From self-service FAQs to scripted chatbots, we’ve been deploying forms of AI in the contact centre for well over a decade. The more recent development of Generative AI saw breathless predictions of entirely automated bots handling most, if not all, customer interactions.
AI is already replacing a substantial amount of routine customer service work. IKEA is a well-known example: its Billie chatbot resolved about 47% of customer enquiries between 2021 and 2023. Rather than simply cut the jobs affected, IKEA reskilled 8,500 call-centre co-workers in remote interior design, digital sales, relationship building and more complex customer enquiries.
That kind of redeployment is becoming part of the AI workforce story. Gartner found 85% of service and support leaders were expanding human agent responsibilities as AI reduced contact volume, while 75% were shifting agents into entirely new roles. AI also creates work of its own: organisations need people to maintain knowledge, monitor outcomes, identify deficiencies and improve the systems. Gartner separately found 58% of service leaders planned to upskill agents into knowledge-management specialist roles.
Outcomes matter more than the technology
Having said that, AI undoubtedly has its place in the modern contact centre. It can help customers find answers, handle telephone enquiries, guide employees through difficult conversations and identify why people keep contacting the business. Some systems can also take action, from changing a booking to processing a refund.
Those capabilities solve different problems. Buying a conversational interface will achieve little if customers still cannot complete a transaction or reach someone who can help.
We should not lose sight of the fact that outcomes are what matter. Qualtrics’ global research across 14 countries, including Australia, found that nearly one in five surveyed consumers who had used AI for customer service reported no benefit.
The aim should be to combine AI and people in ways that solve the customer’s problem well. Understanding where each technology fits is the first buying decision.
For a CX leader, that means separating the business case for helping employees from the case for handing decisions to software. Each needs different evidence and controls.
What does AI in customer service actually include?
Customer-facing AI
AI-enabled self-service helps people find information or complete tasks themselves. Conversational AI lets them explain requests in everyday language through chat. Voice AI brings similar capabilities to telephone interactions.
These categories overlap. A customer might describe a problem to a voice assistant, complete an identity check and then use an existing automated transaction. The whole process does not need to be AI-driven.
AI behind and alongside the conversation
Agent Assist supports employees during and after interactions. Interaction analytics examines conversations for patterns. Autonomous or agentic systems can select and execute actions within defined permissions.
Customers can also bring their own AI assistants into the relationship, creating another channel for businesses to consider.
The distinction matters when choosing suppliers. A platform that handles language well may still need integration, knowledge-management and service-design work before it can deliver a complete customer journey.
Self-service and conversational AI: judge them by resolution
Self-service is broader than chatbots
Intelligent search, dynamic help, personalised portals and transactional tools can all support self-service. Sometimes making an existing function easier to find achieves more than introducing another bot.
Measure what happens next. Industry usage varies, but deflection generally means redirecting demand away from human assistance. Containment means the interaction stays within automation. Resolution means the underlying problem is solved. A customer who abandons a bot and calls tomorrow has not been successfully served.
Modern conversational AI can do more than answer FAQs
Scripted bots follow predefined paths. Natural-language bots recognise requests, generative assistants compose responses, and action-taking agents can use connected systems to complete tasks. These capabilities may coexist in one service.
Ask whether the system uses reliable knowledge, retains context, completes the request and hands over properly when it cannot. Fluent language alone does not establish accuracy. Check repeat contacts and completed transactions alongside the bot’s own success reports.
Check where people enter self-service, which tasks they abandon and whether they can see the next step. Poor navigation and unclear instructions are service-design problems, even when the underlying technology works.
Voice AI is changing the phone channel, not eliminating it
Voice remains important
ContactBabel’s 2026–27 Australia and New Zealand research found that phone calls still account for 64.3% of inbound customer interactions in Australia. Australian consumers also continue to favour the phone for urgent enquiries, so a digital service strategy still needs a dependable telephone experience.
What Voice AI can actually do
Voice AI can guide authentication, identify intent, route calls and handle bookings, routine account servicing or transactions. Reminders and outbound notifications are other potential applications.
Evaluate it with real callers. Can it cope with Australian accents, background noise and interruptions? Does response delay make people speak over it? Can callers correct a mistake?
Authentication, fraud controls and clear disclosure deserve attention. So does vulnerability. A distressed caller needs an appropriate route to a person, even when the stated request sounds routine.
Start with a defined call type and test both successful completion and recovery when something goes wrong. A convincing demonstration in a quiet room with a cooperative speaker is insufficient.
Agent Assist may be AI’s strongest near-term contact centre use case
AI helping the agent rather than replacing the agent
Agent Assist can retrieve knowledge, suggest responses and next actions, summarise customer history, prompt compliance checks and support translation. Afterwards, it can draft call summaries and reduce administrative work.
Employees should be able to inspect the supporting information, correct suggestions and retain control of the conversation.
Measure quality as well as speed. Incorrect summaries or confidently suggested policy exceptions can create work for the next employee.
The productivity evidence is unusually strong
The NBER working paper Generative AI at Work followed more than 5,000 support agents and reported a 14% average productivity increase, measured through issues resolved per hour.
The final, peer-reviewed Quarterly Journal of Economics paper used a refined sample of 5,172 agents and reported approximately 15%. These are versions of the same underlying research, not independent studies. Gains were greatest among less-experienced and lower-skilled employees. The setting was chat support for one software company serving US businesses, so the result should inform a pilot rather than become a universal savings assumption.
The human role is changing
Where automation removes routine work, human queues can contain more exceptions, disputes and emotionally difficult cases. Agents need time, judgement and authority to resolve them, particularly when customers are vulnerable. Workforce planning and coaching should reflect that changing mix. Holding employees to unchanged handling-time targets could punish them for taking on harder work.
Interaction analytics: using every conversation to improve the business
From sampled QA to conversation intelligence
Interaction analytics can examine conversations far beyond the small samples available to manual quality teams. It can identify contact reasons, customer intent, complaints and recurring problems, while flagging possible compliance issues, churn or vulnerability for review. Coverage depends on the channels and recordings available. Automated classifications still need checking.
The biggest value may sit outside the contact centre
ContactBabel’s 2026–27 ANZ research found that 37% of Australian speech-analytics users plan to replace or upgrade their current solution. That suggests analytics is already established, but many organisations are reassessing what the technology can deliver. Its usefulness can extend beyond quality assurance to identifying recurring reasons for contact and problems elsewhere in the customer journey.
If thousands of customers call because a bill is confusing, helping agents explain it faster addresses only part of the problem. Fixing the bill could remove the need to call. Give each recurring issue an owner and check whether changes reduce subsequent demand.
Use findings to improve knowledge and coaching too. Validate important patterns against actual conversations before turning them into changes to policy or customer treatment.
Agentic AI: when the system starts taking action
Chatbot, assistant, workflow or genuine agent?
An assistant answers or recommends. A workflow follows predetermined steps. An agent selects from available tools and actions. A genuinely agentic system can work out steps towards a goal, act, inspect the results and adjust its approach.
Useful automation does not require all that autonomy. A fixed refund workflow may be exactly what a business needs.
The market is ahead of itself
ContactBabel’s 2026–27 ANZ research found that 40% of Australian chatbot users plan to replace or upgrade their current solution, even as chatbots remain the leading investment priority for the next 12 months, cited by 36% of respondents. Buyers need demonstrations of specific capabilities behind the product label.
What action-taking AI looks like
Salesforce reports that Telstra uses Agentforce across interim-service, refund and security workflows, supporting more than 1,000 contact-centre representatives. Its interim-service workflow checks eligibility, calculates refunds and updates records. This illustrates action-taking capability, with some steps guided by workflow. It is supplier-reported evidence, not an independent ROI assessment or proof of unrestricted autonomy.
The decisive question is: what is the AI authorised to do without asking a person?
Set transaction limits and approval requirements, and record actions. A system that issues credits needs a controlled response when eligibility is unclear or a transaction fails halfway through.
Human handoff still matters
Customers want AI and humans
In Gartner’s 2026 survey of 3,566 business and consumer customers, 50% said GenAI made service interactions easier. Yet 87% considered access to a human essential when companies used it for service.
A handoff should preserve the work already done
Pass identity and authentication status, customer intent, previous steps and relevant context to the receiving agent, subject to appropriate security controls. The customer should not have to explain everything twice.
Build escalation around vulnerability, emotional distress, disputes, exceptions and consequential decisions. Uncertain answers or repeated failed attempts should also trigger help. Human access should be visible before the customer becomes frustrated.
Treat transfer as part of the same service journey. Measure whether the receiving team can finish the task, rather than counting the transfer itself as success.
The next customer may be an AI agent
Customers are already starting outside company channels
In the same Gartner survey, respondents were approximately three times more likely to have used third-party GenAI than a company chatbot during their latest service interaction. Among GenAI users, 58% said they had used it to complete a task on their behalf. That signals demand for assistance, without establishing widespread autonomous negotiation.
Agents can increasingly communicate and transact
These are recent standards. A2A, announced in April 2025, supports communication between agents. Google’s Agent Payments Protocol, announced in September 2025, addresses authorisation for agent-led payments. Its Universal Commerce Protocol, announced in January 2026, supports commerce interactions. Stripe and Tempo’s Machine Payments Protocol, launched in March 2026, enables programmatic payments between agents and services.
Why CX leaders should care
If a customer’s AI can buy from you, can it also negotiate, reschedule, complain, request a refund, dispute a charge or cancel the service?
This remains an emerging frontier. Buyers should ask how suppliers will verify delegated authority, limit actions and preserve a record for disputes.
For now, make service information accurate and policies unambiguous. These improvements help whether the next enquiry comes from a person or software acting for them.
What needs to be in place before you buy AI for CX?
Knowledge, data and integration
AI cannot reliably compensate for bad knowledge, fragmented data or broken processes. Establish who maintains approved answers, which customer context is available and which integrations allow transactions. Define permissions, workflow rules and escalation paths before granting access.
Privacy and governance
OAIC guidance recommends privacy impact assessment and scrutiny of how AI products handle personal information. Understand processing locations, retention, third-party access and model-training use.
From 10 December 2026, APP 1.7 requires covered entities to include additional information in their privacy policies where they have arranged for a computer program to make, or do something substantially and directly related to making, a decision that could reasonably be expected to significantly affect an individual’s rights or interests, and personal information is used in that process. Human involvement does not automatically remove that obligation.
Ask suppliers what the system can actually do
- Which customer problem will it solve?
- Does it answer, recommend or act?
- Which systems can it read and change?
- What knowledge supports its responses, and who updates it?
- When does it escalate, and what context transfers?
- What personal information does it use, retain or share?
- How are errors detected, corrected and audited?
- How will we measure resolution and ROI, including implementation and ongoing costs?
Require a demonstration using your own representative cases, including failures and exceptions.
Start a pilot with a baseline for resolution, repeat contact, effort and cost. Include human escalation and ongoing maintenance when calculating returns.
The right AI depends on the work
A straightforward enquiry may suit self-service or a conversational assistant. A routine telephone transaction may suit Voice AI. A disputed charge may need an experienced employee with Agent Assist, while analytics identifies why similar disputes recur. Autonomous agents belong where permissions and controls support their actions.
The goal is better resolution, less unnecessary customer effort and an operating model where AI and people each do the work they are best suited to do.
If you are evaluating conversational AI, Voice AI, Agent Assist, interaction analytics, agentic technology or CX implementation support, Matchboard can provide a shortlist technology vendors with the relevant experience and capabilities.
