Beyond the demo
Most CRM demos show you the best case. This is not that.
Watch any AI-plus-CRM demo from the last twelve months and you see the same thing: a clean dataset, a perfectly phrased prompt, and a result that looks effortless. Then you try it on your own data, with your own workflows, and it does not quite hold up. That gap between the demo and the real brief is where most consultancy leaders feel quietly stuck rather than genuinely ahead.
We connected Claude to HubSpot using the new MCP integration, ran it against real-world sales tasks, and documented exactly what worked, what did not, and what a consultancy leader should do with it today. Here is the honest version.
The plumbing
What MCP is, and why it changes the economics
MCP stands for Model Context Protocol. Think of it as a universal connector standard — the USB-C of AI integrations. Before MCP existed, connecting Claude to your HubSpot data required a developer to build a custom API integration: weeks of work, typically several thousand pounds.
MCP changed that entirely. HubSpot built an MCP server. Claude supports MCP. Authorise the connection in Claude’s settings and the two talk directly through natural language — no code, no middleware, no developer time.
In practice, a Managing Partner who has never written a line of code can now query their entire pipeline, pull cross-object lookups, and generate reports through a chat window. The question is no longer “can we build it?” It is “should we trust it?”
What holds up
What works right now
These are the capabilities that earn their place in your week today:
- Pipeline queries. A prompt like “show me all deals owned by [name] closing in the next 30 days, sorted by amount” returns a formatted, sortable result in seconds — no filters, no saved views, no dashboards. For Monday pipeline reviews, board prep, and client portfolio summaries, this saves real time.
- Cross-object lookups. What used to take three or four clicks and a filtered view now takes one sentence: “show me all deals linked to [company], tell me what stage they are at and what the next step should be” resolves in a single prompt.
- Creating records conversationally. Now that write access is out of beta, you can describe a record in one sentence instead of navigating, creating a contact, logging a note, and setting a task as four separate actions. The caveat is significant — see below.
HubSpot has also published a free prompt library for the Claude connector, with over a dozen ready-made prompts across sales, marketing, and service workflows.
What does not
Where it falls short
The gaps are real, and two of them carry risk:
- No access to company notes. The single biggest gap. If your team stores context in HubSpot notes — meeting summaries, client preferences, project history — the MCP server cannot read it. For any consultancy where relationship context lives in those notes, this is a material limitation.
- Write operations carry hallucination risk. HubSpot’s own documentation warns that large language models are prone to hallucinations and advises reviewing every write action before confirming it, and testing write and delete permissions in a developer sandbox before production. Write operations are usable, but every output needs a human review before it touches your live CRM. This is not set-and-forget, at least not yet.
- Documentation thins out fast. Standard setup works cleanly, but custom scopes, advanced authentication, and multi-user deployment outrun the docs. One community member built their own extended MCP server with 106 custom tools because the official one did not cover enough ground. This is beta software behaving like beta software.
The verdict
Should your consultancy use it?
Use it now for read operations — they are solid. Pipeline queries, cross-object lookups, and portfolio summaries all work reliably and save time on the tasks your senior people do every week.
Use write operations cautiously. If your team already operates with a trust-but-verify mindset on AI outputs — the correct posture at this stage — write features are usable in production with human review in place. If the expectation is automation without checking, this integration is not ready for that.
Test in a sandbox first. HubSpot provides a developer sandbox; it costs nothing to explore write operations there before touching live data. And watch the change log — HubSpot is shipping updates monthly, so what is a limitation today may be resolved in the next release.
The bigger shift
Why this matters, and what governance it demands
279,000 businesses use HubSpot. The vast majority have never built an API integration because it requires developer capacity they do not have. MCP removes that barrier entirely.
This is the KWA pattern in practice. The integration gives you better access to your firm’s data (Knowledge layer), makes existing CRM workflows faster to execute (Workflow layer), and begins to create the conditions for agentic automation to do real work in the background (Agent layer). Read capabilities sit firmly in the Workflow layer now; write operations, once they mature, will push into Agent territory.
No hype, no platform lock-in — just measurable impact on margin and capacity, but only if the governance is in place to use it well.
The bottleneck in AI adoption for professional services has never been technical complexity. It has been the vendor noise, the overblown demos, and the integrations that work in a controlled environment but not on a real client brief. This one is different because it is genuinely useful today and honest about what it cannot do yet. The work for you is the Govern step in the 5 Steps for AI Leadership framework: knowing which tools are approved, what write operations require human sign-off, and how your team is trained to verify AI outputs before they touch live data. Without that, write access to your CRM is a liability, not an asset.