Customer service AI still needs clear routes to human help

Customers need control over when AI stops

Customers are already trying to get past the AI systems businesses put in their way. Twilio research found that 78% of consumers have tried to bypass an AI agent to reach a human. For teams choosing customer experience technology, that behavior makes customer control a basic deployment requirement. An agent needs useful context, a clear route to human help, and a record of what happened during the interaction.

The reasons for seeking a person explain those requirements. Jesse Sumrak, writing on Twilio’s company blog, identified three drivers: uncertainty about whether the customer is talking to a bot, discomfort with automation in high-stakes situations such as medical inquiries and payments, and failure to resolve the problem. Twilio also found that 80% of consumers report a negative AI experience. Its research puts the time an unsuccessful agent consumes before escalation at an average of 81 seconds, adding delay to an unresolved issue.

So customers need an easy way to end automation. Twilio found that 63% want an easy route to a human at any point in a conversation, while 47% of brands provide one. The company recommends identifying the AI at the start, making escalation visible, and passing the conversation’s history and context to the human. Twilio sells communications technology and connects these recommendations to its Conversation Orchestrator and Conversation Memory products, giving it a commercial interest in adoption.

Customer control becomes more important as agents gain permission to change things. During one week in late August 2026, more than a dozen vendors across contact centers and the wider customer experience technology stack announced developments through earnings calls, sports partnerships, products, consumer research, and monitoring tools. The announcements cover both agent execution and supervision, but they address different operating problems. Buyers need to establish what an agent can do, when a person can take over, and how the business will examine the result.

Agents are taking different kinds of action

The first distinction is between supplying information and completing a transaction. Agentic customer experience means software takes steps on a customer’s behalf, such as changing a booking or correcting a ticket. Those steps change the customer’s situation, making the business responsible for the decision and its consequences. Testing therefore has to cover the completed action as well as the conversation that led to it.

Dialpad’s multi-year agreement with the Miami Marlins illustrates that responsibility. As the team’s “Proud AI-Powered Communication Partner,” Dialpad is deploying its Agentic AI platform for fan communications and in-stadium support at loanDepot Park. Proposed actions include reissuing tickets and changing seats, aiming to solve a problem before a frustrated fan leaves. Dialpad, which sells the platform and benefits from demonstrating its value in sports venues, says Austin FC already uses it to handle routine fan requests immediately.

Speed also has to serve the relationship the team wants with its fans. Craig Walker, Dialpad’s co-founder and CEO, described fans as expecting recognition, attention, and quick help when calling about tickets or asking for assistance at the gate. Ruben Smith, enterprise technology strategist for the Marlins, described the communications change as happening “without losing the personal connection.” Their expectations give the deployment two tests: whether it completes the requested transaction and whether the interaction maintains that connection.

Other fan-facing announcements address a different kind of work. For the 2026 US Open, IBM and the United States Tennis Association introduced Serve Quality, a metric using 21 data points across a player’s body and racquet, captured 50 times per second. Their enhanced Match Chat assistant adds photos and video to text responses. These capabilities support large-scale digital engagement through analysis and richer answers, so they need different performance measures from ticket reissuing or seat changes.

Producing those answers still requires substantial processing behind the scenes. Jonathan Adashek, IBM’s senior vice president and chief global affairs officer, described automated workflows turning millions of court data points into usable insights within minutes and hours. IBM supplies the technology and benefits commercially from demonstrating that capability; Brian Ryerson, the USTA’s senior director of digital strategy, presented the collaboration as a major advance in digital fan engagement. Buyers should distinguish analytical output, multimedia assistance, and transactional execution because each has a different definition of success.

Easier agent creation still requires production integration

Once the required capability is clear, the next question is who can build it. Service teams may know the work well but lack specialist AI staff, making development a barrier to deployment. 8×8 sells agent-building tools and has a commercial interest in lowering that barrier. Its customer accounts describe plain-language tools that let service teams create agents directly.

According to 8×8, customer interactions built on 8×8 AI Studio grew nearly 450% since April, with no AI specialists among the teams responsible for that growth. Customers including PrepayPower and LSH Auto describe building working agents through ordinary language without starting a separate IT project. This lets teams express their workflows directly, reducing the need to translate requirements into a specialist development brief. The growth figure measures platform use; customer resolution remains a separate outcome to assess.

Interest in custom agents extends beyond these non-specialist teams. Metrigy analyst Irwin Lazar cited research finding that nearly 72% of organizations prefer agents built internally or with a partner over off-the-shelf products. Lazar attributed that preference to trust and the need for company-specific knowledge. The finding helps explain demand for customization, while 8×8’s examples show how some teams can do it.

But building the agent is only the beginning of the work needed to run it. Hunter Middleton, chief product officer at 8×8, said, “Everyone has seen the AI demo,” and continued: “Far fewer have seen AI survive contact with production, because between the two sits an integration project most organizations were never staffed to run. AI that demos well was never the hard part. AI that holds up in production, at this pace and across this many industries, is.” His warning identifies the gap between expressing a workflow and connecting it to the systems it needs to run reliably.

Deployment conditions are part of that integration work. 8×8 says its newly included provisions cover GDPR compliance, referring to the EU’s General Data Protection Regulation, and HIPAA-readiness, concerning US health-information requirements. A team can describe an agent’s behavior in plain language and still need an acceptable environment for the information it uses. The business must decide what data the agent may access, which actions it may perform, and who will supervise it.

Agents need access to the right data

Those access decisions become more demanding as deployment expands across an enterprise. Verizon says Gemini Enterprise now handles most of its incoming consumer calls and chats each month. Its expanded partnership with Google Cloud follows several years of consolidating legacy data lakes, or large stores of data, into Google’s Agentic Data Cloud. Verizon’s reported scale therefore follows a substantial investment in its data environment.

The partnership is extending that environment into network operations, marketing automation, and cloud security. Alfonso Villanueva, Verizon’s chief transformation officer and EVP of Verizon Consumer, links personalized service to an AI-first approach across the organization. Karthik Narain, Google Cloud’s chief product and business officer, presents Verizon as an example of enterprise-wide AI transformation. Verizon is describing the value of its own investment, while Google Cloud sells the underlying technology; both have an interest in presenting the deployment as successful.

The deployment is described as among telecom’s largest enterprise AI implementations, but its interaction volume answers a specific question. Handling a call or chat establishes that the system participated in the interaction. Autonomous resolution requires evidence that it completed the customer’s task. Buyers evaluating similar deployments need to define which outcome their volume measure represents.

Even a consolidated data environment must let an agent find the particular record it needs for its next action. Zendesk’s addition of custom object record search inside action flows addresses that dependency. An action flow is a sequence of operations for completing a task; the update lets it search related orders, contracts, returns, and entitlements without already holding the record’s ID. Zendesk sells customer service software and benefits from making more of those workflows executable within its platform.

Finding the record determines whether the workflow can continue. A returns process may need the relevant order before deciding the next step, while an entitlement check needs the applicable record before a service decision can proceed. Zendesk’s update moves that lookup into the workflow, reducing the need for a person to retrieve it. The practical gain is access to the right object when an action depends on it.

Operating evidence must follow the customer across systems

After the agent retrieves information and acts, the business needs a connected account of the interaction. Observability means collecting information about system behavior so teams can understand events and diagnose failures. Operata sells this capability through its expanded CX AI Agent Observability offering, covering telephony, contact-center-as-a-service systems, customer relationship management software, and AI platforms. Contact-center-as-a-service systems provide cloud-hosted contact-center functions; Operata brings their metrics, execution traces showing the steps a system took, and logs together with records from the other systems.

Combining those records matters because a successful event in one application can sit inside a failed customer interaction. A transfer may complete technically while leaving the human agent without the conversation’s context. An announcement may register as played while remaining inaudible to the customer. Operata’s commercial proposition is to preserve evidence across the journey so operations teams can investigate those customer-level failures.

AI disclosure makes the difference between a logged event and a customer outcome especially important. Operata positions its offering around Article 50 of the EU AI Act, describing the disclosure requirement for AI systems that interact with people as applicable since August. In Operata’s framing, the evidence should establish that the customer could hear the disclosure and that a transfer preserved context. A playback event records execution; evidence of audibility addresses whether the customer received the information.

Operata also presents this evidence as a way to improve shared human and AI work. Romilly Blackburn, the company’s co-founder and CEO, said: “At Operata, we believe human judgment and AI execution together deliver the best end-to-end customer experience.” Blackburn describes observability as supporting monitoring, assurance, diagnosis, and improvement across both. For the escalation requirement identified by Twilio, this means checking whether the transfer worked and whether the person receiving it could continue the conversation.

Automation can create more room for human service

A successful handoff still leaves a broader design decision: how the business uses the staff capacity technology frees up. A contact center can absorb growth, expand personal outreach, or support employees through difficult conversations. Christian Healthcare Ministries provides an example of capacity redirected toward member relationships. Its reported results come from routing and operational automation, with AI self-service considered as a later extension.

In Five9’s account, the starting problem was preparing for a tenfold increase in membership. In a Five9 customer story written by Julie LeWarne, CHM, a faith-based health cost-sharing organization, had a 10-option phone menu that directed excessive traffic into an operator queue and left employees with substantial manual work. Wayne Pierce, CHM’s contact-center-as-a-service manager, described the need to manage incoming demand more effectively. The membership increase was a planning requirement.

To meet that requirement, CHM worked with integration partner Waterfield Tech to move to the Five9 Intelligent CX Platform. Five9 sells the platform and has a commercial interest in the customer story’s results. According to its account, CHM’s call transfer rate fell from 48% to 4%, a reduction of 44 percentage points. Intelligent routing identified roughly 80% of incoming calls before an employee answered, directing more calls before staff had to intervene.

A separate change improved collection work. Preview dialing, which lets an employee review information before placing a call, helped shorten late-payer collection cycles from 22 days to six. Five9 reports that this contributed to a 400% increase in collections. These figures describe the deployment’s process and business outcomes, providing evidence for the routing and dialing changes involved.

The resulting capacity also allowed CHM to expand personal contact. Employees began making welcome calls to every new member and prayer calls to people facing difficult health circumstances. “Our members just love this,” Pierce said. “They feel loved.” CHM’s decision to use available time for outreach turned an operational change into a different member experience.

That experience shaped CHM’s consideration of AI-powered self-service beyond its existing business hours. The proposed extension would need its own tests of access, escalation, and resolution. The reported routing, dialing, and outreach gains belong to the completed deployment; after-hours self-service remained a future decision.

Human service can also improve through assistance during a live conversation. Verint reports that an unnamed financial institution deployed real-time, empathy-led coaching across 6,500 agents and estimates $28 million in annual value. The system detects emotional signals and prompts more consistent responses during stressful calls involving deployments, family financial needs, and government disruptions. The financial figure is the customer’s estimate of annual value.

Verint, which sells the coaching technology and benefits from demonstrating its value, also reports results from one high-volume scenario. It says the system provided guidance on roughly 160,000 calls and generated more than 1 million positive-reinforcement notifications. Teresa Anania, Verint’s chief customer officer, attributes better interactions to relevant insight and support reaching employees at the right moment. Here the operating question is whether coaching helps the person responsible for the conversation respond effectively.

Procurement must establish value and acceptable operating conditions

These different outcomes require different purchasing evidence. A reduction in call transfers measures an operating change, an annual-value estimate expresses expected financial benefit, and platform-use growth measures adoption. Buyers first need to identify the outcome they are purchasing and the status of the evidence behind it. Compliance-sensitive organizations must also establish acceptable deployment conditions before acting on a commercial benefit.

Omilia’s appointment of Michael Weiss as vice president for Germany, Austria, and Switzerland brings those conditions into focus. Weiss, a four-year Five9 veteran with an enterprise sales background, is tasked with developing a market where banks and insurers remain cautious about replacing legacy interactive voice response menus and first-generation automation. Interactive voice response menus route callers through recorded choices. Omilia sells the replacement technology, so Weiss’s assessment also describes requirements the company must satisfy to win business.

Weiss stated those requirements directly: “The caution isn’t reluctance, it’s a need for proof of ROI and sovereign cloud.” ROI means return on investment, while sovereign cloud here concerns buyers’ requirements for jurisdiction and control over the cloud environment and its data. Nick Delis, Omilia’s chief revenue officer, describes DACH, the Germany, Austria, and Switzerland market, as having demanding compliance obligations and long enterprise buying cycles. He also describes customer relationships as durable once a supplier is established.

These conditions make the business case and the operating environment separate procurement decisions. Buyers need to identify the records a deployment will access, where it will run, and which sovereignty obligations apply. A promising financial return still leaves those requirements to be met. Suppliers therefore need to show both how the deployment produces value and how it fits the buyer’s permitted environment.

The implementation barrier will also vary by organization. The earlier 8×8 examples concern teams that need a simpler way to create an agent, while Weiss describes buyers seeking proof of returns and acceptable sovereign-cloud arrangements. Procurement should identify which constraint is delaying a decision before choosing a remedy. Easier creation, a stronger business case, and acceptable deployment conditions solve different problems.

Businesses must also serve agents acting for customers

An acceptable operating environment must account for agents outside the business’s control as well. BrightEdge’s Agent Edge launch addresses agents that research, compare, and take action for people before those people reach a brand’s website. BrightEdge calls this activity the “Second Web.” The company sells tools for improving brands’ visibility and agent access, giving it a commercial interest in businesses preparing for this shift.

BrightEdge found that 97% of the top 20,000 websites were unprepared to provide information in formats optimized for agents. To interpret human-oriented HTML, the markup used to structure web pages, and navigation, agents must spend tokens, the units of text processed by language models, on material designed for people. Agent Edge supplies content adapted for agent consumption. BrightEdge places this work within generative engine optimization, which aims to improve how brands are found through generative AI services.

Early deployments provide measures of delivery efficiency and discovery. According to BrightEdge, chip designer Arm reduced the payload, the amount of data delivered to agents, by roughly 90% and more than doubled verified agent traffic within weeks. Bloomfire’s Brian Zander reported a 30% increase in direct AI referral traffic as an early Agent Edge customer. Jim Yu, BrightEdge’s CEO and co-founder, argues that understanding and improving these interactions will become more valuable as agents perform more work for users.

Traffic and payload measures help teams assess whether an external agent can reach and process business information. When that agent can also take consequential action, the business must preserve the customer’s authority over the transaction. Marketing and engineering teams therefore need to establish what information the intermediary received, what permission governed its action, and what happened when it acted.

The bottom line

As AI agents move from answering questions to changing customer records and completing transactions, deployment decisions carry greater operational responsibility. The question for leaders is which tasks an agent can complete reliably, within what limits, and with what route to human help. Those conditions should shape the purchase before interaction volume or adoption becomes the measure of success.

The evidence needs to match the intended outcome. A system handling more calls has demonstrated scale, not necessarily resolution. An agent that completes a transaction still needs checks on whether it used the right record and acted within its permissions. A successful transfer must leave the human employee with enough context to continue. These are separate tests, and each needs a clear owner.

Leaders should therefore connect spending approval to evidence from the full customer interaction. That includes production integration, data access, deployment requirements, and records that let teams investigate failures across systems. Easier agent creation can shorten development, but it does not remove these responsibilities.

The business must also decide how to use the capacity automation creates. CHM’s reported expansion of personal outreach shows one option beyond absorbing more demand. The broader opportunity is to make routine service easier while preserving human attention where it matters most. Customers need control over when automation stops; businesses need evidence that the work it completes is worth doing.