AI calling is changing how customer-service teams handle enquiries, but it is not making the human call centre obsolete. Autonomous voice systems can now interpret requests, retrieve account information, qualify an enquiry and complete a limited transaction. That creates a practical question for service leaders: which conversations should AI handle, and which still need a person?
The answer depends on the issue, the customer, the data available and the risk of getting something wrong. A delivery-status request may be suitable for automation. A mortgage application, debt conversation or complaint about a serious service failure requires more care. A controlled rollout treats AI as an additional service layer, not as a replacement for judgement.
What AI calling can do in customer service
AI voice agents can answer common questions, identify the customer’s intent and collect the details needed for the next step. In a designed workflow, they can:
- Answer frequently asked questions using an approved knowledge base.
- Confirm an order, appointment, balance or delivery status.
- Collect basic information before transferring a caller to an adviser.
- Route enquiries according to topic, urgency, language or customer value.
- Update the CRM with notes, outcomes and the next action.
- Make follow-up calls when timing, permissions and campaign rules allow.
These functions are useful because they reduce repetitive work while keeping the service channel open. AI calling can also provide after-hours coverage for predictable requests, although an out-of-hours message should clearly explain what happens next and how to reach a person when an issue is urgent.
For outbound teams, predictive dialling and AI-supported prioritisation can improve calling capacity. A system may rank contact attempts, select the best time to call and give the agent a relevant reason for making contact. The workflow should use accurate campaign and customer data. A higher call rate cannot compensate for an outdated record, an incorrect consent status or an offer that does not fit the recipient’s situation.
The scale of adoption and the gap in results
Gartner predicts that agentic AI will autonomously resolve 80% of common customer-service issues by 2029. In 2025, 85% of customer-service leaders were exploring, piloting or deploying conversational AI, but only 20% of those projects were fully meeting expectations. The difference shows that model access is only the beginning of an implementation.
AI systems need organised knowledge, accurate records and a clear process for exceptions. They also need testing against the situations that occur in a real contact centre: interruptions, unclear speech, multiple questions, repeated transfers, incorrect authentication and customers who explicitly ask for a person. A conversation that is technically contained is not necessarily resolved. The correct measure is whether the customer’s issue was completed without needing to repeat information or make another call.
Salesforce reports that 94% of its customers choose AI agents when offered the option, and that Agentforce has reduced customer-service costs by $100 million. These are vendor-reported results, and the cost figure was accompanied by workforce redeployment. Businesses should compare their own results against a clear baseline rather than assuming that the same saving will apply automatically.
Why human handoff remains essential
Human handoff is not a sign that an AI project has failed. It is a routing decision. Customers often want a person for mortgages, debt, bereavement, vulnerable-customer situations, disputed charges or decisions with significant financial consequences. An agent can interpret nuance, reassure a caller, negotiate an exception and take responsibility for the outcome.
The handoff must be visible and easy to use. Tell the customer what the AI can do, what it cannot do and how to request an agent. Pass the transcript, intent, account context and authentication status to the human adviser so the customer does not begin again. Measure whether the transfer is smooth and whether the agent can solve the issue at first contact.
For related guidance, see our article on predictive dialling for prioritised outbound contact.
Guardrails should cover more than the model itself. They include knowledge retrieval, approved responses, escalation rules, authentication and exception handling. A chatbot that gives a wrong answer, shows evidence from another customer or continues down an incorrect path can create more work than it removes.
What went wrong in published chatbot examples?
Examples reported in the source show why controls matter. In one case, a BBC reporter asked an Evri chatbot about a missing parcel. The chatbot marked it as delivered, displayed proof linked to the wrong address and provided no useful route to continue. In another case, DPD disabled its chatbot after it criticised the company and swore at users. These incidents may reflect knowledge errors, poor retrieval, inadequate testing or a missing boundary, but the customer impact is similar: trust declines.
Other examples show the risk of rules that are too narrow. A Salesforce agent needed explicit training to express sympathy, while an over-rigid rule designed to block discussion of competitors also blocked legitimate questions about Microsoft Teams integration. A system should apply policy to the risk, not blindly to a keyword.
Before deployment, test each intent with normal, awkward and malicious inputs. Include a human review process for high-impact responses. Restrict access to recordings and customer records, and log important actions so a supervisor can investigate what the system saw and did.
A measured rollout for contact-centre AI
A practical first phase should use bounded, low-risk intents. Select two or three frequent requests with reliable answers, such as opening hours, appointment confirmation or delivery-status guidance. Connect the voice service to the CRM and an approved knowledge source, but do not allow it to invent a policy or make an unrestricted account change.
Set a human route for every important stage. During the pilot, compare the AI group with a baseline group or the previous process. Review transcripts and error categories, not just automation totals. The table below gives a simple operating model:
| Workflow stage | AI responsibility | Human control |
|---|---|---|
| Intent capture | Recognise the request and collect missing details. | Review ambiguous or sensitive cases. |
| Information retrieval | Search approved order or account records. | Authorise exceptions and consequential actions. |
| Resolution | Complete routine, bounded transactions. | Handle complaints, vulnerability and complex advice. |
| Handoff | Pass the transcript and context to an agent. | Confirm identity, explain status and close the issue. |
| Measurement | Log outcome, reason for escalation and resolution status. | Investigate errors and improve the workflow. |
Containment should mean resolved-contact containment, not simply a caller accepting a message or ending the call. Track first-contact resolution, accuracy, average handle time, escalation rate, customer satisfaction, cost per resolved contact, conversion where relevant and compliance incidents. A lower headcount number alone is not a sufficient measure of success.
For related guidance, see our article on cloud telephony options for connected service workflows.
Data, compliance and voice-call controls
Voice automation increases the importance of data governance. The source refers to proposed US legislation that would require disclosure of AI use and a transfer to a human on request, as well as a possible EU right to a human by 2028. These are not universal enacted requirements. Deployment plans should be checked against current local law.
Depending on the market, telemarketing consent, do-not-call rules, calling hours, caller identification and recording consent may apply. US requirements may include the TCPA and state rules. UK and European operations may need to consider UK GDPR, EU GDPR, ePrivacy requirements and local recording-consent law.
Minimise the data collected to what the task requires. Document the lawful basis, restrict access to recordings and CRM records, manage international transfers and retain audit logs. Test responses for biased or unsafe outcomes. Make AI disclosure clear where required, and provide an effective route to a human during the hours customers are served.
Planning the next step
AI calling is most credible as capacity support. It can absorb high-volume, repeatable work, prepare context for advisers and keep routine requests moving after hours. People remain central where trust, sensitivity or commercial value is highest. The goal is not the smallest possible contact centre; it is a service model that resolves more issues accurately while preserving human judgement.
Start with one workflow, document its limits, connect only approved data and define the handoff before launch. Then expand based on evidence from resolution, customer satisfaction, handling time and compliance performance. This approach makes it possible to capture the productivity potential of AI voice without treating automation as an untested promise.
Evaluating AI voice or calling automation? Consider the workflows, data permissions, escalation routes and compliance requirements that apply to your service before selecting a solution.
