By
At the July 9 Orange County HubSpot User Group, held online, Eli Portnoy, founder and CEO of BackEngine.ai, tackled a question most teams are living with: why does AI turn into a superpower at some companies and into chaos at most others? Eli has spent about 15 years in data science and AI, and he walked through how AI tools like Claude actually use your data, then shared three workflows you can build with the data you already have.
Part of the Orange County HubSpot User Group series. See the recording and recap from every session.
Watch the full session
Missed the meeting, or want to rewatch the workflows? Here's the full recording from July 9 (about 50 minutes).
Video not loading? Watch the recording here.
Why AI rollouts turn into chaos
Eli has watched dozens of AI deployments, and he says they go one of two ways. In a few companies, AI becomes a force multiplier: productivity, quality and sales go up, and cost goes down. In most, usage is high but the output is inconsistent, and the team ends up producing more mediocre work, faster.
His explanation: large language models are trained on nearly all the data in the world except yours. They know nothing about your customers, deals or processes. He compared it to hiring 500 brilliant graduates and never onboarding them. They'd be eager to help and would mostly create chaos. Ask Claude "Which committed deals this quarter are going sideways?" without your data, and the best it can offer is generic advice on evaluating deals. With your data, it can do the work for you.
Three ways to give AI your company's context
- Paste it in every time. Add the background, transcripts and documents to each prompt. It works and it's the easiest way to start, but it doesn't scale, and results vary wildly from person to person.
- Connect your apps through MCP. Connectors for HubSpot, Gmail, Fathom and others give Claude or ChatGPT direct access in a few clicks. The catch is that they're "raw pipes." Eli compared it to being dropped into a library with no catalog and asked to find the best murder mystery involving aliens: the AI samples what it can and guesses. In BackEngine's own testing, about 23% of answers from direct connectors contained a substantial error.
- Use a knowledge graph. A layer between your tools and the AI that reads, joins and indexes everything (emails, tickets, calls, CRM records) so the AI gets exactly what it needs, like a catalog plus a librarian. BackEngine reported the error rate falling to about 7.6% this way, with token costs cut by roughly $1,300 per person per year.
You can build a graph yourself (for example, code that pulls from your tools daily and exposes an MCP connection) or buy one, which is what BackEngine does. Eli also argued for keeping your data separate from your AI tools, because the leading platform today may not lead in three months, and a separate data layer lets you switch without starting over.
What this means for HubSpot
- Clean data matters more than ever. The richer and better structured your HubSpot data is, the better Claude can use it through the HubSpot connector.
- HubSpot stays the source of truth. A knowledge graph doesn't replace the CRM; it's an index built for the AI on top of it.
- Test how the AI got its answer. Connect Claude to HubSpot, ask it a question, then ask how it answered it and what it looked at. AI is built to give answers we like, so make it show its work.
- Be careful with write-back. Claude can write to HubSpot through MCP, such as scoring prospects weekly into a field. Set permissions deliberately. John has seen a client import 2,000 contacts with no email or phone number.
Three workflows to try
1. A live account dashboard
Ask Claude to build a live artifact that becomes your Monday morning command center: every active account as a row, with tier, ARR, signal count and a letter grade Claude assigns from everything it can see. Because it's live, the data refreshes each time you open it, and you can click into any account for the detail behind the grade.
2. An automatic weekly status update
Set up a scheduled task that runs every Friday at 4 PM, reviews your email, calendar, Slack, Notion and Drive, and sends a 50-word summary of what you moved forward plus what you checked off. Save the reports to Notion and you can later ask Claude what you focused on over the last three months.
3. One-on-ones that aren't status updates
Have Claude track each person's progress against their goals, pulling from email, call transcripts and notes. When both people walk in already knowing where things stand, the meeting can focus on blockers and strategy.
BackEngine shares these prompts and about 100 more in its free prompt library, including HubSpot-specific ones.
Q&A highlights
What's the difference between a prompt and a workflow? A prompt is the instruction you give the AI. A workflow chains steps together and runs on a schedule, like an 8 AM email with a prep guide for each of today's meetings.
Do you need a knowledge graph for these workflows? No. They work with direct connections to HubSpot and your other tools. In Eli's experience, the answers are just less accurate and cost more to produce.
Why are my results with Breeze mediocre? Eli's view was that it's usually the tool, not your prompting, and that the best way to learn Claude or ChatGPT is to experiment and ask the AI itself when you get stuck. John added where Breeze does shine: configuration. He uses it to add properties and adjust pipelines, and logs each change as a task on a HubSpot project record, which gives him a running history of every change to the portal.
A tip for Starter editions: if you can't build custom reports in HubSpot Starter, try building them in Claude.
Join the next OC HUG
The Orange County HubSpot User Group meets monthly, alternating virtual sessions with in-person meetings in Irvine, with guest speakers from HubSpot and its partners. The August meeting covered where to start with HubSpot AI, credits, AEO and Loop Marketing; read the August recap. Join the group on HubSpot User Groups to get an invite to every meeting, browse all past sessions, or get in touch with John Hulina to suggest a topic.
