Syntic
All posts

Published October 4, 2026 in Voice

The Call Center Is Becoming Software

By Syntic AI

Every contact center is built around one physical fact: a person can hold only one conversation at a time. Seats, shift schedules, queue music, IVR menus, wrap-up codes and a quality team that can only ever hear a fraction of the calls are all machinery for rationing that scarce attention.

When speech recognition, reasoning and speech synthesis can run together in real time, that fact stops being the binding constraint. Capacity becomes something you provision, and the operation starts to behave like a software system: designed, versioned, tested and monitored. The hard problems move with it. They are now conversational timing, the telephone network itself, verification, and a body of law that treats answering a call very differently from making one.

An industry built on queueing math

Contact centers have long been run by formula. The standard staffing models, Erlang C and its descendants, treat a call center as a queue: calls arrive at some rate, each takes some time, and the question is how many agents you need so that callers do not wait too long or hang up. Noah Gans, Ger Koole and Avishai Mandelbaum's widely cited 2003 review of the field walks through these queueing models, from Erlang C to the Erlang A variant that accounts for callers who hang up, and the forecasting and scheduling built on top of them.

The math rewards scale and punishes variance. A Monday-morning spike, a billing error or a storm produces a surge that the schedule did not plan for, and the only levers are overtime, longer waits or turning callers away. The IVR tree exists largely to protect the queue: it deflects what it can and sorts the rest so the right scarce person receives it.

The economics follow from the same constraint. When Gartner forecast in 2022 that conversational AI would cut contact center agent labor costs by $80 billion in 2026, its analyst noted that labor "can represent up to 95 percent of contact center costs." The same release estimated roughly 17 million contact center agents worldwide and only about 1.6% of interactions automated with AI at the time.

What has to happen in a few hundred milliseconds

Replacing a seat with software is not mainly a question of whether a model knows the answer. It is a question of whether it can take part in a conversation at human speed.

Human turn-taking is remarkably fast. Research led by Tanya Stivers across ten languages found that the most common gap between one speaker finishing and the next starting falls between 0 and 200 milliseconds. Stephen Levinson and Francisco Torreira point out why that is striking: the language production system needs more than 600 milliseconds to begin even a single word, so listeners must start preparing their reply while the other person is still talking. A caller notices a slow turn long before they could explain what felt wrong.

The cascade and its seams

Many production voice agents use a cascade. Voice activity detection decides when the caller has stopped, speech recognition turns audio into text, a language model decides what to say and which tools to call, and a synthesis model speaks the answer. Every stage adds delay, and the joins lose information. Kyutai's Moshi paper notes that the complexity of such pipelines "induces a latency of several seconds" and that their segmentation into speaker turns "does not take into account overlapping speech, interruptions and interjections."

Speech-to-speech models attack both problems. OpenAI's GPT-4o system card reports audio responses in as little as 232 milliseconds, with an average of 320. Moshi, a full-duplex model that listens and speaks at the same time, reports a theoretical latency of 160 milliseconds and about 200 in practice. Speed alone is not conversation, though. Benchmarks such as Full-Duplex-Bench now test whether a system stays quiet during a pause, offers a well-timed "mm-hm", takes the turn promptly and yields gracefully when interrupted. Those are the behaviors callers judge, and they are still an active research problem.

The phone line is not a clean microphone

The telephone network takes its own share of the budget. ITU-T Recommendation G.114 treats up to 150 milliseconds of one-way transmission delay as acceptable for most applications, and warns that highly interactive tasks can suffer even below that. Much traditional telephony still carries voice in narrowband codecs such as G.711, sampled at 8 kHz, which discards the higher frequencies present in wideband recordings. Add speakerphones, car noise, accents, hold music and people reading out account numbers, and the gap between a demo and a production line becomes clear.

This is why voice is an engineering discipline rather than a feature. Latency has to be budgeted across network, recognition, reasoning, tool calls and synthesis. Interruptions must cancel speech mid-sentence. Long lookups need something natural to say while they run. Building these systems also overlaps with the broader shift toward models that handle sound, text and images natively, which we look at in why the future of AI is multimodal.

Which work moves to software first

The early movement is not always the conversation itself. Several functions that surround the conversation are moving faster.

Routing. IVR trees ask callers to translate their problem into the company's menu. A system that understands open speech can route on intent, and often resolve the request without a transfer at all. The menu becomes a policy, not a maze.

After-call work. Notes, disposition codes and CRM updates have always consumed agent time between calls. Airbnb researchers reported in 2025 on a production system that drafts notes during support cases and learns from agents' edits. In a difference-in-differences analysis, they measured a 3% reduction in agent working time, rising to 9% for the most complex cases, and agents' manual note-writing time fell from 10% to 3% of their work. The gain is real and modest, which is a useful calibration for vendor claims.

Quality assurance. When every call is transcribed and structured, QA can move from listening to a sample of calls to checking every call against policy: required disclosures, prohibited promises, escalation rules. Humans then review the flagged cases and the rules themselves.

Outbound. Reminders, confirmations, collections and lead follow-up are repetitive and easy to schedule, which makes them attractive to automate. They are also the most heavily regulated calls a company makes, as the next sections show.

Once these functions are software, the conversation layer connects to the systems behind it. The call agent books the appointment, issues the refund or opens the ticket, and the same questions of permissions, policy and audit apply as in any other execution layer for business.

The workforce arithmetic

Customer service is one of the largest occupations exposed to this shift. The US Bureau of Labor Statistics counted about 2.67 million customer service representatives in 2025 and projects employment to decline 5% by 2035, citing automation and self-service. It still expects about 289,500 openings a year, all of them from the need to replace people who leave the occupation. This is a high-turnover job, so change can show up first as fewer replacement hires rather than layoffs.

Forecasters have also become more cautious about the agentless contact center. In June 2025, Gartner predicted that by 2027, half of the organizations that planned to significantly reduce their customer service workforce because of AI would abandon those plans. A poll of 163 service leaders behind that forecast found 95% intended to keep human agents. In September 2025 it went further, predicting that no Fortune 500 company will have fully eliminated human customer service by 2028.

The more likely shape is a smaller, more senior human layer. The calls that reach people will be the harder ones: disputes, distress, exceptions and anything requiring discretion. That changes what an agent's job is and how new agents learn it. We explore what that does to company design in the next company you build may have more AI employees than humans.

The law draws a line between answering and calling

In the United States, the Telephone Consumer Protection Act restricts calls made with an "artificial or prerecorded voice." In February 2024, the Federal Communications Commission confirmed in a declaratory ruling that this covers current AI technologies that generate human voices, so such calls require the called party's prior express consent unless an exemption applies. The ruling was explicit that "the TCPA does not allow for any carve out of technologies that purport to provide the equivalent of a live agent." AI voices are "artificial," it said, "because a person is not speaking them."

Enforcement showed how the telephone stack is involved. After AI-cloned robocalls imitating President Biden's voice reached New Hampshire voters before the January 2024 primary, the FCC fined the consultant behind them $6 million. Lingo Telecom, the carrier that transmitted the calls, agreed to pay $1 million in a consent decree over its caller-ID authentication obligations under the STIR/SHAKEN framework.

The FCC then proposed rules in August 2024 that would define "AI-generated calls," require callers to disclose at the start of each call that AI is being used, and require consent language that covers AI calls. The proposal frames the definition around outbound calls and notes that the TCPA's requirements "do not extend to technologies used to answer inbound calls." As of this writing, those rules remain a proposal. Consent rules are also moving: in January 2025, the Eleventh Circuit vacated the FCC's "one-to-one" consent rule the business day before it was to take effect.

The asymmetry matters for operations. Under the federal robocall rules, answering calls with AI is largely outside the TCPA's reach. Placing them is not. A company automating outbound work has to treat consent records, calling hours, opt-outs and identification as part of the system, not paperwork attached to it.

Europe is moving on disclosure and access to people. Article 50 of the EU AI Act, applicable since 2 August 2026, requires AI systems that interact directly with people to be designed so that they are informed they are dealing with an AI, unless that is obvious. Spain's Law 10/2025 on customer service goes further for large companies and essential services. It bans automated answering systems as the only means of customer service and, on the customer's request, requires personal attention, with 95% of those requests attended on average within three minutes. Companies have twelve months from the law's entry into force at the end of December 2025 to adapt.

None of this is legal advice, and the rules vary by country, state and call type. The design implication is consistent, though: disclosure at the start, consent where calling out, and a working path to a human.

Human takeover is an engineered path

In a seat-based center, escalation means a transfer and a caller repeating themselves. In a software center, it can be designed. The system decides when to hand over based on explicit triggers: the caller asks, the confidence drops, the topic is on a restricted list, or a policy requires a person. The human receives the transcript, what has already been checked, and what the AI was about to do. A supervisor can also join a live call before it goes wrong rather than reviewing it afterward.

This is where accountability lives. Someone must own the rules that decide what the AI may do on a call, review the escalations and change the policy when they see a pattern. Spain's requirement for a fast route to a person is a regulatory version of what good operators would build anyway.

Running a contact center like software

If the call center is becoming software, it inherits software's habits. Prompts, policies and voices change through releases rather than training memos. New versions are tested against recorded and synthetic calls before going live. Latency, interruptions, escalation rates and resolution are monitored like uptime. Every call has a transcript and a record of what the system did.

Syntic's AI Call Center is built on this model, with live monitoring, human takeover, transcripts and per-role permissions. Disclosure, consent and escalation rules remain decisions for each operator and their counsel.

The demand side of the shift, what happens when no call goes unanswered, is the subject of what happens when every customer call can be answered. On the operations side, one thing is already clear: the limiting resource in a contact center is no longer how many people are sitting down. It is how well the operation is designed, tested and supervised.

Sources