Blog

What Is a Customer Harness? HubSpot’s Headless CRM Explained

Written by Rob Guinn | Nov 18, 2026, 2:10:11 PM

Here's a question worth asking: what is a CRM for when nobody has to open it anymore?

For a couple of decades, the answer has been obvious. A CRM is a database with screens on top. People open the screens, type things in, click through records, run reports. The value of the CRM was roughly the value of the people clicking around in it, and the vendors competed on how pleasant the clicking was.

Then AI started doing a lot of the clicking. And it turns out AI doesn't care about the screens at all. It cares about three other things, and those three things are what HubSpot is now calling a harness. Brian Halligan put it as "the Customer Harness Company." Dharmesh Shah called it the world's first Customer Relationship Harness.

You can read that as a rebrand. I think it's more useful to read it as the name of a new software category, one that every CRM vendor is going to end up in whether they use the word or not (we're a HubSpot partner, so grade this accordingly). So here's my attempt at describing the category: what it is, why it's showing up now, and why you should care if you run revenue operations for a living.

What Is a Customer Harness?

A harness is the thing an AI plugs into to do customer work. Not the AI itself; that's interchangeable, and it's going to keep being interchangeable. The harness is everything around it that makes the AI useful and safe in your business specifically. One way to express the model is:

  • What it knows. Your contacts, companies, deals, tickets, emails, call transcripts and meeting notes, connected to each other so the AI can tell that this email is about that deal at this company.
  • What it can do. Update records, run workflows, put agents to work, build quotes. The actions, not just the reading.
  • What it's allowed to do. Who can see what, which actions need a person to sign off, and where the line is before anything reaches a customer.

A CRM has always had some version of all three, but they were designed around a human. The data was organized for a person reading a screen, the actions were buttons, and the permissions assumed that people mostly don't go looking. A harness is the same three things redesigned around a reader that reads everything, acts instantly and never gets bored.

The connection between the AI and the harness is MCP, the Model Context Protocol. The best way I've found to think about it is a wall outlet. You don't rewire your house every time you buy a new appliance; anything with a standard plug just works. But the outlet isn't the interesting part. The wiring behind the wall is: which circuits are live, what each one can power, and where the breaker trips. MCP is the outlet. The harness is the wiring. Claude, ChatGPT or HubSpot's own Breeze are the appliances, and you'll swap them more often than you think.

Any AI plugs in over MCP. What it knows, can do and is allowed to do comes from the harness, and what fills the harness comes from you.

Why Is the CRM Going Headless Now?

Three things happened at roughly the same time, and together they make the category inevitable.

First, the models stopped being the moat. Every few months a different model is the best one, and the gap between the leader and the runner-up keeps shrinking. If you're a software company, betting your future on having the best model is a bet you will lose most quarters. Betting on being the place every model works from is a much better trade.

Second, the plumbing got standardized. Before MCP, connecting an AI to your CRM was a custom integration project. Now it's closer to plugging in a cable, which means the question shifts from "can the AI reach my data" to "what does it find when it gets there."

Third, the work moved. A lot of people (me included) now do a large share of their HubSpot work without opening HubSpot. They ask Claude which deals need follow-up, build a quote in a conversation, update records with a sentence. Dharmesh's phrase for this is that sometimes the interface needs to fade into the background, and "headless" is the industry word for it.

My own version starts before I've opened my inbox. Every morning, Claude goes through my email and my HubSpot portal and hands me four things: the emails that actually need me, the deals I should check on, summaries of my calls, and the action items I'd otherwise forget. I haven't opened HubSpot to get any of it. It's all HubSpot data; the work just happens somewhere else.

Once the work leaves the app, the app's screens stop being the product. What's left is the harness.

How Is HubSpot Pricing Headless Access?

HubSpot has now put a price on that, which is about the most revealing thing a software company can do. In HubSpot is headless. Here's how we'll price it, published November 17, HubSpot lays out three ways in:

  • Your seat. Everyday work from Claude, ChatGPT, Gemini or Microsoft Copilot, through HubSpot's connectors, its MCP server and its Agent CLI, is included with the seats you already pay for. Heavy users can add capacity with HubSpot Credits instead of buying a different seat.
  • Your own agents. An agent you've built or bought connects through the same APIs as any other integration, and its calls count against the API capacity already in your plan.
  • A governed environment for agents. For agents doing work at scale, HubSpot is building a separate environment with purpose-built agentic APIs, control over what each agent can access and do, and a record of every action. That work runs on credits.

Notice what is and isn't being charged for. Reaching your own data from an outside AI isn't; HubSpot's framing is that other platforms are effectively taxing customers for using their data with outside AI. What gets charged for is the governed layer: context, control and accountability. HubSpot is pricing the harness, not the head. That's about as clear a statement of the category as you'll ever get from a pricing announcement.

What Does the Future of CRM Look Like With AI?

This is the part where I'm guessing, so take it with the appropriate amount of salt. But if the category plays out the way the logic suggests, a few things follow.

  • You'll run more than one AI on the same harness. Sales uses Claude, support uses Breeze, someone in finance insists on ChatGPT, and they all work from the same customer record under the same rules. The harness is the one thing they share.
  • The interface becomes the exception. You'll still open HubSpot to review, approve and fix things, but most routine work will start with a sentence typed somewhere else. The screens become the place where humans check the AI's work rather than the place where the work happens.
  • Switching costs move. Today, part of what keeps you on a CRM is that your team knows the screens. Tomorrow, what keeps you on it is everything inside the harness: years of connected history, the actions you've exposed, the rules you've tuned. That's a stickier kind of lock-in, and it's worth noticing before you're in it.
  • The admin job changes. The person who used to build page layouts and property groups ends up designing what the AI can see, what it can do, and where it has to stop and ask. That's a harder, more valuable job, and most teams haven't started hiring for it.
  • Harness quality becomes a competitive advantage between companies, not just between vendors. Two companies on the same CRM with the same AI will get very different results, and the difference will be what's strapped into the harness.

What Does the Harness Mean for HubSpot Admins?

If you own HubSpot at your company, the category shift changes what your job actually is.

For years, CRM cleanup has been the work everyone agrees matters and nobody funds. Duplicate companies, lifecycle stages nobody maintains, transcripts that never got associated to the right deal. A human works around all of that by instinct. An AI treats it as truth. In a harness world, messy data isn't a reporting annoyance anymore; it's the thing that decides whether every AI your company buys gives good answers or confident wrong ones.

The same goes for process and permissions. We've watched companies give everyone an AI login and see almost nothing change, because the AI was bolted onto the old way of working. We've watched others roll out a promising new capability and stall in security review because nobody built permission sets for finance data. Neither of those is a model problem. Both are harness problems.

What Are HubSpot Agent Users?

HubSpot is already moving on the permissions part. It just shipped Agent Users (Manage your custom agent's permissions like you manage your team, HubSpot, November 16): every custom agent built in HubSpot now gets its own user identity, managed with the same roles, permission sets, teams and object permissions as your people. Agents no longer run on their builder's full access. Permissions set a ceiling on what each agent can reach, risky actions like delete are restricted by default, and every change an agent makes lands in the Audit Log under the agent's own name, with the before and after values and a way to restore the old one.

This is what the third part of a harness looks like when it's taken seriously. It answers the two questions security teams always ask about agents: what is it allowed to do, and can we prove what it did? Those are exactly the questions that have been stalling rollouts.

Two caveats worth knowing. First, it covers HubSpot's custom agents only. Customer agents, Agent Builder workflows and third-party agents aren't included, so an outside AI plugged in over MCP still works with whatever access the person connecting it has. HubSpot's pricing post says a governed environment for third-party agents is coming, but for now that's a plan, not a setting. Second, the default is broad: agents start with a wide baseline of permissions, so the ceiling is only as low as someone sets it. The controls have arrived. Setting them is still your job.

So here's the uncomfortable, and I think mostly good, implication: the most valuable AI work at most companies over the next couple of years won't look like AI work. It'll look like data modeling, process redesign and access control. The people who do it well will make every AI their company touches better at once. That's a pretty good position to be in.

Where Could the Harness Idea Go Wrong?

Two honest ways this goes differently.

The first is that the AI tools build their own harness. Claude and ChatGPT are getting memory, projects and connectors of their own. If enough customer context ends up living inside the AI tool rather than the CRM, the CRM becomes a dumb database the AI occasionally reads from, and the harness belongs to whoever owns the conversation. I don't think that's the likely outcome, mostly because permissions, workflows and a shared system of record are hard to rebuild and companies don't want their customer history scattered across individual chat accounts. But I'd put real odds on it, not zero.

The second is that "harness" stays a slide-deck word. Categories get named before they get built, and a vendor can call anything a harness. The test is simple: can an AI you choose, plugged in from outside, actually do customer work inside the rules you set? If yes, it's a harness. If it can only read, it's a database with a new label.

Frequently Asked Questions About the Customer Harness

Q: Is a customer harness the same thing as a CRM?

A: It is built from the same three things a CRM has always had: the data it knows, the actions it can take and the permissions that limit it. The difference is the reader. A CRM organized those for a person clicking through screens, while a harness organizes them for an AI that reads everything and acts instantly.

Q: What does it mean that HubSpot is headless?

A: Headless means work in HubSpot can happen without opening HubSpot. An AI such as Claude, ChatGPT, Gemini or Microsoft Copilot reads and updates HubSpot data through HubSpot's connectors, its MCP server or its Agent CLI. The HubSpot interface stays in place for people to review, approve and fix what the AI did.

Q: Does HubSpot charge extra to use Claude or ChatGPT with HubSpot data?

A: Not for everyday work. Under the pricing HubSpot published on November 17, 2026, work from Claude, ChatGPT, Gemini or Microsoft Copilot through HubSpot's connectors, MCP server and Agent CLI is included with existing seats, and heavy users can add capacity with HubSpot Credits. Custom agents count against your plan's API capacity, and a governed environment for agents at scale runs on credits.

Q: Do HubSpot Agent Users apply to third-party AI agents?

A: No. Agent Users cover custom agents built in HubSpot. Customer agents, Agent Builder workflows and third-party agents are not included, so an outside AI connected over MCP works with whatever access the person who connected it has. HubSpot has said a governed environment for third-party agents is coming, but it is not a setting yet.

Q: How should you prepare HubSpot for AI agents?

A: Start with the data an AI will treat as truth: duplicate companies, lifecycle stages nobody maintains and call transcripts that were never associated to the right deal. Then decide which actions agents can take, which need a person to sign off, and build permission sets for sensitive data such as finance before the rollout reaches security review.

What This Means for Your Team: The Harness Will Outlast the AI

You don't need to wait for the category to settle. Next time you evaluate an AI tool, or your CRM, stop asking how good the AI is and start asking about the harness: what will it know, what will it be able to do, and what will it be allowed to do. Then go look at your own portal and answer the same three questions honestly.

The AI you pick will probably change in a year. The harness won't.

Ready to find out what your HubSpot harness knows, can do and is allowed to do? Talk with our team about AI for HubSpot.