Matthias Wickenburg

The through-line

Wickenburg's consistent argument is that in an AI product built on conversations, everything rests on the first step. Attention's features sit on top of call recording and transcription, and he frames the whole stack as a dependency chain: "If the initial transcription quality isn't strong, everything else after that suffers" [1]. That conviction drives a set of practical positions, on continuous vendor evaluation, on when not to self-host, and on treating an infrastructure choice as something customers feel directly in the sales cycle [1].

On transcription as the foundation of the product

He describes Attention as "a combination of insights and workflows, all downstream from call recording and transcription," spanning CRM autofill, coaching scorecards, and generalized insights across hundreds of thousands of conversations [1]. Because those layers are all derivative, errors do not stay contained. A single mistake on the wrong word propagates: "If the transcription messes up a certain keyword or a certain competitor name, the insights are not going to be as strong" [1]. This is why he treats provider selection as a recurring engineering discipline rather than a one-time procurement decision, running a comprehensive benchmark every three months to internally evaluate which models are strongest [1].

On build versus buy, and the real cost of self-hosting

Wickenburg is candid that Attention went down the same road as most technical teams and reconsidered. The appeal was obvious: "we can just plug in Whisper or a similar model, press play and suddenly all our customers will have access to cheap self-hosted transcription" [1]. His verdict is that the model is the easy part and the operations are not. "There's a tremendous amount of overhead in terms of being able to deliver transcription at scale in production to tens of thousands of users simultaneously in a way that's very difficult with a heavyweight model like Whisper" [1]. The decision to work with Gladia turned on accuracy, diarization, and what he calls world-class customer support [1].

On multilingual as an under-served requirement

International customers are a specific pressure point for him, and one he thinks the market handles badly. Attention's users routinely "cycle between English, French, and Spanish on the same conversation," which makes dynamic language detection valuable rather than optional [1]. He singles out support for international languages as "an area that other transcription providers often overlook quite a bit" [1].

On infrastructure quality as a commercial lever

He makes an unusually direct link between a backend dependency and revenue. Transcription is "one of the first things that our clients notice whenever they do a pilot or a PC," so quality shows up in evaluations before any downstream feature gets a fair hearing [1]. And the effect persists past signature: because so many features depend on it, transcription remains an ongoing concern, and a dependable source has helped Attention "both close more deals and maintain clients after the fact" [1].

Takeaways

• Treat transcription as the load-bearing layer of a conversation-AI product, since CRM autofill, coaching scorecards and cross-conversation insights all sit downstream of it [1].
• Re-benchmark transcription providers on a fixed cadence; Attention runs a comprehensive internal evaluation every three months [1].
• Self-hosting an open-source model like Whisper is cheap in theory and expensive in practice once you need production scale for tens of thousands of simultaneous users [1].
• Evaluate providers on accuracy, diarization and support quality together, not accuracy alone [1].
• Misrecognised keywords and competitor names degrade every insight built on top of them, so error tolerance should be judged by downstream impact [1].
• Dynamic language detection matters for customers who switch between English, French and Spanish mid-call, and international language support is commonly neglected by vendors [1].
• Transcription quality is what prospects notice first in a pilot, making it a factor in closing deals as well as retaining accounts [1].