| Regulated or restricted data | Personal phone numbers, anything covered by a customer’s data terms, health or financial details | Not allowed | Only with legal sign-off | Not allowed |
| Forecast, ownership and compensation fields | Deal amount, close date, forecast category, deal owner, commission splits | Not allowed | Allowed for analysis only | Read only; no AI writes |
The last row is the one sales leaders add once they see it. Forecast, ownership and comp fields drive pay and board numbers. An assistant that edits them on a rep’s say-so turns a data-quality issue into a payroll dispute.
What shouldn’t reps let AI do unchecked?
Reps shouldn’t let AI send a customer message, change a forecast or ownership field, merge or delete records, or make a bulk edit without a person reviewing the result first. Drafting, summarizing and research are fine to automate. Anything that reaches a buyer or rewrites the system of record needs a named human to approve it, with a record of who approved what.
The list for the policy, in plain words:
- Send anything to a customer. AI may draft emails, follow-ups and proposals. A rep reads and sends each one.
- Change forecast, ownership or compensation fields. Amount, close date, forecast category, deal owner and splits change only by a person, in the CRM, with the change logged.
- Merge, delete or re-associate records. Duplicates and associations are where silent damage spreads. A person merges, using the CRM’s own merge tools, so the history survives. Our guide to HubSpot duplicate contacts covers why merges go wrong even without AI in the loop.
- Make bulk edits. Any change touching more than a handful of records goes through whoever owns the CRM, not through a rep’s chat window.
- Fill blanks with guesses. If a field is empty, the AI may flag it. It may not invent a value. A made-up job title or close date looks exactly like a real one six months later.
- Put restricted data into an unapproved tool. The data classes table decides, not the rep’s sense of what is sensitive.
Reps who enter data under pressure already shade it. An assistant that writes on request gives that habit a faster pen. Why good reps enter bad data covers the incentives behind it.
An AI acceptable use policy template for sales
Copy this, change the parts in braces, and have whoever owns the CRM and whoever owns security sign it together. Keep it to one page so reps read it. Review it each quarter, because the tools and their connector scopes change faster than most policies do.
{Company} sales team: AI acceptable use policy (version {date}, owner: {name, role}) 1. Approved tools. Reps may use these AI tools for work: {list}. Anything else needs approval from {owner} before it touches customer data. 2. Data rules. Follow the data classes table. When unsure which class a field belongs to, treat it as the stricter one and ask {owner}. 3. Human review. A person reviews each customer-facing message before it is sent, and each change to forecast, ownership or compensation fields before it is saved. 4. CRM connectors. Connectors are read-only by default. Write access is granted by {owner}, per tool and per object, in writing, and listed in the connector register. Reps do not connect AI tools to the CRM themselves. 5. Not without review. No AI sends to customers, merges or deletes records, makes bulk edits, or fills empty fields with guesses. 6. Audit. {Owner} reviews the connector register and the CRM’s change history for AI-made edits monthly, and can revoke a connector at any time. 7. Reporting. If an AI tool sends, changes or exposes something it should not have, tell {owner} the same day. Reporting a mistake is not a disciplinary issue; hiding one is.
Section 7 does more work than it looks. A policy that punishes people for reporting AI mistakes means you tend to hear about them last.
How do you know what a connector changed?
You know what a connector changed only if you planned for it before turning it on. Keep a register of each approved connector, its scopes and its owner. Give each connector its own integration user where your CRM allows it, so its edits show up under a name you can filter on. Then review that user’s change history monthly and spot-check a sample of the records it touched.
A connector register needs four columns: the tool, what it can read, what it can write, and the person who approved it. If you cannot fill in the third column from your CRM’s settings, assume the answer is “more than you meant” until you check.
The confidence gap here is wide. Validity’s State of CRM Data Management in 2025, a survey of 602 CRM users and admins, found 45% of companies’ CRM data isn’t prepared for AI, while 54% are already deploying GenAI tools on those foundations. If your team is in that 54%, ask whether anyone could list the records an AI tool edited last month. Across the 150+ B2B companies we’ve worked inside, roughly 7 in 10 had material CRM data accuracy problems.
Read-only until you can see what it writes
You can let each rep decide what to connect and find out what changed when a number looks wrong, or you can set connectors to read-only, name an owner, keep a register and review the change history monthly. The first is faster this quarter. The second is the one you can explain to your CFO when the forecast moves and someone asks why.
If you want to know how much of your CRM would hold up before an assistant starts writing to it, book your free diagnostic.