We hosted our second webinar last week, this time with the team from Lorikeet on the future of AI-driven customer service. It was a good session and the questions that came through resulted in some excellent insights from the Lorikeet team.
Why this topic, and why Lorikeet
One of the principles we hold strongly at Inlight is composability. When we build digital products and ecosystems for clients, we deliberately choose best-of-breed technology that is modular, API-first, and does one thing exceptionally well. We don't try to build everything ourselves, and we don't recommend platforms that lock clients into a single vendor stack.
AI-driven customer service keeps coming up in conversations with clients across healthcare, financial services, retail, utilities and travel. There is genuine appetite to move, but also real uncertainty about what good actually looks like and how to get there responsibly. Lorikeet caught our attention because they started in the hardest verticals first and built their platform around the assumption that AI needs to take real actions in complex systems, not just answer questions from a set of FAQs. Their platform and capability fits our composability philosophy well, and it made them a natural partner for this conversation.
Where things actually stand
Michelle from Lorikeet opened with an honest assessment of the current state. Despite years of investment and a lot of noise, the majority of customer support interactions are still handled entirely by humans. Most AI deployments today are FAQ-only. The chatbot handles the easy questions and everything else gets bounced to a human queue. The customer ends up re-explaining their problem, waiting on hold, and arriving exactly where they started.
The models are not the bottleneck anymore. The gap is in how AI is being deployed.
Lorikeet recently surveyed over 1,000 consumers globally on their attitudes toward AI in customer service. One finding that surprised people in the room is that 61% of consumers said they are comfortable sharing sensitive information with AI. In areas like health, finance and personal circumstances, people often find it easier to speak to an AI than a human. The barrier to adoption is less about consumer trust than most organisations assume. You can read the full research here: Consumer Attitudes Toward AI Customer Service Report
What the next wave looks like
The session covered three things that are changing what is possible, particularly for organisations in regulated industries.
The first is agents that actually resolve things. As Michelle put it, if an AI can't take action, it is probably just a search bar with manners. The shift from retrieving information to executing actions in core systems such as billing platforms, CRMs, policy engines is what separates a genuinely useful agent from a sophisticated FAQ bot.
The second is the human handoff as a design decision, not a fallback. When AI hands off to a human, that transition should stay inside the same conversation, carry full context, and teach the AI what to do next time. Done well, it also gives organisations a complete audit trail of agent reasoning, something that is genuinely valuable in regulated environments.
The third is voice. It is still early, but latency and naturalness have improved significantly. Bella noted that customers often do not realise they are speaking to an AI, and some thank the voice agent at the end of the call. It is also worth noting that voice is more accessible than chat for many users, older Australians in particular, who may be less comfortable with text-based interfaces, and non-English speakers, since a well-configured voice agent can converse in any language. In industries where the phone is still the primary channel for high-stakes interactions, this is worth paying close attention to.
If AI-driven customer service is on your radar and you are trying to work out where it fits in your broader technology ecosystem, we are happy to have that conversation. Get in touch with the Inlight team.
Check out the research
Lorikeet's consumer research report covers attitudes toward AI in customer service in detail, including the data behind the stats referenced in this session. Read the full report here.
Questions from the session
Thoughts on declaring to a customer that they are speaking to AI at the start of a chat or call, and how to handle it if they ask directly?
Transparency tends to reduce what Lorikeet calls "refusal", which is customers trying to bypass the AI to reach a human. In Australia the rules around disclosure are less clear than in some US states, but Lorikeet leans toward transparency. Something as simple as "I'm an intelligent assistant, I'll see if I can help" is usually enough, especially when paired with an agent that quickly proves it knows something about the customer. When a customer does ask for a human, Lorikeet uses a tactic called "give AI a chance", sharing the real wait time and offering to help in the meantime. Most people give it a try.
Where agents overlap with existing customer portals, are you seeing a shift in how customers prefer to interact?
Not a wholesale replacement, more of a yes and. Customers tend to ask questions wherever they are, and behaviours are ingrained but shifting as AI interactions become more normalised. The bigger point is that conversational AI is built on the same logic and APIs already powering the portal. The ROI case is usually: we have already built the self-serve pathways, now how do we give people who prefer to chat a different way through them.
As agents move from answering questions to taking actions in core systems, how are organisations thinking about approval layers, auditability and accountability?
Lorikeet maps this to how you would onboard a new human team member, starting with read-only access and gradually expanding permissions as trust is established. At a technical level, the platform is designed to speak only to the specific slices of data enabled through each API call, so the scope of what the agent can touch is intentional and auditable from the start.
What approach works best when core systems cannot support increased real-time demand from AI?
Start with the FAQ layer to free up human capacity, then use that breathing room to work on backend resilience. Even when an agent cannot resolve something directly, it can gather everything a human will need, so the handoff is warm and the customer does not repeat themselves. An agent with good knowledge and a clear blueprint for escalation is far more effective than one that can only surface articles.
How do you handle accents and dialects, particularly for Australian users interacting with voice models trained on American data?
Lorikeet makes deliberate choices about which third-party voice models to use for which tasks, and clients can select voices that reflect their user base. Failover mechanisms are built in so that after a couple of failed attempts, the agent acknowledges the issue and offers alternatives, such as switching to keypad input for things like claim numbers, which is often faster anyway.
Are you concerned about social engineering, for example an agent exposing sensitive data based on an unverified input?
The demo simplified authentication for clarity, but a production agent would not give out account information on an email address alone. Lorikeet works within whatever authentication flows a client already uses. More broadly, the platform defaults to not knowing information exists unless it has been explicitly configured to share it, and the instinct to be helpful that general-purpose LLMs have is intentionally engineered against.
