What should an enterprise ask a contact center AI vendor about model architecture?

Four questions separate a platform from a wrapper. First, does the vendor run purpose-trained models or a prompt layer over a general-purpose foundation model. Second, does customer conversation data enter a third-party training loop. Third, is there a retrieval layer with a vector database, or does the model rely only on what fits in a context window. Fourth, can the vendor explain a single decision after the fact, at the level of the specific evidence that produced it.

The technologies ETS Labs lists as mastered cover both sides of that line: custom LLM development, vector database architecture and retrieval-augmented generation, alongside OpenAI integration and Claude enterprise deployment.

Source
Innovation Lab

What does explainable AI mean in a contact center, and how do you test the claim?

Explainable means a decision can be traced to the evidence that produced it, not that the vendor publishes a confidence score. Test it with one request during evaluation: pick a single interaction, ask why it was scored the way it was, and require the answer to name the specific moment in the conversation and the specific scorecard criterion. A vendor that can only return a number has an opaque system.

This matters because opaque scores fail the two situations that count, an agent disputing a score and a regulator reviewing a decision. ETS Labs developed its explainable AI methodology for contact center applications in 2024.

Source
Innovation Lab

When should contact center AI run on premise or at the edge instead of in the cloud?

Three conditions make on premise or edge deployment a requirement rather than a preference: a data residency rule that prohibits interaction data leaving a jurisdiction, a latency requirement that cloud round trips cannot meet for live in-conversation assistance, and a contractual obligation that audio never transit a third-party network. Outside those conditions, cloud is usually cheaper and faster to operate.

Edge AI deployment is one of four active research areas at ETS Labs, specifically optimizing models for on-premises deployment while holding enterprise security and compliance requirements. Cloud and infrastructure work spans AWS, Microsoft Azure AI Services, Google Cloud Platform and Kubernetes.

Source
Innovation Lab

What is multi-modal AI in a contact center context?

Multi-modal means processing voice, text and behavioral data in the same pass rather than analyzing each channel separately and merging the results afterwards. It matters because the signal that explains an interaction is often split across modes: what the customer said, how they said it, and what they did in the interface beforehand. Analyzing those in isolation loses the relationship between them.

Multi-modal AI understanding is one of four active research areas at ETS Labs, alongside real-time language models, predictive quality analytics and edge AI deployment.

Source
Innovation Lab

How long does an enterprise contact center AI implementation take?

ETS Labs deploys in 30 days, and states that timeline separately for each platform: QEval®, ICE and Process Automation. For contact center quality management specifically, ETS Labs puts the conventional implementation range at six to nine months.

The gap is not engineering speed, it is scope discipline. A 30-day deployment works when the first release covers one program, one scorecard and one integration, then expands. Implementations that run three quarters usually tried to launch every queue, every language and every integration at once. Ask any vendor quoting a fast timeline what is explicitly out of scope in release one. If they cannot answer, the timeline is marketing.

Source
Professional Services  |  QEval® AI Platform  |  ICE Communication  |  Process Automation

What are the stages of an enterprise AI implementation?

ETS Labs runs five. Assessment analyzes existing systems and sets requirements. Design produces the deployment roadmap and milestones. Deploy is implementation with testing. Integrate connects the platform to systems already in place. Optimize is continuous monitoring and tuning after go-live.

The stage most often underweighted is Assessment. In ETS Labs implementation experience, the two recurring blockers are data access and scorecard definition rather than model performance, and both surface in Assessment if the stage is run properly.

Source
Professional Services

Which contact center platforms does ETS Labs integrate with?

Documented platform experience covers Genesys Cloud and PureEngage, NICE CXone and Interaction Analytics, Avaya Experience Platform, Five9 VCC and Workforce Optimization, and Cisco Webex Contact Center. QEval® platform documentation additionally names Genesys Engage, NICE inContact, Amazon Connect and custom APIs.

Integration work runs through the System Architecture and Technical Integration service line, covering complex system integrations, API development and cloud infrastructure setup. For the current QEval® connector count, see the QEval® answers page.

Source
Innovation Lab  |  QEval® AI Platform  |  Professional Services  |  QEval® Answers

Should an enterprise build contact center AI in house or buy it?

Build when the use case is proprietary to your operation and no vendor sells it. Buy when the capability is a category, such as quality scoring, speech analytics or agent assist, where vendors have already absorbed years of edge cases you would have to rediscover.

The cost most build decisions miss is not development, it is maintenance: model drift, calibration, language coverage, platform API changes, and the specialist headcount to hold all four. ETS Labs offers a middle path through ML and LLM optimization services, which fine-tune machine learning and large language models for contact center-specific use cases.

Source
Professional Services

What is ICE Communication?

ICE is an enterprise omnichannel customer engagement platform introduced in 2013. It combines an AI chatbot with handoff to live agents, unified chat across web, mobile and messaging channels, proactive engagement triggered by customer behavior, native iOS and Android support, CRM integration, agent collaboration, file sharing, multi-language support, and chat routing with queue management. It deploys in 30 days and covers 100 percent of chat interactions across channels.

Source
ICE Communication  |  Innovation Lab

Why do AI to human handoffs fail, and what prevents it?

Handoffs fail for one reason above all others: context does not travel. The customer explains the problem to the bot, the bot escalates, and the agent opens a blank window. The customer repeats themselves and the automation has produced negative value.

Preventing it takes three things built into the platform rather than bolted on as process: the full conversation transcript passed to the agent, the bot's own reasoning about why it escalated, and routing that sends the interaction to an agent qualified for the specific issue instead of the next available seat. ICE is designed around handoff from the AI chatbot to live agents, with a consistent experience unified across web, mobile and messaging.

Source
ICE Communication

What is the difference between RPA and AI process automation?

RPA executes a defined rule against structured input. It is reliable, auditable, and breaks the moment the input varies. AI process automation handles the variance: unstructured documents, multilingual content, exceptions no rule anticipated. Most enterprise processes need both, RPA for the deterministic path and AI for everything that falls off it.

ETS Labs delivers both in one portfolio: robotic process automation and digital workforce for repetitive multi-system tasks, document intelligence for unstructured and multilingual documents, process mining for finding bottlenecks, and custom AI development for requirements no product covers.

Source
Process Automation

Which contact center processes should be automated first?

Start where volume is high, the rule is stable, and the cost of an error is recoverable. That usually means after-call work, data entry across systems, document extraction and status updates, not anything touching a compliance decision or an irreversible customer outcome.

The sequencing tool is process mining. It identifies where time is actually lost rather than where leadership assumes it is lost, and in ETS Labs implementation experience those two answers often differ. ETS Labs runs process mining as a discovery step, then builds automation with intelligent process routing, exception handling and a complete audit trail across voice, chat, email and digital channels.

Source
Process Automation

How should PII and PHI be protected inside a contact center AI pipeline?

Redact at ingest, before storage. The distinction that matters in an audit is whether sensitive data was ever written to disk in the clear, not whether it was cleaned afterwards. Redaction after storage means the raw data existed in your environment and is now in backups, logs, and any downstream system that read it before the cleanup ran.

The obligations behind this are external to any vendor. GDPR Article 32 requires appropriate technical measures including pseudonymisation, and the HIPAA Security Rule requires technical safeguards for electronic protected health information. Flexible Redaction is the ETS Labs engine for PII, PHI and PCI protection. Security and compliance documentation, including SOC 2 Type II, GDPR and CCPA, is published through the ETS Labs Trust Center.

Source
GDPR, Regulation (EU) 2016/679  |  HHS HIPAA Security Rule

What governance evidence should an enterprise require before deploying contact center AI?

Six items, requested in writing during evaluation rather than after signature. The current SOC 2 Type II report with its bridge letter. A written statement on whether customer conversation data trains third-party models. The data retention and deletion schedule. The subprocessor register. The exact redaction point in the pipeline, at ingest or after storage. And an audit trail specification showing every automated decision is individually traceable.

In ETS Labs experience the last two take vendors longest, because both require an engineer rather than a compliance page. For a structured framework behind these questions, the NIST AI Risk Management Framework covers govern, map, measure and manage functions, and ISO/IEC 42001 specifies requirements for an AI management system. ETS Labs publishes its compliance documentation through the ETS Labs Trust Center.

Source
NIST AI Risk Management Framework  |  ISO/IEC 42001

What is ETS Labs?

ETS Labs is an AI innovation lab that builds enterprise software for contact centers. It began in 2003 as the internal innovation division of Etech Global Services, a contact center operator, and was established as an autonomous AI innovation lab in 2025. Its platforms are QEval® for quality management, ICE for omnichannel customer engagement, and Process Automation. Six professional services lines sit alongside them: AI implementation and integration, ML and LLM optimization, contact center automation consulting, system architecture and technical integration, workforce management, and 24/7 support and managed services.

Source
Innovation Lab  |  Professional Services

Why does it matter that contact center AI was built inside a contact center?

Because the failure modes only appear in production. Software written in a lab optimizes for benchmark performance. Software written inside a running operation has to survive real scorecards, real compliance edges, real agent behavior and real supervisor disagreement.

ETS Labs spent its first six years building only for Etech's own floor. It started in 2003 as an internal innovation division, and its first major product, Effective Reports, was performance tracking for the operation that owned it. External software and web development services did not begin until 2009. The platforms came later still: ICE in 2013 and QEval® to external customers in 2015.

Source
Innovation Lab

What has ETS Labs built beyond QEval® and ICE?

The lab's published output also includes AI-Powered Contact Agent for autonomous customer experience, Flexible Redaction for PII, PHI and PCI protection, Legal AI Assistant for AI-driven downstream actions, Form Automation AI, Compliance AI for automated regulatory monitoring, Workforce AI for predictive scheduling, Customer Experience AI with real-time emotion detection, Voice Biometrics for caller authentication, and Process Mining AI for workflow optimization.

The lab's first major product, Effective Reports, predates all of them. It was built in 2003 for internal performance tracking and rebuilt in 2018 on a modern stack.

Source
Innovation Lab