Claap and Claude Code for $7M internal operations

We run the operations of a $7M ARR agency on two things: Claap and Claude Code.

The part people expect me to say is that Claude reads our call transcripts. It does not. It never opens one unless someone asks for a specific quote.

That distinction is the whole reason this works. Across roughly 300 clients, the bottleneck was never recording the calls. Recording is solved. The bottleneck was getting what happened on those calls into a shape something could query.

Here is how the system is built and what it produces.

1. A dataset over a transcript

Claap tags every call as it lands. Objections, pain points, competitor mentions, feature requests, all against standard categories that stay the same from one call to the next.

That labelling step is what turns a pile of recordings into something queryable. When Claude asks for every objection raised this quarter, it gets a set of labelled records back. It is not summarising 200 conversations and hoping the summary held.

The difference shows up in three places.

→ The same question asked twice returns the same answer

→ Q1 and Q3 can be compared, because both were labelled the same way

→ Nothing depends on how well a model compressed an hour of audio

Ask an AI to read 200 transcripts and tell you what prospects push back on, and you get a plausible paragraph. Ask it to count labelled objections across a quarter, and you get a number you can act on. The second one survives contact with a leadership meeting.

Everything below runs on that foundation.

2. Capture without a bot

A dataset with holes is not a dataset. If half the calls never get recorded, every query underneath is wrong in a way nobody notices.

So capture covers everything. Google Meet, Zoom, and Microsoft Teams for video. Phone calls come in through lemlist or Aircall depending on what the client already runs, which matters more than it sounds. Asking a client to switch dialers to make your reporting work is a losing conversation.

The piece that surprised our team most is that none of this requires a bot joining the call. Claap's desktop capture, currently in beta, records from the machine instead of sending a participant into the meeting. There is no visible attendee and no "recording bot has joined" moment for a prospect to react to.

Phone coverage only helps if you know who is worth calling in the first place, which is its own problem upstream of any of this.

You can pull the decision-makers at your target accounts, with their contact details, for free:

People Finder Tool

3. Objections in one query

This is the query that changed how we run our own pipeline reviews.

One question returns every objection raised across the quarter, already grouped. Not a sample, not the ones a manager happened to sit in on, and not the ones reps chose to log in the CRM. All of them.

The value is that the ranking comes from the calls rather than from whoever spoke loudest in the last pipeline review. If price is the blocker, the count says so. If it is timing, or an approval step nobody mapped, the count says that instead. Pain points and feature requests come back the same way, which is how product hears the same signal sales hears, in the same week, without a meeting to reconcile the two.

The output feeds straight back into outbound. Objection language is customer language, and customer language outperforms anything written from a positioning doc.

The sharpest cold email copy we write comes from lines prospects said on calls, not from a brainstorm.

You can rewrite your own copy against what your buyers respond to, for free:

Email Copy Optimizer Tool

4. Client reviews that publish themselves

Client reviews used to eat an afternoon per account. Someone opened six recordings, scrubbed through them, cross-checked the email thread, and wrote the thing from memory.

Now Claude pulls a client's calls and the emails sitting on the same deal timeline, drafts the review against both, and publishes it back into Claap as a shareable link. The review lives where the calls live, so the next person to open the account gets the summary and the source material in the same place.

Two things came out of that beyond time saved. Client sentiment stopped being a gut read, because the review is built from what was said rather than what the account owner remembers. And the reviews became training material, since a new hire can read six months of a real account and see how the relationship was handled at each stage.

5. Handoffs nobody has to chase

The handoff from sales to delivery is where agency context goes to die. The discovery call had the diagnosis, the close call had the promises, and the email thread had the caveats. Delivery inherits a CRM note.

The system now assembles the internal brief before delivery asks for it. Discovery call, close call, and the email thread go in. What comes out is the diagnosis, what was committed to, and the constraints nobody wrote down.

The uncomfortable part of building this was seeing how much had been getting lost. A promise made on a close call that delivery never heard about is the kind of gap that turns a good first month into a difficult second one.

6. Coaching against any framework

Call scoring is not new. Scoring every call is.

Claude scores calls against MEDDIC, SPICED, or whatever framework a team already uses, and the frameworks are interchangeable because the scoring runs on labelled call data rather than a bespoke prompt per methodology.

The value is not the individual score. It is the trend across the team. When the same gap shows up in 40 calls across six reps, that is a training problem, and it gets fixed once. Reviewing one rep at a time, that same gap reads as six separate coaching notes and never gets named.

Managers also stop grading the three calls they had time to listen to, which were never a representative sample of anything.

7. Two minutes to connect

The setup is one line in the terminal and an OAuth click. Claap runs a remote MCP server at api.claap.io/mcp, so connecting it to Claude Code means pointing at that endpoint and signing in. There is no API key to generate and no integration work to schedule.

Permissions carry over from Claap. Whatever a person cannot see in Claap, their agent cannot see either, which is the answer to the question every ops lead asks first.

From there the tools register themselves and the queries above work immediately. Our whole rollout took about two minutes, which felt disproportionate to what it replaced.

8. Structure beats prompting

None of this is a claim that AI got better at reading transcripts. The models were already good enough. What changed is that the calls stopped arriving as unstructured audio and started arriving as labelled records.

Once that is true, the queries stop being clever and start being boring. Every objection this quarter. Every deal where a specific competitor came up. Every call where the champion never got named. Boring queries against clean data beat impressive prompts against messy data every time.

The teams that get the least out of AI in sales are the ones pointing it at raw recordings and asking for insight. The teams that get the most fixed the shape of the data first.

You can see how your full GTM motion performs across content, outbound, and conversion, for free:

GTM Reports Tool