TL;DR. Before any AI tool, build the foundational data layer. Record every customer call, structure the transcripts so they're queryable, layer your ICP, persona, and messaging docs on top. Once this exists, every other GTM-AI use case becomes 10x more powerful. Most teams skip this step and wonder why their AI projects stall.
A lot of AI use cases get pitched at GTM teams right now: outbound automation, content production, lead scoring, agentic SDRs. They all sound impressive in the demo.
Most of them are downstream of one decision the team usually hasn't made yet. Until you make that decision, every AI investment hits a ceiling within a few weeks. The demo looks magical, the team uses it for a month, and then it quietly fades.
So before any tool, there is one thing you have to do.
Start with the foundation
Record every customer call and transcribe it. Put it in one structured place your AI can read.
That's your foundation. It's not a single tool you buy, it's a layer that sit underneath every AI workflow you will ever run. Once this is in place, the same workflows that used to feel generic and disconnected start to feel like they actually understand your business. The reason why customer calls are so foundational in your AI efforts is because they are the one place where key input for your GTM efforts live: customer feedback, objections, insights on your client journey, messaging do’s and don’t’s. Everything important to building your knowledge on how to run GTM.
We will run through how to set up this layer below
Why most AI projects in GTM stall
The reason most AI projects in GTM teams stall is simple: the AI has no idea who you are. Every new chat session starts blank. Someone on your team re-types "we're [Company], we sell [thing] to [persona], our ICP is..." and that context loses to fatigue by the third session. The AI never sees a stable picture of the business, so it answers like a stranger every time.
A foundational data layer fixes this. It's the brain the AI consults at the start of every task. It's what lets you stop typing your context every time and start asking real questions:
"What did customers say about pricing in the last 30 days?"
"Which deals stalled around the integration question?"
"Which phrases do customers actually use to describe the problem we solve?"
Capture the knowledge where it surfaces the most
Two things are important to build, and they are easy, not hard.
Layer 1: Capture every call
Pick a recording tool: Mojo, Granola, Otter, Fathom, Demodesk, tl;dv. The specific tool matters less than committing to the rule. Every customer call gets recorded with permission, and every transcript lands in one location: a Google Drive folder, a Supabase table, an Airtable. Pick one and commit.
The goal of this layer is to stop losing what was said in the room. Most teams accumulate hundreds of hours of customer conversation a year and retain none of it as queryable data. In the age of AI, this has to stop.
Layer 2: Structure the calls so they're queryable
Once recordings are flowing, build the metadata around them. Tag each call with account, persona, deal stage, and topic. Add your ICP, persona definitions, messaging guidelines, and recent product context as separate documents the AI can reference at the start of every chat.
The goal of this layer is to give the AI everything it needs to know about your business in a structured form, so it stops asking you the same context questions over and over. A few hours of structure work here compounds into months of saved context-typing later.
This is the layer most teams skip. They record calls but never structure them. Then their AI can read one call but can't answer "what's changed in customer language over the last month?"
Most teams skip the foundation, and that's the whole problem
Most teams skip building this foundation because it doesn't feel important enough. It is the most important layer. The compounding starts the moment you commit to it, and the gap between teams that did this and teams that didn't widens every week.
Continue layering structure and interpretation on top of capture as the call volume grows, and within two months you'll be running queries against your own customer conversations that your competitors cannot match.
→ This week's how-to is one tactical example of what becomes possible once the foundation exists: how to ensure your Claude posts don't sound like AI slop.
FAQ
Why is recording calls the first step, not the last?
Because every other GTM-AI use case depends on the AI knowing who your customers are, what they care about, and how they talk. Without that grounding, the AI is an outsider every time you open a chat. Recordings, structure, and interpretation give it the context one time, so every downstream use case starts from a stronger base.
Do I need a paid call-recording tool, or can I use Zoom transcripts?
Either works to start. The specific tool matters less than the rule that every call gets recorded. Mojo, Granola, Fathom, Otter, and Read are all reasonable. Zoom transcripts work too if you have a process to extract and store them. Pick one and commit by next week, the question of "which is best" is a distraction.
What do I do if my prospects refuse to be recorded?
Honour the refusal, both for the relationship and for compliance. The system still works on the calls you do record. You don't need 100% coverage to get value, you need a representative cross-section of your pipeline. A team that records 70% of its calls is years ahead of one that records 0%.
How long does it take before this pays off?
The capture layer pays off immediately, you stop losing what was said. The structure layer starts paying off after about four weeks, enough data to spot patterns. The interpretation layer compounds from month two onward, the more calls flow in, the better its answers get. Treat it like an asset, not a project.
