PSOhub Blog

Headless CRM and AI Harness: A Guide for HubSpot Users

Written by Bas van der Horst | November 19, 2026

TL;DR: A headless CRM is a CRM that AI tools like Claude and ChatGPT can work in without opening a screen. HubSpot is going headless. Your AI only becomes really useful when the tools after the deal are open too.

On 12 November 2026, Brian Halligan came back as CEO of HubSpot and called it a refounding. The plan is a headless CRM: a HubSpot your AI can work in directly. That's good news for every HubSpot user, but there's a catch. Your AI only sees what's connected, and in most service firms the work that decides your margin happens after the deal, in tools that aren't ready for it.

Below I explain the three words HubSpot now uses, what they mean for your business, and how to check the rest of your stack.

What do human, headless and harness mean?

Human, headless and harness are the three ways into software that HubSpot now talks about. Human is you, working in the screens. Headless means other software can use every function without a screen. Harness is an AI assistant that turns what you say into the right action, checks with you first and keeps a record.

HubSpot co-founder and CTO Dharmesh Shah gave each word a one-line definition in his Day 4 notes on the refounding. Headless is the programmatic interface to HubSpot. Harness is the conversational AI interface. Human is the classic interface you know today.

I find it easiest to explain with my electric car, because there are three ways to get in.

  1. Human. I see my car, press the button on my key and it unlocks. I do it myself, by hand.
  2. Headless. I walk out of the supermarket with my arms full of groceries. I can't reach my key and I don't need to, because the car notices the key in my pocket and opens. Nobody pressed a button.
  3. Harness. I forgot my key, so I open the app on my phone. The app knows it's me. I tell it to open the car and start it, it asks if I'm sure, I say yes, and it happens.
Human, headless and harness compared
  Human Headless Harness
Who does the work You, in the screens Other software, through an API or MCP An AI assistant, on your instruction
How it starts A click A call from another system A sentence in plain language
Typical use Planning, reviewing, approving Syncing data, automations Questions, updates, admin
What keeps it safe Your login and role The same rules the screens use A check, a preview, your yes and a log

 

Human, headless and harness are three ways into the same software. Whichever way a request comes in, the business software applies the same rules. Graphic: PSOhub.

Two things make this work. The app has no way into the car of its own: the key signal and the app signal end up at the same system in the car. That's why a harness always needs a headless layer underneath it.

The second thing matters even more. The car decides, not the app. Whether I press the button, walk up with the key or ask the app, the car checks whether this is an approved key or phone and whether I'm allowed to do this. Those rules are set once, in the car, and every way in has to follow them.

Your software should work the same way. The rules belong in your business software, not in the AI tool you talk to.

What is a headless CRM?

A headless CRM is a CRM where every function can be used by other software, without anyone opening a screen. In practice that means a full API and an MCP server, so Claude or ChatGPT can look up a deal, update a contact or log a call for you. You keep the screens, but they're no longer the only way in.

MCP (Model Context Protocol) is the standard way AI tools connect to business software. HubSpot's MCP server is what lets Claude or ChatGPT work in your CRM today. The three terms below get mixed up a lot, so this is how they relate.

API, MCP server and headless compared
Term What it is Who uses it What it means for you
API A technical way into the software Developers and integration tools Systems can sync data, but someone has to build the link
MCP server A standard way in, built for AI tools Claude, ChatGPT and other MCP clients Your AI can connect without custom code
Headless The choice to open every function through those ways in The whole business Almost everything can be done without opening the app

An API or an MCP server on its own doesn't make a product headless. Plenty of tools have an MCP server that reads a few fields and nothing else. A headless product lets software reach all of it, actions included, and every action follows the same rules as the screens.

You'll also see the term agentic CRM. It describes the same direction from another angle: a CRM built for AI agents to do part of the work. Headless is what makes that possible.

HubSpot isn't alone here. Salesforce announced Headless 360 on 15 April 2026, opening its platform through APIs, MCP tools and command-line commands so people and AI agents can work without the Salesforce browser interface.

What is an AI harness, and which part belongs in your software?

An AI harness is everything around a language model that makes it useful and safe in your business. A model can write and reason, but on its own it can't do work in your systems. The harness decides what data it sees, which actions it can take, what it may do without asking, what it remembers and who changed what.

The parts of an AI harness and where they sit
Part of the harness What it does Where it sits
Conversation and reasoning Understands what you mean Your AI tool (Claude, ChatGPT)
Memory Keeps what's needed between tasks Your AI tool
Context Decides which data the AI can see Your business software
Tools Decides which actions the AI can call Your business software
Permissions and rules Decides what it may do, for whom Your business software
Log Records what changed and who did it Your business software

Part of a harness sits in the AI tool you already use, and that's the part most people picture. The other part has to sit in the business software, and I think it matters more than people expect. Claude doesn't know your rates, your project types or which projects are closed. The platform that holds that data does, so that's the only place an action can be checked properly.

HubSpot says it will build what it calls the first Customer Relationship Harness: a conversational interface with your customer context already built in.

Why is HubSpot going headless now?

Three things came together in 2026. AI models got good enough to do real work across several steps, rather than only write text. MCP became the common plug between AI tools and business software. And the big CRM vendors moved: Salesforce launched Headless 360 in April, and HubSpot's refounding in November put headless and harness at the center of its plans.

HubSpot had already opened its CRM to Claude and ChatGPT through MCP and built its own AI agents. The refounding speeds that up. In the same Day 4 notes, Dharmesh wrote that a lot of HubSpot is already available through MCP and the API, and that the aim is to get much closer to all of it.

HubSpot has also said how it will charge for this. In a newsroom post on 17 November 2026, it said every seat includes enough headless access for everyday work from Claude or ChatGPT at no extra cost. Heavier use, and agents that need full governance and traceability, run on HubSpot Credits. You still pay your AI provider for the model itself.

Some companies are already a long way down this road. SaaStr founder Jason Lemkin wrote on 15 November 2026 that no one on his human team logs into their CRM anymore. Most firms won't work like that next year, but it shows where things are going. Software used to compete on who had the nicest screens. Now it competes on who gives AI the most useful context, the safest actions and the best controls.

Why does your AI only see half the picture after the deal?

HubSpot is built for the part before the deal is signed. In a service firm, the questions that decide your margin sit on the delivery side: who does the work, how many hours it takes and what you can invoice. If that side is spread over several tools, your AI can't see it properly and fills the gaps with guesses.

Take a firm that delivers work in projects, like an agency, a consultancy or a software company running implementations. HubSpot knows the deal. Everything after the deal usually lives in a project tool, a planning spreadsheet, a separate time tracker, the accounting package and a lot of email and Slack messages.

A hypothetical example. The owner of a 25-person agency on HubSpot asks Claude a normal managing director question: can we take on the Acme deal next month and still hit our margin? To answer, Claude has to reach six places. Some have no MCP connection, each one names clients and projects its own way, and nothing checks whether the plan matches the budget. Claude still answers, and it sounds confident, but part of that answer is a guess.

Put the same question to a firm that runs delivery in one platform connected to HubSpot. The deal, the planning, the hours, the budget and the invoices share one data model, so Claude needs two connections instead of six. Both return clean, linked data, and the answer is one the owner can act on.

Our guide on how to improve project profitability shows where margin leaks between those tools.

How do you get the rest of your stack ready?

Getting your stack ready for AI takes four steps, in this order. Find out where the work after the deal lives, bring it into one platform connected to HubSpot, make sure that platform is headless, and put a harness on top that knows your rules. Notice that the first two steps have nothing to do with AI.

  1. Find where the work after the deal lives: List every tool your team uses between Closed Won and the paid invoice, including spreadsheets and the inbox.
  2. Bring that work into one place connected to HubSpot: For a project-based firm this is a PSA, short for professional services automation: tasks, resource planning, hours, budgets, risk management and invoicing in one data model. Our list of the features a PSA should have helps you check the options.
  3. Make sure that platform is headless: Take a simple request: book 4 hours on the Acme project today. Headless means there's a book-time action that software can call without a screen, and that action follows the same hard rules as the screen. No hours on a closed project, and no hours on a project you're not part of. If an AI can write hours straight into a database field past those rules, that's a back door, not headless. Ask your vendor whether their MCP connection can read and act with the same rules as the screens.
  4. Put a harness on top that knows your rules: "The Acme project" might be three projects. Which task is it? Billable or not, at which rate, and is there budget left? The harness works that out, asks you what's missing, shows what it's about to do and waits for your yes. Afterwards it logs what changed and who did it.

If you're still choosing tools for step 2, our overview of HubSpot delivery stack integrations compares options for projects, time tracking and invoicing.

See steps 2 to 4 on your own setup. In a discovery call we look at how your work after the deal runs today and show what HubSpot and PSOhub look like with your own AI connected. Book a discovery call.

Built-in AI agents or bring your own AI: what's the trade-off?

Some vendors build their own AI agent into their app. Others open the app so you can bring your own AI, like Claude or ChatGPT, through MCP. Both look good in a demo. The difference shows up later: built-in agents usually cost more with every question, while your own AI is one subscription across all your tools.

Built-in AI agents vs bring your own AI
  Built-in AI agent Bring your own AI through MCP
How you pay Credits or tokens per question, summary or agent run, in each app One AI subscription, plus whatever each vendor charges for access
As use grows The bill grows with it, app by app The AI cost stays one line
Which model The vendor picks You pick, and you get each new model as it ships
Questions across tools Limited to that app's data One AI can work across every connected tool
Where the rules live In the app Should still be in the app, so check it

To be fair, your own AI isn't free either. A Claude or ChatGPT seat costs money. But it's one cost that works across your stack instead of a new meter in every app, and each new generation of models gets better and cheaper to run. That's why I think bringing your own is the smarter choice.

HubSpot's new pricing shows the two roads can live side by side. So for every tool in your stack, find out what's in your licence, what gets metered and what your AI subscription costs.

How do you judge a tool's AI readiness?

You can judge a tool's AI readiness on five levels. Level 0 is screens only. Levels 1 to 3 are headless in growing steps, ending with an AI that can read and act under the same rules as the screens. Level 4 adds a harness: help choosing the right action, a preview, your approval and a log.

The five levels of AI readiness
Level What it means What your AI can do
0. Screens only Human only Nothing
1. Open API Headless, first step Nothing directly; developers can connect things
2. MCP, read Headless, second step Answer questions, often from only part of the data
3. MCP, read and act Headless in full Create, change and book, through the same rules as the screens
4. Harness on top Harness Pick the right action, ask what's missing, preview, wait for your yes, log the change

 

What your AI can do at each level of AI readiness: read, act, follow the same rules as the screens, and preview before you say yes. Graphic: PSOhub.

More and more tools reach level 2, but many still give your AI only part of the data. Reading is easy; acting is where it counts. The best harnesses also understand roles, because a resource manager, a project manager and a finance lead ask different questions and need different actions.

Six questions to ask every vendor next to HubSpot

Questions 1 and 2 are about the road you take. Questions 3 and 4 test headless, and questions 5 and 6 test the harness.

  1. Can you bring your own AI, like Claude or ChatGPT, through an MCP server? Or are you stuck with the vendor's built-in agents?
  2. What's included in your licence, what gets metered in credits, and what do you pay your AI provider?
  3. Can your AI read all of your data, or only a few fields? And can it also take action, like creating, changing and booking things?
  4. When it takes action, does it follow the same rules as someone working in the screens, or can it slip in through a back door?
  5. Does the platform help your AI pick the right action, and ask what's missing when it isn't sure?
  6. Does it show a preview, wait for your yes, and log who changed what?

If a vendor can't give a clear answer on 3 to 6, their AI is a demo, not a colleague. It's AI bolted onto software that was built for screens. For a quick AI readiness assessment of your own firm, take the AI Readiness Check. It takes about four minutes and you get a report.

How does PSOhub answer the checklist?

PSOhub is a professional services automation platform for HubSpot users, with quotes, projects, resource planning, time tracking, risk management, budgets and invoicing in one place. On question 1 we chose the second road: you bring your own AI. Agentic AI Insights reads your data, and Agentic AI Actions prepares changes and waits for your approval.

How PSOhub answers the six vendor questions
Question How PSOhub answers it
1. Bring your own AI? Yes, through MCP with Claude, ChatGPT and other MCP tools. There's no AI agent built into PSOhub
2. Licence and metering PSOhub never bills tokens or credits. Actions is a flat add-on, so the bill doesn't change with use
3. Read and act? Insights reads projects, hours, planning and invoices. Actions creates and changes things like time entries, planning and invoices
4. Same rules as the screens? Yes. Each person connects with their own PSOhub login, so the AI only sees and does what their role allows
5. Help picking the right action? Yes, through the action layer described below
6. Preview, approval and log? Yes. Nothing changes until you say yes, and the change shows in the activity feed

 

When your AI wants to do something in PSOhub, like create an invoice or move someone's planning, it never writes straight into the database. It goes through our action layer.

  1. Know the setup: The AI first gets the fields, defaults and options of your organisation: your rates, your project types, your rules.
  2. Check first: Every proposed action is checked before anything changes. If something is missing, the AI hears what, along with the valid options. If something is wrong, it hears why.
  3. Preview: You see exactly what will happen, including where each value came from: a document, you, a default or a calculation.
  4. Confirm: Nothing changes until you say yes.
  5. Log: The change shows up in the activity feed, with who made it.

Think back to the car, where the car decides and the app doesn't. Your AI tool doesn't need to know how PSOhub works or what your rules are, because our layer does and applies them every time. That's what makes AI safe to use on real invoices and real people's planning. And since deals, projects, hours and invoices share one data model through the HubSpot integration, your AI can tell you which won deals actually made money.

The alerts inside PSOhub run on algorithms, not AI. They watch your data around the clock, behave the same way every time and cost nothing extra.

The trade-off: To use Agentic AI Actions you need a paid Claude or ChatGPT subscription. Most firms already have one. And if you want a chat window built into the app itself, we're not your vendor.

FAQ

Is HubSpot getting rid of its screens?

No. Human, the classic interface, is one of the three ways into HubSpot that Dharmesh Shah describes, next to headless and harness. Some work is simply better on a screen, like a resource planner moving people across weeks or a finance lead approving a batch of invoices. Good screens stay. They just stop being the bottleneck.

Do I need a developer to make my stack headless?

Usually not. Headless is something your vendors build, and your job is to choose tools that are headless and connect them. With MCP, connecting Claude or ChatGPT to a tool often comes down to adding a connector and signing in with your own account. Older tools that only offer an API may still need some technical help.

Is it safe to let AI change data in my systems?

That depends on where the rules sit. If the AI goes through the same checks as the screens, with your own login and role, a preview and your approval, every change is checked and logged before it happens. If it can write straight into the database, it's a back door. Questions 4 and 6 of the checklist test exactly this.

Do I have to replace the tools I already use next to HubSpot?

Not all of them. Keep the tools that give clear answers to questions 3 to 6 of the checklist. Look first at the spreadsheets and single-purpose tools that hold your delivery data, because that's where your AI is most likely to guess. Your accounting package can usually stay: a PSA like PSOhub prepares the invoice and posts it there.

Is this relevant for a firm with only 15 people?

Yes. In a small firm the owner is often the one asking the margin question and also the one who has to piece the answer together from several tools. Connecting one AI to two well-connected systems is a modest project at that size, and it means the numbers no longer live only in the owner's head.

What does HubSpot charge for using Claude or ChatGPT with its CRM?

HubSpot said on 17 November 2026 that every seat includes enough headless access for everyday work from Claude or ChatGPT at no extra cost. Heavier use and agents that need full governance run on HubSpot Credits, and you pay your AI provider separately. Check HubSpot's pricing page before you budget, because not every detail was published yet.

Run the checklist on your stack before you add more AI

The HubSpot refounding is good news for every HubSpot user, and it pays off fully when the rest of your stack follows. Start with the six questions on each tool next to HubSpot, or take the AI Readiness Check to see where your firm stands.

Want to see what this looks like with HubSpot, PSOhub and your own AI? Book a discovery call. If you'd rather look around on your own first, start a free trial.

PSOhub is a PSA platform for professional services firms that run on HubSpot. Quotes, projects, resource planning, time tracking, budgets and invoicing sit in one place, open to the AI your company already uses.