Omilia

Named connectors reached at the click of a button, with agents taking authorized action behind them. Out-of-the-box CCaaS integrations are named on the vendor's own platform page, including NICE CXone and RingCentral; the wider set covers NICE CXone, Genesys, Amazon Connect, RingCentral and Talkdesk on the contact-center side, with Salesforce, Microsoft Dynamics 365 and SAP on the CRM side.

The action layer is documented: agents connect to customer APIs, MCP servers, CRMs and knowledge bases, with authorized integrations defined per agent and executed inside OCP's certified perimeter, and Task Agents reach enterprise APIs and MCP servers through the orchestration layer. The work is transactional: authentication, payment capture and balance transfers are completed inside the automated flow rather than passed to a person.

Sourceomilia.com self-learning agentic CX platform and enterprise voice AI pagesread 2026-09-06

A named orchestration layer routes between distinct agent types and execution modes. Omilia owns every layer through to its own agentic orchestration engine, running a proprietary chain from speech recognition to natural language understanding to a Dialog Manager to text-to-speech, with Service Agents, Security Agents, Authentication Agents and CSR Copilot operating as distinct types.

Composition works at two levels: miniApps are task-specific dialogue components with their own disambiguation and error recovery, assembled into a conversation, and the Orchestrator routes between deterministic miniApp paths and agentic execution so a team can blend both without re-platforming. Pathfinder blends agentic reasoning with deterministic planning.

Agents are described resolving conversations even when customers change topic or interrupt mid-interaction, and escalation to a human is a defined end state when a task falls outside an agent's authorized scope.

Sourceomilia.com enterprise voice AI and platform pagesread 2026-09-06

Knowledge bases are first-class objects in the platform, built, attached and crawled. Omilia states the platform can build new knowledge bases and retrieve information from documents or APIs, ships an OCP Knowledge Engine as a named component of its autonomous tier, and describes agents connecting to customer APIs, MCP servers, CRMs and knowledge bases with authorized integrations defined per agent.

The management surface is documented down to the verbs: Pathfinder projects are created, URLs are crawled into FAQs, and knowledge bases are attached to specific agents, listed alongside agents and orchestrator apps as discoverable platform objects. Ingest, index, attach and retrieve add up to a maintained retrieval structure that persists and stays queryable. Omilia publishes no index type, embedding model, chunking or refresh cadence, so the corpus and its management are documented but the mechanism is not.

Sourceomilia.com self-learning agentic CX platform and enterprise voice AI pagesread 2026-09-06

Approval sits in the path by design, not as an escape hatch. Omilia gates its own self-learning loop: the offline pipeline discovers resolution patterns and drafts automation without touching the live runtime, human experts review and refine the suggestions before launch, and once validated the new or optimized agents are deployed to production under human supervision. So the platform cannot change its own behavior in front of customers without a person releasing it.

At run time the limits are explicit too: each agent operates within configured authorization limits and escalates automatically to a human for any task outside its defined scope, so containment is a configured rule rather than a model judgment. Glass box AI is the stated design philosophy (full transparency and human oversight in every AI decision), and the Agentic Adoption Framework names human-in-the-loop as the structure around adoption.

Sourceomilia.com self-learning agentic CX platform and enterprise voice AI pagesread 2026-09-06

Access works as a per agent authorization model, and the certifications are published. Each agent operates within explicitly configured authorization limits and escalates to a human for any task outside its defined scope, with authorized integrations defined per agent and executed inside OCP's certified perimeter. That is least privilege applied to the actors that take action.

The certifications and frameworks named are PCI DSS including Level 1, AICPA SOC 2 Type II, ISO/IEC 27001 and UK NCSC Cyber Essentials, alongside EU and UK GDPR, CCPA and HIPAA. The anti-fraud stack (passive voice biometrics, liveness detection, blocklisting, speaker change detection and ANI spoofing analysis) authenticates the buyer's end customers rather than governing who inside the buying organization can do what.

Sourceomilia.com enterprise voice AI and platform pagesread 2026-09-06

Reconstruction, not just reporting, and the vendor sells it as such. Omilia states that every action, system call and decision is logged with full attribution, giving a traceable record of every automated outcome for compliance and operational review, plus a full interaction audit trail for every resolved task.

Glass box is the stated design philosophy: visibility into how the AI reaches conclusions, routes conversations and applies changes, so compliance teams in regulated industries can audit AI behavior and demonstrate regulatory adherence. Alongside the runtime trace sit OCP Monitor for tracking interaction data including call dialogs and web chat, OCP Reporting for dashboards, and raw application data accessible under GDPR with near real-time streams. Decision-level explainability, change auditing and per-call traces together are unusually complete coverage.

Sourceomilia.com enterprise voice AI and conversational insights pagesread 2026-09-06

Context holds within and across an interaction, and no memory layer is described. Omilia maintains context as a customer moves between voice, chat, messaging apps, mobile apps and smart speakers, so a conversation that starts on the phone and continues in chat is one interaction rather than two. Agents handle topic changes and interruptions mid-interaction, which requires state carried across turns. That is conversation state, which every conversational platform has.

No memory layer with a stated scope and lifetime is documented: nothing says what an agent retains about a caller between separate contacts, whether memory is scoped per customer, account or workflow, how long it persists, or how an operator inspects or deletes it. An analyst assessment lists memory among the platform's agentic capabilities, but Omilia's own pages do not describe one.

Sourceomilia.com self-learning agentic CX platform and voice agents pagesread 2026-09-06

The customer's own data center is a supported deployment. Omilia runs either on its own managed AWS infrastructure or on dedicated bare-metal nodes inside the customer's data center, and states that in both modes no customer utterance data leaves the certified environment, tying the arrangement directly to PCI Level 1 requirements.

On-premise deployment is offered explicitly for Tier 1 financial institutions and government agencies with strict data residency mandates. Running on the customer's own hardware, rather than in their cloud account, is the harder version of a customer-controlled deployment.

Sourceomilia.com enterprise voice AI pageread 2026-09-06

A pre-built component library plus named agent types, both ready to adopt. miniApps are the distinctive asset: task-specific dialogue components pre-trained and pre-tuned by Omilia to handle one job well, such as understanding an intent or validating data, each shipping with its own disambiguation and error-recovery handling, which a team drags into a conversation rather than building.

Above them sit named agent types delivered with the platform (Service Agents, Security Agents, Authentication Agents and CSR Copilot), each doing its own job independently of the others. The self-learning pipeline extends the set at run time by auto-generating Playbooks from discovered resolution patterns. Low-code and no-code tooling is positioned to remove design and development effort and reach production in days.

Sourceomilia.com self-learning agentic CX platform and platform pagesread 2026-09-06

Voice is the primary arrival path, and it is harder to serve than text. Work reaches the agent through the telephone line by way of the customer's CCaaS platform, and alongside it through web chat, messaging apps, mobile apps and smart speakers, each an independent inbound queue rather than another view of one surface, with context maintained as a customer moves between them.

Voice Agents and Chat Agents ship as separately named products across voice and digital, and the platform supports a wide range of languages. No event catalog, schedule or outbound campaign surface is documented, so the agent answers rather than reaches out, and the self-learning pipeline runs offline on completed interactions rather than starting new ones.

Sourceomilia.com self-learning agentic CX platform, voice agents and chat agents pagesread 2026-09-06

No customer or admin model choice, and the absence is a stated design position rather than an omission. OCP is vertically integrated, owning every layer from the acoustic model through natural language understanding and the Dialog Manager to its own Lexis text-to-speech and the orchestration engine. Omilia states its proprietary models are fine-tuned for enterprise voice rather than adapted from general-purpose frontier models.

It contrasts itself with platforms assembled from third-party speech, language and speech-synthesis APIs, and gives the reason: eliminating third-party data flows so that no customer utterance leaves the certified perimeter, which is what lets it hold PCI Level 1 and serve on-premise government deployments. So the closed stack is the product's security argument, not a gap in it, and the vendor explains the trade-off openly.

Sourceomilia.com enterprise voice AI pageread 2026-09-06

An official MCP server plus a CLI, with no published API or SDK reference behind them. Omilia ships MCP Tools as an official server under its own GitHub organization, installed by a wizard through npx, running locally and talking to OCP over authenticated HTTPS, and supported on Claude Desktop, Claude Code, Cursor, VS Code and Codex. It exposes the platform itself rather than a bridge to the customer's systems.

It lists and searches OCP groups, agents, knowledge bases, Pathfinder projects, Orchestrator apps, miniapps, phone numbers, variable collections and dialog logs; creates and deploys orchestrator apps; creates and updates Concierge and Task agents; configures sub-agents and escalation queues; creates and edits WebService miniapps; and attaches knowledge bases. So Omilia's platform can be operated from outside. No public API reference, SDK or developer portal is published. The authenticated HTTPS calls the server makes imply an underlying API, but an implied interface is not a documented one.

Sourcegithub.com/omilia/mcp, omilia.com enterprise voice AI pageread 2026-09-06

Agents are tested before release and scored against real outcomes. Pre-production testing validates agent behavior against real-world interaction patterns before any configuration reaches live customer traffic, the offline self-learning pipeline generates improvements without modifying the live runtime, and validated agents are deployed to production under human supervision.

Before going live, agents are tested with real and simulated conversations, including replaying historical data to predict performance, and the platform measures success against real agent outcomes and tunes toward accuracy and satisfaction. Replaying historical interactions against a candidate configuration and scoring it against what human agents actually achieved is a test harness with a comparable result, applied to the customer's own agent configuration. No pass rate, threshold or named test set format is published alongside the harness, so the mechanism is documented and its calibration is not.

Sourceomilia.com self-learning agentic CX platform and enterprise voice AI pagesread 2026-09-06

No browser, desktop or remote computer control is documented. Omilia's agents reach external systems through connected CCaaS platforms, customer APIs, MCP servers, CRMs and knowledge bases, with authorized integrations defined per agent and executed inside the certified perimeter. Their customer-facing surfaces (telephony, chat, messaging apps, mobile apps and smart speakers) are channels the agent speaks through rather than interfaces it operates.

The architecture makes this structural: the security argument rests on no data leaving the certified perimeter, which is inconsistent with an agent driving arbitrary third-party interfaces. Adjacent but different: voice agents navigate telephony and complete authentication and payment capture inside a call, which is real work through a human-facing medium, but a telephone is not a browser or desktop.

Sourceomilia.com enterprise voice AI and platform pagesread 2026-09-06