← All postsHow Do You Use ChatGPT for Sales?

How Do You Use ChatGPT for Sales?

Carlos GarciaCarlos GarciaPublished

How Do You Use ChatGPT for Sales?

Most advice about using ChatGPT for sales is a list of prompts. Paste this to write a cold email, paste that to handle an objection. It is not wrong, exactly, but it describes the smallest and least interesting thing the tool does, and it ignores the part that actually changes a week of selling: connecting it to the systems where your deals already live.

That part has moved fast. ChatGPT can now be connected to external applications, and at least one major CRM documents a connector that does not just read your pipeline but creates and updates records in it. That is a different proposition from a prompt library, and it comes with a different set of questions about approval, auditing and what happens when the model gets something wrong.

This guide covers what ChatGPT is genuinely good at in a sales workflow, how to set it up properly, how the CRM connection actually behaves, and the specific things that OpenAI's own documentation does not tell you. Everything about the integrations is taken from first-party pages, and where those pages contradict each other or stay silent, that is reported rather than smoothed over.

Free audit

Where do you stand in AI search?

A manual SEO + AI search audit of your site: how ChatGPT, Gemini, Perplexity and Claude read and cite it today, and what to fix first. Report within 48 hours, no credit card.

Get my free audit

Or have it all done: AI SEO services, $999 once

Can ChatGPT Actually Help a Sales Team?

Yes, in four areas, and the four are unevenly useful.

It is strong at preparation. Reading a long document and pulling out the five things you need before a call is the task it does best, and it is the task salespeople do most often and most badly, usually in the ninety seconds before the call starts.

It is strong at first drafts that you then rewrite. Not because the drafting is excellent, but because a mediocre draft is enormously easier to improve than a blank page, and the time it saves is real even when you keep almost none of its words.

It is useful at summarizing and extracting next steps, which is the admin tax on selling. Turning a messy set of notes into a short recap with owners and dates is mechanical work that nobody enjoys and that nobody checks carefully, which makes it a good fit.

The unevenness is the point. A tool that is excellent at three tasks and unreliable at a fourth is still worth having, provided you know which is which before you put it in front of a customer.

It is weakest exactly where sales teams want it to be strongest: knowing things. It does not know your pricing, your competitors' current terms or what your customer said last quarter unless you give it that information or connect it to a system that holds it. Treating it as a knowledge source rather than a reasoning and drafting tool is the single most common way sales teams get burned by it.

Selling a product nobody can find in search is the hard way to do it. Get a free SEO audit from SEO Stuff and see how much of your pipeline you could be sourcing organically.

What ChatGPT Does Well in a Sales Workflow

Here is what each of those four areas looks like in practice, rather than in the abstract.

Research before a call

The useful version of this is not "tell me about this company." It is handing over the material you already have and asking narrow questions of it: the prospect's last three email replies, a transcript of the discovery call, the deal record. Ask what the stated objection was, who else was named as a decision maker, and what was promised but never delivered.

That framing matters because it keeps the model working on text you control. The failure mode of open-ended company research is that the answer sounds authoritative and is assembled from whatever the model happens to have absorbed, which may be three years old.

Drafting that keeps your voice

The way to get useful drafts is to feed it two or three of your own emails that worked and ask it to match them, rather than asking for "a professional follow-up email." Generic requests produce generic output, and the generic register of AI-written sales email is now recognizable to anyone who gets more than ten a day.

Ask for three variants rather than one. Picking between options is faster than editing a single draft, and it surfaces the angle you had not considered more often than a single attempt does.

Summarizing and extracting next steps

Give it notes, ask for a recap of no more than six lines, with every commitment made by either side and who owns it. Then read it. The point of the exercise is that you check a short list rather than reconstruct a long conversation, and that only works if you do the checking.

A useful discipline here is to ask for the recap in a fixed shape every time: commitments, owners, dates, open questions. Consistency makes the output scannable, and scannable output is output you will actually read rather than file.

Objection and pricing preparation

Describe the objection in the prospect's own words and ask for the three strongest counters and the one case where the objection is simply correct. That last part is the valuable one. A tool that only argues your side is a tool that walks you into the deal you should not have chased.

None of the four is a prompt trick. All four are about what you put in, which is why the setup below matters more than any wording you could paste.

How to Set ChatGPT Up for Sales Work, Step by Step

Setup is where most of the value sits, and most people skip it.

Step 1: Decide which account you are using. OpenAI states that data from its business tiers is excluded from model training by default. Its business data page says: "By default, we do not use data from ChatGPT Enterprise, ChatGPT Business, ChatGPT Edu, ChatGPT for Healthcare, ChatGPT for Teachers, or our API platform—including inputs or outputs—for training or improving our models." If you are pasting customer information into a personal account, you are making a different decision than you think you are.

Step 2: Write your context once, not every time. Keep a short block of text covering what you sell, who you sell to, the three objections you hear most and the tone you write in. Paste it at the start of a working thread rather than re-explaining your business in every conversation.

Step 3: Give it your own best examples. Two or three emails that actually got replies, one call recap you were happy with, and your current pricing framing. This is the difference between output that sounds like you and output that sounds like everyone.

Step 4: Connect the applications where your work lives. In ChatGPT, these are reached through Settings. OpenAI's own help page gives the path as "Open Settings > Plugins, and select the app or the plugin that includes it," then installing and connecting it. Once connected, you can bring one into a conversation by mentioning it with @ or by selecting it from the + menu.

Step 5: Connect your CRM, if it supports it. HubSpot documents its own connector for ChatGPT, and the steps are on HubSpot's side of the integration rather than OpenAI's. HubSpot's instructions run: in ChatGPT, select the profile icon, then Settings, then Apps, select HubSpot and click Connect, continue to HubSpot, authenticate, pick the account, choose permissions, and click Connect App.

Step 6: Set the permission level deliberately. This is the step that matters most and takes the least time, and it gets its own section below.

Step 6b: Test it on a record that does not matter. Create a throwaway contact, ask for a property update, and watch what the approval prompt actually shows you. Five minutes here tells you more about the integration's behavior than any documentation, including this article.

Step 7: Check what your administrator has already decided. OpenAI's page notes that in Business workspaces apps are enabled by default with administrator control, and that Enterprise workspaces can restrict new connections. Whether you can connect anything at all may not be your call.

Your best-converting content should be working while you sleep. A free SEO audit from SEO Stuff will tell you which pages are closest to ranking and worth the effort.

How the HubSpot Connector Actually Writes to Your CRM

This is the part that makes ChatGPT a sales tool rather than a writing tool, and it is also the part most worth reading carefully before you switch it on.

HubSpot's page is explicit that the connector writes. It states that the connector can "Create and update HubSpot CRM records, including contacts and deals, from the ChatGPT chat window," and its FAQ answers the question of whether ChatGPT can update HubSpot data with a plain yes: it can create and update CRM records and log activities from the chat window. The actions it names include creating contacts and deals, updating record properties such as deal stages, logging notes, emails, calls, meetings and tasks, and creating follow-up tasks.

The write permission is surfaced in ChatGPT as "Write tools," and HubSpot recommends setting it to "Needs Approval" rather than "Always allow." Its own warning is that with "Always allow" selected, edits "may occur without asking for approval." With approval required, ChatGPT shows you the proposed change and you confirm it.

Take that recommendation. The asymmetry is stark: the upside of automatic writes is saving one confirmation click, and the downside is a model quietly moving deal stages in your pipeline because it misread an instruction.

It is also worth understanding why HubSpot, rather than OpenAI, is the one spelling this out. The connector is HubSpot's product, built against ChatGPT, so HubSpot is the party that knows and documents which of its objects accept writes. That division of labor explains several of the gaps in the next section.

A few concrete limits are documented and worth knowing before you plan a workflow around them.

  • Bulk operations are capped at 10 records at a time for creates and updates.
  • Custom validation rules, including pipeline stage validations and association label validations, are not applied when changes come through the connector. Your CRM's own guardrails do not protect you here.
  • Every action is attributed in your HubSpot audit log to both the user and the ChatGPT connector, so there is a trail.
  • Access is bounded by your existing HubSpot permissions: users "can access only the data allowed by their existing HubSpot user permissions."
  • Super Admins, or users with App Marketplace Access permission, can connect it without prior approval; everyone else needs a Super Admin to approve it and configure its data permissions.
  • The object tables show no delete capability for any object, though the page never says in words that deletion is impossible.

The validation-rules point deserves a second reading if your team relies on them. Many sales operations teams use required fields and stage validations precisely to stop deals being moved without the supporting information. A write path that bypasses those rules is a hole in a control somebody deliberately put there, and it is worth telling them about before you switch it on rather than after.

HubSpot also says the connector respects your data center location, and that for Enterprise, Business or Education ChatGPT plans, connector data is not used for training; on Free, Go, Pro or Plus, users may opt out, and if they do not, OpenAI "may use connector data to improve connector performance" while not otherwise using it for general model training.

What OpenAI's Own Pages Do Not Tell You

Four gaps are worth naming, because each one costs time if you assume otherwise.

OpenAI's page never says which applications can write. It states that apps can "Take supported actions, such as creating or updating information," and that "Sensitive actions may require approval or be denied." It offers permission levels of "Always ask," "Allow read actions," "Allow low-risk actions," and "Allow all actions," noting that the last "carries elevated risk." What it does not do anywhere is tell you whether a given application is read-only or read-write. You find that out from the other vendor's documentation, or by connecting it and looking.

OpenAI does not name the major CRMs at all. Neither HubSpot nor Salesforce appears on its connected-apps page. The HubSpot integration is documented in detail by HubSpot and not by OpenAI, which means the authoritative page for a ChatGPT feature is on someone else's website. If you search OpenAI's help center for your CRM and find nothing, that is not evidence the integration does not exist.

There is no plan matrix. OpenAI's page says availability "depends on the app, your ChatGPT plan, region, workspace, role, model, and the ChatGPT interface you use," which is seven variables and no table. There is no way to determine from that page whether your plan gets a given application.

Even the vocabulary has drifted. The article's own web address describes connectors, while the page itself is titled "Connected apps in ChatGPT" and talks about apps and plugins, with "connector" appearing nowhere in the text. HubSpot, meanwhile, calls its integration a connector throughout. Two first-party sources, two names for one thing.

None of these gaps is a reason not to use the integration. They are a reason to read two vendors' pages instead of one, and to assume nothing about capability that you have not either read or tested.

One more gap is specific to sales teams who share their work. OpenAI's help page for ChatGPT Business answers the question of whether shared links can be disabled entirely with a flat "No," adding that the capability is available in ChatGPT Enterprise. If your concern is a rep sharing a chat containing pricing or customer detail outside the company, a Business workspace does not give an administrator a switch for that.

Not sure where your organic traffic is leaking? Get a free SEO audit from SEO Stuff and find the pages worth fixing first.

Where It Falls Short for Sales

Some of these are the model's limits and some are the integrations' limits, and it is worth keeping them apart.

Accuracy is the obvious one, and the vendors say so themselves. HubSpot's page advises human review and notes that AI output may be wrong. In a sales context the specific risk is not a bad sentence but a confident number: a discount, a renewal date, a contract term stated fluently and incorrectly to a customer.

Documentation can contradict itself, which makes planning harder. HubSpot's own page says in its FAQ that "Custom objects, unstructured data, and associated engagements are all not yet available," while its object table lists custom objects as readable, creatable and updatable. The page does not reconcile the two. If your pipeline depends on a custom object, test it rather than trusting either statement.

Dates are hard to pin down on OpenAI's side. Its help pages show relative timestamps such as "Updated: 9 days ago" rather than absolute ones, so you cannot tell from a page whether a limit you read six weeks ago still holds. HubSpot's page, by contrast, carries a visible last-updated date of November 12, 2026, which is the better practice of the two.

OpenAI's business pages do not fully agree about data sharing either. Its ChatGPT Business help page says flatly that OpenAI "won't train on your workspace's data," while its consumer data page says organizations "can choose to share data when that option is available." Both are first-party, and neither explains the other. Treat "never trains" as the default rather than as an absolute, and check your own workspace settings.

Permissions can also go stale in a way that looks like a bug. HubSpot notes that after a Super Admin changes data permissions for new objects, each user needs to reconnect the connector. If a colleague's connector suddenly cannot see something yours can, that is the first thing to check.

And the connector does not do everything the demo implies. HubSpot notes that for drafted campaigns or workflows, "ChatGPT will provide the drafted content for you to implement in HubSpot." The drafting is automated; the building is not.

How ChatGPT Compares to the Alternatives for Sales

There are three realistic shapes of answer here, and which wins depends on where your selling day actually happens.

  • ChatGPT plus a CRM connector. The strongest option if your CRM supports it, because the pipeline data comes into the conversation and the resulting updates go back out with an approval step and an audit trail. HubSpot's documentation of its own connector is the most detailed first-party account of this that currently exists.
  • An assistant bundled into the office suite you already use. Wins on friction rather than capability. If your day is email, calendar and documents in one vendor's suite, an assistant sitting inside those windows gets used, and the best tool is the one you do not have to switch to.
  • Your CRM's own built-in AI. Wins on safety and loses on flexibility. It understands your data model natively and cannot be talked into something strange, but it will only ever do what the CRM ships.

Notice that none of the three is distinguished mainly by model quality. For the work described in this article the models are close enough that the integration, the permissions and the place the window opens decide the outcome.

The verdict: if your CRM is HubSpot, ChatGPT with the HubSpot connector set to require approval is the best-documented path available today, and the audit-log attribution is the detail that makes it defensible to whoever owns your data policy. If your sales work lives inside an office suite all day, pick the assistant already in those windows and accept a less capable model in exchange for actually using it. And if you have no CRM connector available, do not force it: use ChatGPT for preparation, drafting and recaps, and keep writing to your CRM by hand. The prompt-library version of this tool is still worth having. It is just not the version that changes your week.

Pipeline from search compounds in a way that outbound does not. Start with a free SEO audit from SEO Stuff and build a channel that keeps paying.

Final Thoughts

The useful mental model is that ChatGPT is a good reasoner with no knowledge of your business, and that every step of setup is about closing that gap. Context block, your own best examples, then a connection to the system that holds the truth. Do those three things and the output stops sounding like a template; skip them and no prompt will rescue it.

If you take one decision away from this, make it the permission setting. "Needs Approval" on write actions costs you a click per change and removes the entire class of problem where a model edits your pipeline without being asked. HubSpot recommends it, and the recommendation is correct.

And read the other vendor's documentation, not just OpenAI's. The most detailed, most current and most specific page about using ChatGPT with a CRM is published by the CRM company. If you are weighing up which account tier your team needs before any of this, our guide to what ChatGPT Enterprise actually includes covers the administrative and data commitments that sit underneath everything described here.

Want this done for you?

SEO Stuff gets businesses cited by ChatGPT, Gemini, Perplexity and Claude, and ranking in Google. Start with the free audit, a call, or the package.

See where you stand

A free, manual SEO + AI search audit of your site, with a report within 48 hours. No credit card, no sales call.

Get my free audit

Talk it through

A free 15-minute call with the founder about whether AI search is a meaningful opportunity for your business.

Free strategy call

Have it done

The Done-For-You Package: audit, 10 pages of content, 3 DR50+ placements, dashboard. $999 once, delivered in 21 business days.

See the package