Encore AI’s bet on customer calls is a smart one, and a risky one

Encore AI just raised $30 million in Series A funding led by Team8 to push a simple idea: the best training data for customer-facing AI agents is already sitting inside a company’s own calls, emails, texts, and CRM records.

That’s a strong thesis. It’s also hard to do well.

The startup, founded in 2022 as Insait IO by CEO Dvir Ginzburg, started with recommendation software for financial advisers and relationship managers. It has since rebranded as Encore AI and turned that work into what it calls “interaction mining,” a system that studies customer conversations, pulls out what works, and uses those patterns to train voice and text agents.

The pitch is straightforward. If your best sales rep, support agent, or relationship manager has a playbook that closes deals or resolves issues, teach the AI from that instead of from generic prompts and canned FAQ content.

Training on company behavior

Encore’s approach sits in a more useful place than the usual “AI agent for customer support” pitch. Plenty of vendors are wrapping LLMs around a knowledge base and stopping there. That can handle simple triage, but it often breaks down when nuance matters, especially in regulated or relationship-heavy businesses like financial services.

Encore says its platform ingests call recordings, emails, text messages, and CRM data. It then breaks conversations into stages and tries to figure out which exchanges moved a deal, a case, or a support interaction forward. So the model isn’t just learning what was said. It’s learning when it was said, by whom, and in what sequence.

That matters.

In a sales cycle or customer-success workflow, timing can matter as much as content. A well-placed clarification early on may matter more than a polished closing line later. If Encore can isolate those patterns reliably, it can train agents that behave more like experienced staff and less like generic chatbots with company jargon.

Ginzburg described the output as a set of successful playbooks. He said the agents sometimes even mirror the jokes, anecdotes, and examples used by human relationship managers because those pieces show up in the interaction data as working patterns.

That’s interesting. It’s also a little unnerving, which is probably the right reaction.

A system trained on live customer interactions can capture real nuance. It can also absorb bad habits if the scoring and segmentation are sloppy. If a company’s best-performing rep is also the one who bends policy, the model may learn the trick, not the rulebook. That’s a different failure mode from standard retrieval-based AI, where the usual problem is hallucination or stale content.

What interaction mining actually means

The phrase sounds cleaner than the underlying work. What Encore is doing is a mix of conversation analytics, outcome attribution, and supervised training data generation.

The hard part isn’t storing calls. Plenty of systems do that. The hard part is mapping unstructured interactions to outcomes. A successful sale, an escalated ticket, a churned account, a policy exception. Those labels often live in fragmented CRM fields or never get recorded cleanly at all.

If Encore can connect transcripts to CRM state and later business results, it can start identifying features that correlate with success. That gives it a path to training data that’s much richer than a static FAQ or product manual. It also gives customers a way to inspect their own process debt. Which stages are stalling? Which handoffs are failing? Which phrasing tends to derail a conversation?

That’s useful for teams that already know their process is messy and want more than a dashboard.

There’s a catch, though. Correlation is easy to overstate. If one rep closes more deals, that doesn’t always mean their wording caused the close. They may get better-qualified leads. They may work a stronger territory. They may just be more senior. If the system doesn’t control for those confounders, it can mistake coincidence for best practice.

For engineers, that’s the real challenge here: turning noisy, multi-modal interaction data into training signals stable enough to automate against.

Why financial services is the obvious first market

Encore says it has more than 40 enterprise customers globally, most of them financial institutions. That makes sense. Banks, insurers, and wealth managers care about consistency, auditability, and making every interaction count.

They also have something smaller startups rarely get: lots of recorded customer conversations, plus compliance pressure to keep them.

That combination is good fuel for a system like this. It gives the model enough data to learn from and gives the buyer a clear reason to care about process quality. A small improvement in conversion, retention, or call resolution can justify a serious software budget.

The same vertical focus also shows where the friction will be.

Financial institutions are cautious about data movement, model behavior, and human oversight. A tool that listens to customer calls, reads emails and texts, and then proposes or generates responses has to fit into a much tighter security and compliance envelope than a typical SaaS deployment. Access controls, retention policies, PII handling, audit logs, model isolation, and approval workflows all matter. A slick demo won’t get far without them.

If Encore is winning these customers, it’s probably doing more than selling AI. It’s solving the boring plumbing around ingestion, governance, and deployment. That’s the part most startups underestimate.

A real edge, but not a permanent one

Encore’s early advantage comes from owning the data-to-agent loop. It can look at conversation history, infer what worked, and feed that back into the agent layer. That gives it a feedback loop generic LLM wrappers don’t have.

The moat is narrower than it sounds.

Large CRM vendors like Salesforce, HubSpot, Zoho, and SAP already sit on huge piles of customer data. They also control the systems of record where outcomes live. If they decide conversational history should be a first-class training signal, they can move quickly. They have distribution, integrations, and enterprise trust.

Ginzburg’s argument is that data alone won’t be enough for them, because they’d have to rework the implementation stack to make conversation history the foundation for their agents. That’s probably true. It’s also only partly reassuring.

Big vendors don’t need to copy Encore’s exact architecture to make the pitch harder. They can bundle voice intelligence, call summarization, agent assist, and outcome-linked coaching into products customers already buy. That’s often how platform competitors squeeze startups: not by matching them feature for feature, but by making the adjacent feature good enough.

Encore’s defense is specialization. It’s building around specific workflows, especially in financial services, where generic automation tends to produce generic results. That can work. But it only lasts if the company keeps shipping deeper process intelligence faster than the platforms can catch up.

The agent part deserves scrutiny

Encore says its agents can work alongside employees or operate autonomously, and can communicate by voice or text. That covers a lot of ground, from recommendation surfaces to customer-facing automation.

The autonomous part is where things get tricky.

Voice agents are still unforgiving. Latency, turn-taking, escalation logic, identity verification, and policy constraints all have to work in real time. If the model stalls or improvises badly, the failure is immediate and audible. Text agents get a little more slack, but only a little. In financial services, one wrong recommendation can create support headaches or regulatory exposure.

That means the quality bar is much higher than “sounds human.” The agent has to know when to stop talking, when to hand off, and when to refuse.

If Encore is using mined interaction data to train routing logic and response selection, that’s a meaningful technical asset. If it’s mostly using the data to make the model sound like the company’s employees, the value is thinner. Tone matters less than accuracy, policy compliance, and repeatable outcomes.

And yes, the joke examples are memorable. They’re also the least important part of the product.

What the money buys

Encore says it will use the Series A to expand U.S. sales operations and deploy the platform with more large financial institutions. That’s the right move if the product is already getting pull from enterprise buyers. This market is still early, but enterprise AI in regulated sectors tends to reward patience and punish overreach.

The round included Planven, Lukatz, and Garage, plus some banks and insurers. Some of those financial institutions reportedly used the product first, then decided to invest. That’s a strong signal. It suggests the product got past the demo stage and into actual workflows, which is where a lot of AI startup claims fall apart.

Encore’s revenue story is moving in the right direction too. Ginzburg says ARR has grown more than 5x since the seed round less than 18 months ago, though he didn’t disclose exact numbers. That’s not enough to judge scale, but it does suggest real customer pull rather than one-off experimentation.

The larger question is whether the company can keep its training loop defensible as the market matures. The answer probably comes down to three things: the quality of its outcome attribution, the depth of its compliance and security stack, and whether it can keep turning interaction history into something that actually changes agent behavior.

That’s hard to fake.

And in enterprise AI, hard is often where durable businesses start.