ChatGPT for sales: instructions, memory, and data controls

A serious ChatGPT sales setup separates four kinds of context: stable seller defaults, remembered preferences, account- or deal-specific work and current task evidence. It also defines what data can be used, which apps are permitted and which outputs require review.

The goal is not to make ChatGPT “know everything about you.” The goal is to give it the minimum reliable context for the current sales job without blurring accounts, exposing restricted information or turning memory into a shadow CRM.

Before you paste sales data into ChatGPT

Answer these questions:

1. Is this a personal or managed business workspace?
2. Does company policy allow this customer, prospect, employee or commercial data in ChatGPT?
3. Is the information required for the task, or can it be removed or de-identified?
4. Which memory, retention and training settings apply?
5. Are connected apps permitted, and what can they read or do?
6. Who must review the output before it reaches a customer or system of record?

For ChatGPT Business, Enterprise, Edu and other covered business products, OpenAI states that organization data is not used for model training by default. Personal Free, Plus and Pro workspaces use separate data controls; OpenAI currently documents an option to turn off “Improve the model for everyone.” Your company’s own rules may be stricter than OpenAI’s defaults.

Put each kind of context in the right place

Context	Best home	Examples	Do not put here
Stable seller defaults	Custom Instructions	Role, sales motion, preferred output format, evidence and review rules	One customer’s confidential details
Lightweight personalization	Memory, if enabled and approved	Communication preferences or durable personal working context	CRM facts, deal stages, pricing or complete playbooks
One account, deal or territory	Project	Files, project instructions, chats, current-state brief	Unrelated accounts or general seller preferences
Current job evidence	The prompt or Work brief	Transcript, notes, source list, exact objective and constraints	Assumptions presented as facts
External system data	Approved app/plugin	Drive files, messages or other supported sources within permissions	Credentials pasted into chat or broad access “just in case”
Official customer/deal record	CRM or system of record	Contacts, stage, forecast, activity, commitments	ChatGPT memory or a stale Project summary

Set Custom Instructions for reliable defaults

Custom Instructions should contain context that improves many sales conversations and changes infrequently.

Sales Custom Instructions template

I am a [role] selling [product/category] to [ICP and typical buyer roles] through [sales motion].  
 
My default working standards:  
- Lead with the sales decision or outcome, not generic copy.  
- Use only facts I provide or sources available in the current task.  
- Separate verified fact, buyer statement, inference, hypothesis and unknown.  
- Do not invent contacts, quotes, priorities, metrics, customer proof, MEDDPICC evidence or commercial terms.  
- Ask for missing information when it materially changes the result.  
- For external messages, use concise, direct language and preserve the buyer’s wording where supplied.  
- For analysis, show evidence, risk, next action, owner and review required.  
- Never send messages, write to systems, quote pricing or create commitments without explicit approval.  
 
Preferred output conventions: [tables/bullets/briefs]. Avoid: [brand or style constraints].

Do not paste the entire company playbook into this field. Link or upload approved material where the relevant capability can use it, and keep the global defaults compact.

At the 30 August 2026 verification date, OpenAI documents Custom Instruction limits of 1,500 characters for Free and Go and 5,000 characters for Plus, Pro, Enterprise, Business and Education. Treat these limits as volatile and recheck them rather than designing the page around the numbers.

Use memory as personalization, not a database

When enabled, memory can remember useful context from chats, files and connected apps. That can reduce repetition, but it is not complete, immutable or suitable as the authoritative record for sales activity.

Use memory for durable preferences that are safe to reuse. Periodically inspect what ChatGPT remembers, delete incorrect or unnecessary items and turn memory off where the task or policy requires a clean boundary.

Do not rely on memory for:

• current deal stage or forecast category;
• contact and stakeholder records;
• customer commitments;
• approved pricing and discounts;
• renewal dates or contractual obligations;
• qualification evidence;
• sensitive customer information.

Use the CRM for those facts and a Project for the working context.

Use Projects for account and deal context

Create a Project when the subject will generate several conversations or files. Add project-specific instructions, a dated current-state brief and only the source material required for the work.

Choose the Project memory boundary deliberately. OpenAI currently allows eligible personal Projects to switch between default and project-only memory later; shared Projects automatically use project-only memory and cannot switch back. See the complete sales Project guide.

Review data controls

Open Settings and review the controls available to the account or workspace. Confirm:

• whether new conversations may be used to improve models;
• memory and reference-chat-history settings;
• retention or deletion behavior;
• export and account controls;
• Temporary Chat behavior for appropriate cases;
• workspace-level restrictions and administrator policies.

Personal users who do not want new conversations used for training should follow OpenAI’s current opt-out instructions. Managed-workspace users should follow company policy and administrator controls rather than assuming the personal-workspace UI applies.

Connect apps deliberately

Connect an app only when a specific sales job requires it.

1. Read the current plugin/app listing and privacy terms.
2. Check workspace enablement and role access.
3. Review the provider permissions requested.
4. Choose the minimum scope.
5. Test retrieval with non-sensitive material.
6. Test any action separately and confirm approval behavior.
7. Record an owner and review date.
8. Disconnect the app when the use case ends.

See the current plugins, skills and apps hierarchy.

Separate accounts and deals

Do not mix unrelated accounts in one long general chat. Use a dedicated Project for a strategic account or opportunity and project-only memory where isolation is required.

For lightweight one-off work:

• provide only the required context in the prompt;
• remove identifying details where possible;
• state the source date;
• label assumptions;
• record approved outcomes back in the CRM.

Add a review rule to every high-risk job

Job	Required review
Account or executive research	Verify sources and avoid private-personality inference
Customer-facing message	Check recipient, facts, claims, tone, CTA and commitments
CRM update	Confirm field mapping, source evidence and record owner
Forecast or pipeline analysis	Validate data timestamp, calculations and methodology
Pricing or negotiation	Check approved commercial boundaries and give/get rules
Proposal, RFP, security or legal response	Obtain subject-matter and formal approval
Connected-app action	Confirm target, scope, side effects and authorization

A 15-minute sales setup sequence

1. Confirm the correct workspace and company policy.
2. Review Data Controls and memory.
3. Add compact sales Custom Instructions.
4. Create one Project for a real account or deal if persistent context is required.
5. Add only approved, current files.
6. Run one canonical Promptifi prompt.
7. Check whether the output labels evidence and uncertainty correctly.
8. Record the useful result in the proper system of record.

Then expand deliberately into Deep Research, Work, skills, apps or Scheduled Tasks based on the job—not because the feature exists.

Start with a safe default and one real use case

Configure the evidence and approval rules, then try a bounded task such as meeting preparation or transcript analysis with approved material.

Browse 50 ChatGPT sales prompts or choose a capability from the ChatGPT for Sales hub.

FAQ

What should a salesperson put in ChatGPT Custom Instructions?

Use stable role, market, sales-motion, evidence, output and review preferences. Keep account- and deal-specific information in a Project or the current task.

Should sales reps enable ChatGPT memory?

Only if it is allowed and the convenience is useful. Treat memory as personalization, inspect it periodically and do not use it as a CRM or authoritative deal record.

Does OpenAI train on ChatGPT Business or Enterprise data?

OpenAI states that organization data from covered business products is not used for training by default. Review the current business-data commitment and your own policy.

Can I change a Project’s memory mode later?

OpenAI currently allows eligible personal Projects to switch between default and project-only memory. Shared Projects automatically use project-only memory and cannot switch to default.