Voice Will Become the Default Interface for Enterprise AI

Luiz Domingos describes enterprise AI as having crossed from experimentation into operational infrastructure. The consequential change is not simply that companies use better models. It is that AI is being embedded into communications workflows where latency, data control, auditability, and human accountability determine whether the system can be used at all.

His strongest interface prediction is that voice will become the natural way people interact with enterprise AI. Companies, in his view, will not primarily “text” or “chat” with AI over the long run. Workers will talk to it. The reason is practical: enterprise communications already run through voice, meetings, contact centers, alerts, messages, and workflows that happen while people are doing other work. For many workers, especially outside office settings, the screen is the wrong interface.

That prediction sits alongside a deployment claim. Domingos argues that cloud-only AI is inadequate for important categories of enterprise communication, especially real-time voice and regulated industries. Voice interactions have latency requirements that generic cloud deployments may not satisfy. Regulated sectors have privacy, sovereignty, auditability, and liability requirements that may prevent conversational data from being processed in unconstrained public-cloud systems.

Mitel, where Domingos is CTO, is a long-established enterprise communications company with products across communication, collaboration, contact centers, and vertical applications. Domingos says the company has more than 50 years in the market, serves 70 million users, and began investigating AI around 2017 and 2018, including natural language processing and Dialogflow-style contact center applications. He treats the period after ChatGPT’s public arrival in 2022 as materially different: AI moved from novelty and planning into day-to-day use.

Two or three years ago, he says, much of the enterprise discussion was still “slideware”: proofs of concept, strategy decks, and innovation-lab exercises meant to understand what generative AI might be worth. Today, the discussion has shifted toward execution. Companies want AI embedded in business processes, with outcomes that can be measured and compliance treated as part of the deployment rather than a later control.

That shift changes the executive question. Domingos says the argument is no longer mainly about “which model is the smartest.” It is about “which deployment brings ROI,” which deployment respects enterprise governance, and which deployment reduces friction in daily operations. In communications, he argues, the value can become visible quickly because the work is close to measurable operational outcomes: contact center handle time, call summarization, routing quality, customer satisfaction, and documentation burden.

The same deployment logic shapes his advice on modernization. Companies that layer intelligent automation on top of outdated systems, fragmented APIs, and inconsistent data do not create efficiency, he says; they raise complexity. Modernization does not require ripping everything out. It does require becoming API-first, decoupling the communications layer from the workflow layer, building modular AI services for ingestion, retrieval, and orchestration, and plugging those services into existing systems with clean data pipelines and clear governance.

Domingos uses contact centers as the clearest example of AI already producing measurable value. In that environment, AI can act as an assistant to the human agent: transcribing calls, analyzing sentiment, summarizing interactions, providing guidance on how to respond, and helping the agent reach a better customer outcome faster. It can also screen calls before they reach a human agent and support intelligent routing.

The defining constraint is speed. In a contact center, “every second matters.” If an agent is in a live conversation and an AI assistant needs two seconds to transcribe, interpret, and suggest a reply, Domingos says that is already too late. The recommendation has to appear close to immediately because the customer interaction is unfolding in real time.

That is one reason he argues for edge AI in communications. In Mitel’s usage, “edge” does not mainly mean an AI model running on a phone, though Domingos says endpoint AI is one scenario. He is mostly referring to bringing inference closer to the customer’s operations: a local data center, an on-premise cloud, or a constrained secure environment. The point is to reduce latency and avoid dependence on network access, bandwidth, or distant inference calls when performance directly affects the conversation.

Small time savings compound in this setting. In early agentic workflows, Mitel is using Workflow Studio in contact centers to summarize calls, write information back into databases, and create tickets in systems such as ServiceNow. Domingos characterizes those as initial implementations, but he emphasizes their immediate value. If an agent saves 10 or 20 seconds on summarizing a call or typing a ticket, that matters inside a contact center’s operating model.

Contact center AI also makes Domingos’s distinction between assistance and action visible. An AI assistant can process conversational data and recommend next steps. It can give the agent relevant information, summarize what happened, or suggest where a call should be routed. Domingos does not call that agentic AI. In his definition, agentic AI begins when the system takes precise actions inside an enterprise workflow: writing the ticket, updating the database, triggering a follow-up, or moving the process forward across systems.

The terminology is especially easy to confuse in communications because “agent” already refers to the person in the contact center. A contact center agent may have an AI agent or AI assistant, but that does not make the system agentic by nature. Recommendation and insight are different from action. Intelligent routing, for example, is not agentic in his account; it can be handled through generative capabilities and rules that determine the best routing destination for a call, chat, social media inquiry, or other request.

For unified communications, the emphasis shifts. Contact centers are about real-time response, agent productivity, and customer satisfaction. Unified communications, as Domingos defines it, is broader than voice calls. It includes the channels and modes through which people collaborate: calls, chat, conferencing, collaboration sessions, multimedia, omnichannel communications, and the act of bringing the right people and information into a business process.

In that environment, AI’s job is to help the enterprise use its own knowledge more effectively. Domingos points to knowledge bases and retrieval-augmented generation as the relevant pattern. A support person who previously asked an internal expert how a product worked can instead use an AI assistant sitting on top of a company knowledge base. The system can provide guidance, relevant facts, and fragments of language that the support person can use in a response. Without enterprise augmentation, he says, generative AI is “plain vanilla AI.” It lacks the specific knowledge, culture, and workflows that businesses need.

The voice-first prediction is tied to the conditions of frontline work. Luiz Domingos says frontline workers are the group most underserved by digital transformation. Knowledge workers have many digital tools; frontline workers often still lack real-time digital tools that match their environment.

Nurses in a hospital are his main example. They may need alerts, alarms, messages from doctors, emergency-room notifications, and guidance while physically caring for patients. They are not sitting at a desktop, and they cannot always browse an app. For those workers, the relevant interface is event-driven and immediate: notifications, alarms, critical-situation alerts, and voice interaction.

A nurse at a patient’s bed should not have to look down at a screen to interact with an AI system. A field or frontline worker should not have to stop the physical task in order to operate a GUI. In that setting, voice-first AI is not just a convenience; it is the interface that fits the work.

This is also where Mitel’s communications background matters in Domingos’s account. The company does not present itself as building its own text-to-speech or speech models. Instead, Domingos says Mitel consumes text-to-speech, natural language processing, and speech synthesis models through an AI services platform designed to be model-agnostic. The platform provides a middle layer that connects applications to language models through a standard interface.

That abstraction is important because model choice and deployment location can vary. A customer may want a public-cloud model, a sovereign “secure cloud” deployment for certain European customers, or an on-premise edge deployment. Domingos says Mitel wants applications to have one implementation while customers choose the model and deployment pattern required by latency, regulation, or internal policy.

His five-year prediction is that voice interfaces will become table stakes alongside real-time transcription, summarization, and AI-powered routing. He compares transcription and summarization to call recording: once premium, eventually standard. In the future he expects them to be part of ordinary subscriptions rather than exceptional features. Similarly, he expects no serious contact center or customer-support infrastructure to lack AI-powered routing.

Voice is the more expansive change. Domingos imagines using a commute to interact naturally with an assistant before reaching the office, doing the things he would otherwise need an application screen to do. The screen does not disappear from every workflow in his account, but the primary interaction shifts: instead of browsing panels in an application or navigating a GUI, a worker speaks to an assistant.

Mitel’s agentic work is centered on enterprise workflows, not autonomous general-purpose agents. Domingos returns to Workflow Studio as the example: a tool for creating communications workflows that combine communication, AI consultation, data retrieval, and action. The data retrieval layer can involve servers collecting information from different systems. The action layer can create tickets, write records, or advance a process.

The helpdesk ticket example captures the pattern. A person calls IT support with a problem. In a traditional workflow, the support person understands the request and manually creates a ticket. With agentic AI, the system can help create and log the ticket automatically, reducing manual work and ensuring the request enters the process for follow-up. Domingos extends the same idea beyond support and contact centers to finance departments and other enterprise functions where communication triggers administrative workflows.

But the enthusiasm is bounded by governance. Domingos is explicit that, in risky and regulated workflows, enterprises are not inclined to let agentic systems operate without supervision. The concern is not abstract discomfort. It is reputational risk, litigation, regulatory liability, auditability, and accountability when something goes wrong. In the kinds of regulated or customer-facing workflows he describes, he says no enterprise wants an AI workflow acting on its own without a human in the loop or at least a person giving a “thumbs up.”

The hard question is how much autonomy to give the system, where to require supervision, and how to make the decision trail auditable. Domingos says this is the real discussion he sees among enterprise technology leaders. At an IT Expo he attended, he says the main concern was not the beauty of agentic capabilities but governance, liability, and auditability. CIOs are trying to understand how to deploy AI responsibly and reliably, and in many cases they do not yet know how.

That uncertainty has architectural consequences. Domingos says Mitel wants governance frameworks built directly into the platform architecture, including monitoring, auditability, explainability, and human-in-the-loop mechanisms as needed for a given vertical solution. The responsible-AI layer is not merely policy language around the system; it has to be present in deployment options, data handling, logs, controls, and workflow design.

Domingos frames communications data as unusually sensitive because it is made of conversations: calls, chats, meetings, customer interactions, and internal exchanges. Those conversations can contain enterprise secrets, customer data, and information that may later become relevant in court or regulatory review. When AI processes them, the system must handle the data with the same care the enterprise already owes to the underlying communication.

Finance is his clearest compliance example. In trading and other transaction-driven businesses, calls may need to be recorded for compliance reasons, not just training. If a regulator later reviews a bank transaction or stock-market transaction, the enterprise may need to locate the relevant call, understand what happened, and provide information back. Historically, finding and interpreting the right call could be difficult. With AI, recorded calls can be transcribed and searched, making review more practical. But the same capability also raises the stakes around data handling, auditability, and liability.

European privacy and AI regulation are another example in his account. Domingos refers to GDPR and the European AI Act as regulatory regimes that impose requirements on AI providers and companies using AI products, particularly around how customer and individual data is processed, used, and protected. The enterprise must protect that data not only from misuse but from actors trying to exploit it.

That is why customers ask whether their knowledge base, customer call transcript, or chat data is being sent to a public cloud and becoming part of broader model knowledge. Domingos describes this as a major concern: enterprises want to know whether the information remains isolated for their use, rather than being lost into or “polluted” by the public cloud. Edge AI and constrained deployments, in his view, give customers a way to choose more isolation and stronger protection, with different cost layers.

Mitel does not claim to solve compliance with one universal product. Domingos says the company’s products have capabilities that enable compliant designs, such as the ability to deploy AI applications and language models across different infrastructures to satisfy sovereignty or privacy requirements. But each customer and vertical has its own compliance definition. Mitel’s services organization may provide advisory work: understanding requirements, helping create a compliance framework, recommending applications, or implementing a customer-specific solution.

In this framing, compliance is not just keeping a record of calls. It includes data sovereignty, privacy, explainability, auditing, accountability, and the ability to demonstrate that what was implemented is actually compliant. Agentic AI makes the issue harder because it raises the question of where the decision was made and who is responsible for it.

Architecture is only part of deployment. Domingos says the other part is the services and partner model needed to make the technology fit actual customer environments. Some customers are self-sufficient and maintain Mitel systems themselves, but he characterizes them as a small minority. Most customers need help translating a use case into a deployed configuration.

Those partners bring service experts who understand customer use cases, configure systems, update them, and operate them. Mitel also has its own professional services organization, which performs similar work, typically for larger enterprises.

The pattern is practical rather than abstract. Mitel may provide a contact center product with AI capabilities already built in: summarization, sentiment analysis, assistance, transcription, routing, and related functions. On top of that product layer, partners or Mitel’s services teams may deploy and later operate the solution for the customer. For customers with heavier compliance or vertical requirements, the work may include advisory services and implementation support rather than product configuration alone.

This delivery model matters because Domingos’s modernization argument assumes that AI adoption is entangled with existing systems. Communications workflows touch ticketing systems, CRM, ERP, databases, compliance archives, and human approval processes. The customer is not simply buying a model. It is deciding where the system runs, what data it can access, what action it may take, who approves those actions, and how the resulting records can be audited.

Energy and cost also enter the deployment question. Domingos treats energy consumption as a real concern, though not equally salient in every market. Customers in environmentally conscious countries, especially in Europe, raise the issue more often. Elsewhere, customers may care more directly about cost, but energy still affects cost. He acknowledges that large models demand significant energy and points to public discussion about power generation, including nuclear power plants being restored or brought back to operation for energy needs.

His response is to right-size the system. The relevant question, he says, is how much AI power a function or workflow actually requires. Enterprises do not always need the largest language model for a given task. Smaller, specialized models may be more appropriate. Architecture should minimize redundant compute and avoid unnecessary attempts across multiple paths when a task can be completed in a single pass.

Edge AI contributes to that argument by reducing unnecessary data transport and bringing inference closer to the work. Processing near the customer’s operations can avoid moving conversational data back and forth across networks, while also supporting the latency, sovereignty, and isolation requirements Domingos has already described. For him, the same architectural choice can serve cost, performance, compliance, and sustainability when the model and deployment footprint are matched to the task.

Domingos is skeptical that pilots alone produce real AI ROI. Pilots can be useful, but he says they often fail to surface the full complexity of deployment. Many companies run dozens of AI experiments before aligning them with measurable business impact. The better starting point, in his view, is to map the workflow itself.

That means asking where work slows down, where manual steps occur, where too many people touch the same data, where errors are introduced, where compliance risk appears, and which steps AI can perform. Only after that should the enterprise define the foundation, select the workflow, and choose the right deployment. If a company begins with a pilot before understanding the value and the source of friction, it may never know what return on investment it is supposed to produce.

The metrics Domingos names are practical: resolution time, cost per interaction, revenue impact, time savings, efficiency gains, productivity improvements, customer satisfaction, and employee satisfaction. Some can be measured precisely. Others, such as satisfaction, combine quantitative scoring with qualitative judgment. In contact centers, customer satisfaction scores can be tracked, but the underlying question is still whether customers are happier with the service. For employees, success includes whether the system reduces burden rather than simply adding another tool.

Mitel’s advisory process begins with communications workflows, but Domingos says those workflows often touch CRM, ERP, ticketing systems, and broader IT infrastructure. A discovery workshop starts with the customer’s use cases: what is not working, where automation could matter, where manual steps create friction, and what business impact should be measured. From there, Mitel may define targets and KPIs, then design pilots or implementations using Workflow Studio, agentic features, or customer-specific deployments.

The organizational implications extend beyond communications teams. Domingos says IT is “morphing into something different” because AI is not just a technology upgrade; it becomes a kind of organizational redesign. As AI automates scheduling, analytics, and workflow orchestration, boundaries between IT, operations, and other departments begin to break down. AI systems cut across domains, so governance and ownership must cut across domains as well.

That requires AI literacy beyond data science teams. Operational leaders need to understand what AI is doing, how they interact with it, and where its value and risks sit. Domingos argues for cross-functional workflow ownership and governance councils that manage risk, ethics, and compliance holistically. Organizations that succeed with AI, in his view, will not merely deploy tools; they will build an organizational construct around the new way work crosses systems.

Domingos describes Mitel’s next stage as a move from AI features to AI-native architectures. In practice, that means deeper embedding of AI across the portfolio rather than a standalone AI product beside communications infrastructure. Agentic workflow orchestration should become more mature, allowing communications to trigger cross-system workflows across the enterprise. Vertical-specific AI should also become more important because, in Domingos’s view, generic AI will not solve industry-specific problems.

He expects edge capabilities to strengthen, especially premise-based AI for real-time intelligence and regulated use cases. He also expects governance frameworks to be built more directly into architecture: monitoring, auditability, explainability, and human-in-the-loop controls where a vertical solution requires them.

The destination, as Domingos lays it out, is AI as the intelligence layer inside enterprise communications: contact centers, unified communications, workflow orchestration, routing, transcription, summarization, compliance review, and vertical applications. Some of today’s premium features become standard. Some agentic actions become accepted where supervision and auditability are sufficient. And if his interface prediction is right, many of those interactions will happen through voice rather than a screen.