Walk into most enterprise call centers today and you'll find something curious. Beneath the talk of artificial intelligence and automation targets sits hardware that predates the iPhone—hulking on-premise phone systems from Avaya, Cisco, or Mitel, their blinking server racks managing billions of customer conversations through technology fundamentally unchanged since the early 2000s.
The collision between AI's future and telephony's past has created what one could call the contact center industry's defining puzzle of 2026. While executives face mounting pressure to automate customer interactions—and vendors promise AI voice agents capable of handling 70% of routine calls—a stubborn infrastructure reality keeps getting in the way. Callab AI, a startup attempting to bridge this gap, claims that roughly 58% of the call center industry still runs on these legacy systems—though independent verification of this figure is limited and other industry data varies. Not cloud platforms. Not modern API-driven services. Physical boxes, often sitting in corporate basements.
For CIOs and contact center executives, the question has shifted. It's no longer whether to adopt AI voice automation. It's whether doing so requires ripping out systems that took years and millions of dollars to deploy, or whether there's a less disruptive path forward.
The answer—like most things in enterprise IT—is complicated.
When "Legacy" Means "Still Paying for It"
The cloud migration story that's dominated enterprise software for the past decade turns out to have a significant asterisk when it comes to contact centers. Yes, cloud-based Contact Center as a Service platforms have made inroads—Metrigy's MetriCast 2025 research, drawing on surveys of 1,397 IT leaders, found that 34% of operations used CCaaS as their primary platform as of mid-2025. But dig into the numbers and a different picture emerges. The majority of agent licenses remain tethered to on-premise infrastructure. Among large enterprises with 1,000 to 2,500 employees, on-premise usage hovers around 49%. For organizations exceeding 2,500 employees, it's 47.9%.
A separate survey from Calabrio conducted in January 2025 found that 31% of organizations still depended on on-premise contact center platforms, with 35% using on-premise workforce management and quality assurance tools.
The reasons aren't simply corporate inertia or IT departments stuck in the past. When Metrigy asked why companies kept infrastructure on-premise, the answers revealed genuine strategic concerns: security (58.7% cited this), reliability requirements (56.6%), cost considerations (53.2%), and the need for deep customization (43.4%).
These aren't small deployments gathering dust. In 2023, Avaya held 28.9% of on-premise platform revenue, according to Metrigy market share data. Genesys claimed 11.9%, Cisco 11.2%. We're talking about billions in capital investment and—perhaps more importantly—systems that actually work. They handle peak call volumes during holiday shopping rushes. They integrate with CRM platforms. They meet compliance requirements that took years to certify and auditors to sign off on.
Replacing all of that to chase an AI trend, however promising, is not a decision any CFO takes lightly.
The Engineering Workaround
Enter SIP trunking, which has quietly become the connective tissue between legacy telephony and modern AI voice platforms. The architecture sounds straightforward enough: an on-premise PBX routes calls using SIP—Session Initiation Protocol, the internet telephony standard—to an AI voice gateway. That gateway handles the conversation through automatic speech recognition, large language model processing, and text-to-speech synthesis. If the AI can't resolve the issue, it transfers the call back to the automatic call distributor with context preserved through SIP headers and CRM API integration.
In practice, though, the devil lives in milliseconds.
Engineers building these integrations obsess over latency. They target end-to-end response times below 300 to 500 milliseconds to maintain what humans perceive as natural conversation. Barge-in detection—the system's ability to recognize when a caller interrupts—needs to happen in under 150 to 200 milliseconds, or the interaction starts feeling stilted. Session Border Controllers mediate between the on-premise environment and cloud-based AI services, handling codec transcoding (typically Opus on the WebRTC side, G.711 or G.729 for legacy trunks), topology hiding, and security.
Google's Dialogflow documentation details SIP trunk and SBC configurations down to certified firmware versions. AudioCodes positions its Live Hub platform as working across "your PBX/SIP/Teams/CCaaS/WebRTC" environments—a kind of universal translator for enterprise telephony. Cognigy's Voice Gateway explicitly supports carrier and SBC integration. These aren't theoretical configurations dreamed up in vendor labs. They're production-grade paths that companies like PSEG, the New Jersey utility, have used to modernize IVR systems while keeping core telephony infrastructure exactly where it was.
Whether this represents genuine innovation or an elaborate technical Band-Aid depends somewhat on your perspective.
The Pressure Cooker
Ninety-one percent of customer service leaders report being under pressure to implement AI this year, according to a Gartner survey released in February 2026. The pressure arrives from multiple directions simultaneously. Boards expect efficiency gains. Competitors tout automation rates in earnings calls. And there's the genuine opportunity—not just hype—to improve customer experience while reducing operational costs.
Major platforms are racing to capture this moment. NICE announced in June that "agentic AI" would be native to its CXone core. Genesys reported 4x year-over-year growth in Agent Copilot adoption during fiscal 2025, alongside the rollout of more than 150 new AI features. Five9 launched a joint enterprise AI solution with Google Cloud in January. Amazon Connect introduced AI agents for real-time assistance and end-customer self-service. Twilio made Flex embeddable in April, allowing companies to integrate contact center capabilities more flexibly.
Yet for organizations with significant on-premise investments, the gap between these vendor announcements and operational reality can feel vast. Migrating to a cloud platform means more than software licensing and agent training. It means renegotiating carrier contracts. Recertifying compliance frameworks. Potentially disrupting operations that must maintain 24/7 uptime—try explaining to your CEO why customer calls went dark during a migration window.
The vendors know this. Which is why many are now selling the bridge rather than insisting on the leap.
When the Math Gets Messy
The economics of AI voice automation have turned out to be more complex than the early sales pitches suggested. Gartner issued a cautionary prediction in January 2026 that deserves attention: GenAI cost per resolution in customer service may exceed $3 by 2030—potentially surpassing what it costs to route calls to offshore human agents. The firm notes bluntly that "full automation will be prohibitively expensive for most organizations," and advises that "leading organizations will use AI to drive customer engagement rather than to cut costs."
This represents something of a conceptual pivot for an industry that initially sold AI voice primarily on deflection metrics and automation percentages—getting calls resolved without human agents touching them. McKinsey's widely-cited research on GenAI in customer care, published in 2024, documented early successes in reducing handle times and improving quality scores. But as implementations have scaled, the industry seems to be discovering that AI works best not as wholesale replacement but as an amplification layer. Assistance rather than elimination.
Five9's CEO, discussing the company's AI trajectory in May, emphasized the need to "marry voice, digital, and AI under one roof"—a unified platform vision rather than point solutions bolted onto legacy systems. Forrester's 2025-2026 predictions noted that many firms were quietly walking back aggressive agent reduction targets, recognizing that contact centers remain leading adopters of operational AI precisely because human judgment remains essential for complex scenarios.
For on-premise environments, the cost equation includes factors that cloud migrations can temporarily obscure. A week-long integration via SIP trunk—as some vendors claim their platforms enable—carries a different risk profile than a multi-quarter platform replacement. The question becomes whether incremental automation of, say, 70% of routine inquiries delivers sufficient ROI without the capital expense and operational risk of infrastructure replacement.
The answer probably varies by how much you spent on your current system, and how recently.
Compliance Gets Complicated
Regulatory frameworks are hardening around AI voice use cases, creating technical requirements that affect on-premise and cloud deployments alike—though not always in the ways vendors emphasize.
The FCC's February 2024 declaratory ruling classified AI-generated voices in robocalls as "artificial or prerecorded" under the Telephone Consumer Protection Act, making them illegal without explicit consent. STIR/SHAKEN attestation requirements tightened throughout 2025, with enforcement actions targeting spam labeling and traceback capabilities. Simply put: AI voice agents need to identify themselves, and the call routing needs to prove authenticity.
PCI DSS version 4.0 and 4.0.1 became mandatory on March 31, 2025, creating specific headaches for voice payments. Contact centers handling credit card data over the phone must implement DTMF masking, ensure CVV and PAN data don't leak into recordings or speech recognition transcripts, and carefully scope which components of their AI pipeline touch regulated data. The architecture choices matter here—whether speech recognition happens before or after payment data capture, how recording and analytics systems segment sensitive interactions.
The EU's AI Act adds yet another layer. Article 50 transparency obligations, which apply from August 2, require that chatbots and certain AI systems disclose they're AI when interacting with people. The European Commission issued draft guidance in May. For multinational contact centers, this means conversation design must include clear disclosure mechanisms, logging to demonstrate compliance, and human escalation paths that meet regulatory expectations.
These requirements don't inherently favor cloud or on-premise architectures. But they do demand rigorous design. An AI voice platform that taps into legacy PBX infrastructure via SIP still needs DTMF masking, still needs SIPREC-compatible recording for compliance, still needs to ensure that speech recognition and LLM storage don't inadvertently expand the regulatory scope. The question isn't whether to comply—that's not optional—it's how integration architecture affects where the compliance boundary sits.
The Long View
Gartner predicts that agentic AI will autonomously resolve 80% of common customer service issues by 2029. The same firm also states flatly that "none of the Fortune 500 will have fully eliminated human customer service by 2028."
The apparent contradiction captures something real about the industry's actual trajectory. Automation grows, yes. But so does the complexity of what remains for humans to handle, and the cost structures evolve in ways that shift the value proposition from pure labor arbitrage to experience differentiation. It's a more nuanced story than the one many early AI voice vendors told.
For enterprises with substantial on-premise contact center infrastructure, the next 24 months present something of a decision window. Cloud platforms are building increasingly sophisticated AI capabilities, and legacy vendors like Avaya have responded with strategies like Infinity—launched in April 2025—explicitly designed to bridge cloud and on-premise estates. The technical path to integrating AI voice agents without wholesale migration exists and is maturing. SIP trunk integrations, SBC configurations, and agent handoff protocols are well-documented and production-proven.
What remains less certain is whether incremental automation via these bridges delivers enough strategic advantage, or whether competitive pressure eventually forces broader platform modernization anyway. The answer likely varies by organization size, industry vertical, and how aggressively competitors move. A utility with stringent uptime requirements and regulatory complexity may calculate differently than a retail brand competing on customer experience innovation.
The broader market dynamics suggest hybridization rather than wholesale replacement. IBM's recent contact center automation trends guide emphasizes increasing usage of conversational interfaces, not the elimination of existing infrastructure. Forrester's research notes that contact centers are leading adopters of operational AI precisely because they can layer it onto existing workflows incrementally.
Perhaps the more interesting question—though one fewer executives are asking yet—is what happens when on-premise infrastructure finally does require replacement. Not because of AI, but because of mundane reasons: vendor end-of-life timelines, hardware depreciation cycles, the resignation of the last engineer who understands how everything's wired together.
At that point, the migration decision becomes genuinely unavoidable. Companies that have spent the intervening years learning how AI voice performs in their specific environment, what automation percentages prove realistic, and which workflows genuinely benefit from agentic systems will make far better platform choices than those rushing to implement AI while simultaneously migrating core infrastructure. That sounds almost obvious when stated plainly. Yet watching the current market, it's not clear how many organizations are taking that longer view.
For now, the promise of AI voice without rip-and-replace addresses a real market need: letting organizations experiment, learn, and capture value while deferring the larger strategic decision about contact center architecture. Whether that proves a bridge to the future or merely a costly detour depends less on the technology itself than on how quickly the broader industry's economics and customer expectations evolve.
The infrastructure will eventually change. Perhaps the question isn't whether it needs to change before AI can be deployed, but whether delaying that change teaches you something worth knowing—or just postpones an inevitable reckoning with a larger bill attached.
