How do I track competitor mentions across all calls?

Oh, you're asking all the right questions. I've audited three of these systems for my team. The "how it works" part is usually a two-stage process: a broad keyword scan and then a smaller contextual model trying to filter out casual mentions from competitive deal conversations. But you're right to be skeptical.

> Does it just flag the word "Salesforce"
Sometimes, yes, that's literally it. I've seen tools flag "Oh, our Salesforce instance is down" as a competitor mention. The false positive rate without tuning is embarrassingly high. The ones I trust let me build a custom exclusion list for internal references, and they show me a confidence score for each flag. You have to ask them directly about the SLA for accuracy. Most won't give you a number, which is a huge red flag for compliance.

And the data storage question is crucial. If they're using a third-party ASR provider, your call data might be stored in two additional places you never agreed to. I always ask for a data flow diagram now. It's the only way to see if it's truly siloed.

That's a really good point about the confidence scores. I was just looking at a demo for one of these tools, and they wouldn't show me any of the "maybe" flags, just the high-confidence ones. It felt like they were hiding the messy middle where most of the mistakes probably happen.

The data flow diagram idea is smart, and scary. If they're using a generic speech-to-text service, how can they promise it's not being used to train other models? I'm definitely adding that to my list of questions for vendors.

You've nailed the exact fear that made me roll my own solution for my team last year. The "false positive" problem is so real. I had a rep discussing a case study saying "we replaced Salesforce," and the system tagged it as a competitor mention, which looked terrible in the deal report.

My advice? Demand to see the audit log for a test call. Upload a recording where you say "Yeah, our internal Salesforce CRM is a pain" and another where you say "We're evaluating Salesforce against your solution." A good system will show you the difference in confidence scores and let you create a rule to mute mentions with "our" or "internal" before the competitor name. A bad one will flag them both as identical high-risk events.

Without that granular control, you're just building a list of every time the word is said, which is useless noise.

You're asking the right questions from the start, which I appreciate. I've been in those compliance reviews where a VP is holding up a report asking why we flagged 200 "competitor" mentions last quarter when half were us complaining about our own tools.

The "how it actually works" part is sadly opaque with most vendors. In my experience, it's usually a basic keyword match first, and the contextual layer is just a filter for common stop phrases. The real issue is that data storage question. If they're using a third-party speech-to-text engine, your client's call audio might be touching systems completely outside the vendor's own SOC 2 scope. You have to ask for the data flow diagram specifically for the audio processing pipeline, not just the application layer.

And you're spot-on about the SLA. Most won't commit to a specific accuracy percentage because they can't. I once got a vendor to admit their "95% accuracy" claim was for speech-to-text transcription clarity, not for competitor identification accuracy. That's a crucial distinction for your contracts.

Your best leverage is to demand a pilot using your own recorded test calls with scripted edge cases. If they won't support that level of validation, walk away.

Your skepticism about the black box is the correct starting point. The false positive issue you mentioned is typically rooted in the initial transcription layer. Most vendors don't own their own speech-to-text engine, so the raw transcript often comes from a generic service like Google or AWS that has no concept of your business context. The "competitor detection" is just a secondary script running on that messy transcript.

For granular controls, you need to look for platforms that treat the competitor list as a rules engine, not just a keyword list. The better ones allow you to set proximity rules, like ignoring "Salesforce" if the words "our" or "internal" appear within three words prior. Few offer an SLA for accuracy because they can't control the upstream transcription errors.

The data storage question is critical. If they use a third-party STT, your audio is processed in that vendor's environment, which may have different data retention and training policies. You must ask for attestation that no audio or transcript data is retained by the speech-to-text provider for model training, ever. Most sales demos gloss over this pipeline entirely.

You're right to zero in on the false positive rate. I've seen a deal review get delayed because a rep said "their platform is no Oracle" as a compliment, and the system logged it as a competitor threat. The confidence scores are everything.

On the data storage point, you need to ask if the audio is transcribed in your own cloud instance or sent to a third-party vendor's general model. Most won't volunteer that their "AI" is just a wrapper around a generic speech-to-text API where your data is absolutely part of the training pool unless you pay extra for a private model.

For granular controls, the only system I've trusted lets me build exclusion patterns, not just word lists. Like, ignore "X" if it's preceded by "like" or "not" within the same sentence. Without that, you're just building a report full of noise.

You've put your finger on the exact worry that stops a lot of us from trusting these systems. The "black box" problem is real.

When you ask about granular controls, you're hitting on the right requirement. The best answer I've gotten from a vendor is that you can define a competitor list and set up exclusion patterns, like ignoring "Salesforce" if "our" appears within three words before it. But you're right to push for the SLA on accuracy - most won't give you one because they can't fully control the upstream transcription service.

That data storage question is the real deal-breaker for many clients. If they're using a generic speech-to-text API, your call data is likely part of a general training pool unless you pay a hefty premium for a private model. Always ask for the specific data flow diagram for audio, not just the app.