Encore AI's $30M: your call recordings are training data

Encore AI raised a $30 million Series A led by Team8 to do something more interesting than "AI answers your phone." It analyzes the conversations your best salesperson already has — calls, emails, texts, CRM notes — figures out which moves actually closed, and encodes those into agent playbooks. The pitch is that your call recordings are the training data, and that's the part small businesses should sit with, because you're already generating the asset and most likely throwing it away.

What actually happened

Per TechCrunch, Encore was founded in 2022 as Insait IO by CEO Dvir Ginzburg, originally selling recommendation software to financial advisers. It now runs what the company calls interaction mining across call recordings, email, text, and CRM data, then deploys agents that either work autonomously or sit alongside a human, suggesting the next line mid-conversation.

The round was led by Team8 with Planven, Lukatz, and The Garage participating — plus several banks and insurers that started as customers and then invested. Encore reports 40-plus enterprise customers, mostly financial institutions, and says ARR grew more than 5x since its seed under 18 months ago. Revenue and valuation are undisclosed. Ginzburg's framing: "we literally run by the playbooks that we see working."

One number to treat carefully. A company release cites a lending client hitting 10x ROI within months. That's a vendor-supplied figure from an unnamed customer. It's marketing until you can reproduce it.

Why your interaction data matters for your business

Every voice AI vendor is renting the same three or four frontier models. Assort Health raised $120M, Gradium raised $100M on latency — none of them have a model advantage. The differentiator is what the agent knows about how deals actually close in your business, and that only exists in your recordings.

So start capturing it now, independent of any purchase. Turn on recording where it's legal and disclosed. Run transcription — Whisper self-hosted or Deepgram's API, both cheap at small volume — and store transcripts as text in object storage you own. Then do the boring part that makes the data valuable: label the outcome. Booked, quoted, closed, lost, and why. A transcript with no outcome attached is a log file. A transcript with an outcome is a training example.

Do that for six months and you can hand a labeled corpus to any vendor, or fine-tune against it yourself. Skip it, and you'll be paying someone to mine data you never kept — and their model will walk when the contract ends.

Sources: TechCrunch, PR Newswire.