Encore AI

What it is

Encore AI builds customer-facing AI agents by copying your best people instead of asking you to write prompts. Its patented “Interaction Mining” reads a company’s historical calls, chats, emails and CRM records, splits each customer journey into stages, and isolates the moves that preceded a closed loan, a renewed policy or a recovered balance. Those moves become playbooks, and the playbooks run in two modes: Autopilot, an autonomous voice and chat agent that takes a journey from first contact to close, and Wingman, a live assistant that pushes discovery questions, objection responses and mandatory disclosures onto a human rep’s screen during the call. A third product, Journeys, replaces static lead forms with conversational landing pages.

The pitch is aimed at the revenue line, not the cost line. CEO Dvir Ginzburg put it directly in the Series A announcement: today’s enterprise AI industry is “optimized for cost reduction.” Encore’s use cases are framed as funnel stages — qualify, convert, onboard, upsell, recover — rather than as tickets to deflect.

The company was founded in 2022 as Insait IO by Ginzburg, who holds a Ph.D. in geometric deep learning from Tel Aviv University and previously worked on recommendation systems at Microsoft. Its first product recommended financial products to advisers and relationship managers, which explains the customer base. It announced a $30M Series A on 29 July 2026, led by Team8 — the company release names Planven and The Garage as co-leads — with Lukatz and several banks and insurers that were already customers. Calcalist names Israeli insurer Harel and Bank Leumi among those customer-investors and puts headcount at about 50 — roughly 35 in Israel, the rest in New York, where the company is headquartered, and Australia. TechCrunch reported more than 40 enterprise customers, mostly financial institutions; Calcalist’s figure was deployments at more than 30 organisations across Israel, Australia, Europe and the US.

Why it shows up in CX and revenue stacks

• The training data is your own outcomes, not a knowledge base. Most agent platforms start from documentation and prompts and then tune. Encore starts from interactions already labelled by what happened next in the CRM. If your top tenth of collectors recovers twice the balance of the median, the claim is that the platform can find out why and make the agent do it. The product pages say there is no prompt engineering, script or knowledge base to write, and that mining runs on every interaction with no sampling.
• Assist and autonomy share one playbook. Wingman and Autopilot run off the same mined playbooks, so a bank can put the assistant in front of human reps first, watch which playbook steps move conversion, and only then hand the easy segment to the autonomous agent. That staged path is the safer one in lending and insurance, where an autonomous agent making a wrong disclosure is a regulatory event, not a CSAT dip.
• It is delivered by forward-deployed engineers, with a handover. The deployment model is three phases: scoping and a production pilot on the use case with the clearest revenue upside, co-building integrations and agent customisation with your team, then testing, guardrail sign-off and “keys and handover.” Workflows move from “Run by Encore AI” to “Run by your team.” For a bank without an in-house conversational-AI team that is the realistic way to get to production; it also means the first months are a services engagement.
• The deployment and compliance options suit regulated buyers. The Autopilot and Wingman pages list on-premises, private cloud, VPC or fully managed Encore cloud, plus audit trails that log every agent decision, per-workflow guardrails for topic, tone and compliance, publish approvals and role-based permissions. The site claims ISO 27001, ISO 42001, SOC 2 Type 2, GDPR and PCI DSS, and points to a trust centre at trust.gainencore.ai.

Pricing reality

There is no public price, no rate card, no self-serve tier and no marketplace listing we could find; every source, including third-party directories, says “contact for pricing.” Expect an enterprise contract that starts with a paid pilot on a single funnel and carries a forward-deployed engineering component.

The useful anchor is what adjacent enterprise tools publish. Cresta’s AWS Marketplace listing prices Agent Assist for Voice at $150,000 for 12 months covering up to 100,000 calls, or $1.50 a call, and that is for assistance on calls a human still handles. Treat that as the order of magnitude a large regulated contact centre should expect for an assist-plus-agent platform in the same segment — a peer anchor, not an Encore quote. The only performance number Encore has published against cost is a “10x ROI within months” at one of its largest lending clients, unnamed and without a baseline.

Because the product is sold on revenue, push the commercial model toward revenue. Ask for pricing tied to the outcome you are buying — funded loans, recovered balances, completed applications — with the pilot fee credited against year one, rather than a seat or minute meter that pays the vendor the same whether conversion moves or not.

Best for

Heads of sales, collections or customer operations at banks, consumer lenders and insurers — the segment Encore was built in — who run a high-volume phone and chat funnel where outcomes are recorded in the CRM, and whose question is “how do we get every rep and every agent to sell, renew or recover like our top performers?” rather than “how do we answer more tickets with fewer people.” The strongest fit has years of recorded calls with consent, outcomes linked to each interaction, and an internal team too small to build agents from scratch.

Skip it if your goal is support deflection, if your interaction archive is thin or not linked to outcomes, if you need to evaluate the product without a sales cycle, or if you sit outside financial services and healthcare, where the reference base is shallow.

Versus the alternatives

The volume in this market sits with the CCaaS incumbents, NICE and Genesys, whose bundled AI covers agent assist and quality management inside a contract you already signed; start there if your problem is coverage, not conversion. Cresta is the closest like-for-like — assist and autonomous agents on the same models, with named connectors into NICE CXone, Genesys, Five9 and Amazon Connect — and is the pick for a voice-heavy contact centre of several hundred seats that needs a supervisor console across humans and AI. Sierra, the fastest-growing entrant at roughly $200M ARR by May 2026, and Decagon are the picks when the objective is end-to-end resolution of service conversations. Parloa wins when voice automation at scale on a predictable meter decides the deal, and Wonderful when you need agents in many non-English markets quickly. Level AI covers scoring every interaction for QA. For B2B sales teams working opportunities rather than consumer funnels, Gong is the revenue-intelligence default. Encore’s case against all of them is narrow: it is the one whose agents are built from your own top performers’ recorded behaviour and pointed at revenue events in a regulated consumer funnel. If that is not your problem, one of the others fits better.

Watch-outs

• The method only works on a large, outcome-linked archive. Interaction Mining needs recorded conversations joined to what happened afterwards; a call archive with no CRM outcome per call, or consent that does not cover model training, gives the platform nothing to learn from. Guard: before the pilot, audit one funnel — count recorded interactions from the last 12 months, check what share link to a CRM outcome, and get legal sign-off that existing recording consent covers training — and make “minimum data available” a written precondition of the pilot fee.
• Copying top performers can copy their worst habits. The rep who converts most may be the one who pushes hardest or skips a disclosure. Automating that behaviour in lending, collections or insurance sales turns one person’s risk into a systemic one, and outbound AI voice calls fall under the US TCPA rules on artificial voices. Guard: route every mined playbook through compliance review before publication, using the platform’s publish approvals; require the mandatory-disclosure checks to be hard guardrails, not suggestions; and sample autonomous conversations weekly against your existing conduct-risk QA.
• The headline results are vendor-published, and several references are investors. The 30% and 33% conversion lifts on the site and the “10x ROI” lender are all Encore’s own claims, and named customers Harel and Bank Leumi invested in the Series A. Guard: run a holdout test — a control group of the same funnel without the agent — scored on your own conversion definition, and ask for at least one reference customer in your vertical that is not a shareholder.
• This is a small company selling a services-heavy model. About 50 staff are carrying forward-deployed engineering across 30-plus enterprise deployments on three continents, and the public pages do not name specific CRM, telephony or CCaaS integrations. Guard: get the integration list for your exact stack in writing, name the forward-deployed engineers and their allocation in the SOW, and contract for export of the mined playbooks and agent configurations so the handover phase leaves you with something portable.
• Deployment claims vary by source. The vendor’s product pages list on-premises and VPC options, while at least one third-party directory lists the product as cloud-only. Guard: confirm the deployment topology, data residency and model hosting for your contract in the security questionnaire, and pull the actual SOC 2 Type 2 report and ISO certificates from the trust centre rather than relying on badges.