I ask an AI something I know it has and it just… doesn't go get it.
Right now I'm asking it for a growth strategy for our brand partnerships, and getting the least expected answer: I don’t know the name of a single brand you’ve worked with.
I thought Claude knew everything about me. Turns out that's not how it works.
AI access to your information is like an assistant that owns a key to your filing cabinet. That key doesn't mean they've read what's inside, or know which drawer to open. Access, searching, reading, interpreting — those are different steps, and any one of them can quietly fail. When an AI says "I don't know" or "I couldn't find it," that's a report on how the search went. It's not proof the thing isn't there.
Lucky me that my COO holds more context than any AI. And she actually knows which drawer to open.
Claude does know our brand deals. It proofreads a sponsor newsletter, or comes up with angles and hooks for an Instagram campaign. Almost every deal we've run has passed through it in some form. So when I asked about our partnerships, I assumed it would just pull that together. It didn't. It likely looked for "sponsorships" or "brands," but inside a project built around content, those words don't always sit where the deals actually live.
I rewrote the prompt. Added specifics, said how the answer should look, told it where to pull from. It came back with tighter formatting and a list of categories: B2B SaaS, fintech, HR tools. Still generic.
I asked for brand names. I got advice that could fit any creator in my niche.
This issue is brought to you by HubSpot.
The problem was never the prompt. It's the way the data is stored and reaches it.
So I opened HubSpot, the platform where our sponsorship data actually lives: terms, budgets, the whole record.
I asked the AI built into that platform to write the strategy there — Breeze, HubSpot's built-in assistant.
Here's the honest limit first: I can't show you the partnership answer. That data stays closed, for the same reason it's valuable. So let me show you the same idea on a pipeline I can open — how we book podcast guests. Same HubSpot, a different corner of it. Here a closed deal isn't money, it's a guest's signed filming date.
One pipeline, every stage — a closed deal here is a guest's signed filming date.
Every guest conversation, every stage, every deal sits in one place, and it updates itself the moment someone on the team moves something. I asked Breeze the same kind of question I'd asked Claude: based on what's actually in here, what should I do next?
Same question I gave Claude — answered from what's actually in the system today.
It told me exactly where the work was: it read our real stages: Reaching Out, In Conversation, Revisit Later, and told me what to do in each. It grouped my deals by stage and gave the exact next step for each one. The line that stuck: my pipeline is crowded early and thin right at the finish, so the move is to push the few deals one step from closing before chasing new ones.
A to-do list I could act on before lunch.
The gap wasn't intelligence. Both are the same kind of AI. The gap was the data underneath. Claude runs on rules it was given once, and when something changes we forget to go back and tell it, so its memory blurs. HubSpot sees the change the moment it happens. Breeze works off records that are true this morning, because they're the same records the team touched yesterday. Fresher data, more accurate answer. That's the whole thing.
Breeze is the ask-a-question side of this. HubSpot's Fall Spotlight is the whole set: agents that don't just answer but actually run the work on your real customers, deals, and conversations, instead of a summary you pasted into a chat window. That's the part I'm digging into next.
I still use Claude every day — for thinking out loud, for drafts, for a second opinion when I want range, not precision. But the moment I need a real decision made on what's true right now, I go where the data lives and ask the AI sitting on top of it.
Marina 💜
Thanks to HubSpot for sponsoring this newsletter.
