Omilia

What it is

Omilia is a voice-first conversational AI platform for large, regulated contact centers — the bank, insurer, or government agency that takes a million calls a day and has to pass a security review before anything answers them. Dimitris Vassos founded it in 2002, in Cyprus, as a contact-center integrator; it now runs from Athens and is opening its first U.S. office. Unlike most AI-native voice vendors, it owns the whole speech path: its own speech recognition, voice biometrics, dialog manager, fine-tuned small language models, and, since July 2026, its own generative text-to-speech model, Lexis.

On 6 August 2026 it raised a $67M Series B led by Expedition Growth Capital — only its second round, after $20M from Grafton Capital in 2020 — and reported live ARR above $60M, more than ten times what it was at the Series A, with roughly 500 employees. Named customers include Capital One, Discover, RBC, PSEG, Purolator, the UK Department for Work and Pensions, and Taco Bell, which in July 2026 extended Omilia’s drive-thru ordering agent to 890+ U.S. restaurants across 38 states. The vendor cites a Tier 1 U.S. bank running over a million calls a day and an insurer above 600,000.

Why it shows up in CX and support stacks

• It is the analyst-backed pick for regulated voice. Forrester’s Wave for Conversational AI Platforms for Customer Service (22 April 2026) named three Leaders — NiCE Cognigy, Kore.ai, and Omilia — and called it a fit for “highly regulated or security-first organizations.” Gartner’s 2026 Magic Quadrant for Conversational AI Platforms placed it as a Visionary for the second year. PolyAI was a Niche Player in the same Quadrant.
• It deploys where cloud-only vendors cannot. Omilia runs as managed SaaS on AWS or as dedicated bare-metal inside your own data center, and lists PCI DSS Level 1, SOC 2 Type II, ISO 27001, HIPAA, GDPR, and FedRAMP readiness. Because the speech stack is in-house, audio does not leave the platform for a third-party ASR, TTS, or LLM API. For a bank whose risk team blocks external model calls, that settles the shortlist.
• Fraud and authentication come in the same box. Voice biometrics authenticate the caller passively while they talk, and TalkGuard, its anti-fraud product, is what Gartner singled out as genuinely distinctive. Most AI-native voice competitors leave this layer to a separate vendor such as Pindrop.
• It plugs into the CCaaS you already run. Native connectors cover NICE CXone, Genesys Cloud, Amazon Connect, RingCentral, Talkdesk, and 8x8, plus a CSR CoPilot for the calls that still reach a human.

Pricing reality

No rate card. Omilia’s enterprise model is outcome-based: one price per resolved interaction, with no token metering and no compute overage, so a call that takes fifteen reasoning steps costs the same as one that takes three. That is the right incentive for a buyer, and it is the reverse of PolyAI’s per-minute meter.

The only public number is the AWS Marketplace listing for its miniApps (pre-built dialog components for address capture, dates, amounts): $0.025 per 20-second increment, about $0.075 a minute, pay-as-you-go. A three-minute call on that meter costs about $0.23. Use it as your reference point: if an enterprise per-resolution quote for three-minute calls lands well above that, you are paying a premium to move containment risk onto the vendor, which can still be worth it. We found no independent contract-value data; at the scale Omilia’s customers run, expect a negotiated multi-year enterprise agreement, not a card swipe.

Best for

Contact-center and CX leaders at Tier 1 and Tier 2 banks, insurers, utilities, and public-sector agencies who handle millions of voice calls a year in North America or Europe, need on-premise or no-external-API deployment, and want authentication and fraud screening in the same platform as the agent. The ROI is strongest where call volume is high enough that a few points of containment cover an enterprise contract.

Skip it if most of your contact volume is chat, email, or tickets — Omilia is built for the phone line; if you are a mid-market team that needs to self-serve and launch within a quarter; if your operations sit mainly in Asia-Pacific, where it appointed its first regional head only in September 2026; or if your roadmap depends on running the latest frontier LLM on every turn.

Versus the alternatives

The market-share leaders are the platforms Gartner ranks above Omilia. Kore.ai is the only vendor that is a Leader in both the Forrester and Gartner reports; pick it when the same platform has to cover employee service and digital channels as well as voice. If you already run NICE CXone, NiCE Cognigy is the incumbent option with one less vendor to onboard. Google and Salesforce win when your contact center is already built on their cloud or CRM.

Among AI-native voice specialists, PolyAI is the pick when the caller’s question lives in a hospitality or healthcare system of record such as OpenTable or Epic, and you accept per-minute billing. Parloa fits EU financial services buyers and contact centers outsourced to the large BPOs. The fastest-growing entrant is Sierra, at roughly $200M ARR by May 2026, which suits a consumer brand that wants one agent across chat and voice and is comfortable with LLM orchestration and outcome-based billing. If the job is mostly chat deflection, start with Decagon or Ada instead.

If none fit, keep your current IVR and put Omilia’s miniApps or your CCaaS vendor’s native AI on the two or three highest-volume intents. That tells you what containment is worth before you sign anything multi-year.

Watch-outs

• “Resolved” is a definition you negotiate, not a fact. The outcome-based price only protects you if a call that ends in a callback or an angry transfer does not count as a resolution. Guard: write the resolution definition into the order form (for example, no repeat contact on the same intent within 72 hours, no transfer to a human), require a monthly sample of billed resolutions you can audit, and agree a dispute process before go-live.
• Implementation and support pace is the recurring complaint. Omilia scores 4.7 on Gartner Peer Insights, but reviewers point to slipped delivery timelines, slow responses on post-go-live incident tickets, and development tools with “pretty significant quirks.” Guard: tie payments to delivery milestones, put response times for severity-1 incidents in the SLA, and ask for a reference customer who went live in the past 12 months.
• A proprietary stack also means a closed one. Owning ASR, TTS, and the language models is what removes token costs and third-party data egress, and it is also why you cannot swap in a newer frontier model on your own schedule. Guard: test open-ended, multi-intent calls in the pilot rather than scripted flows only, and confirm in writing that you can export dialog definitions, transcripts, and voice-biometric enrollments when you leave.
• A public-facing voice agent will be attacked by its own callers. Taco Bell, whose drive-thru agent Omilia has run since 2023, publicly reassessed the rollout in 2025 after viral prank orders, including one for 18,000 cups of water, before extending it in July 2026. Guard: set quantity and value caps, define a fast path to a human on repeated failures, and red-team the agent with adversarial callers before launch, not after the first viral clip.