Last February, a fintech CX leader I know sat across from her board and presented a slide titled "AI-First Support: Year One Results." The company had deployed one of the marquee AI chatbot vendors twelve months earlier. Deflection rate: 48%. Headline cost savings: $1.3 million. But then she showed the next slide. Customer satisfaction had dropped eleven points. Escalation-to-resolution time had increased because agents receiving handoffs had zero context. Repeat contacts on the same issue were up 34%. The board asked one question: what are we doing about this?

She is not an outlier. She is the new normal.

The graduation cohort.

Most AI customer service vendors are still selling to first-time buyers: companies that have never deployed AI for support and are dazzled by a slick demo. But a growing segment of the market has already bought, deployed, measured, and been disappointed. They tried Zendesk AI, or Fin, or Ada, or a custom build wired to OpenAI. They got a deflection engine dressed up as automation. Now they are shopping for their second platform.

This cohort is different. They have scar tissue. They know what a vendor demo hides. They know the difference between a contained conversation and a resolved one. And they are asking harder questions than they did the first time around.

Gartner found that only 8% of customers used a chatbot in their most recent service interaction, and just 25% of those said they would use one again. That is not a technology adoption problem. That is a product quality problem created by the first generation of tools.

Deflection is not resolution.

The original sin of first-gen chatbots was measuring success by how many conversations they kept away from human agents. Deflection. Containment. The metric itself reveals the design philosophy: the AI exists to protect your team from customers, not to help customers.

This worked just well enough to sell. A VP could show the CFO a deflection rate and claim savings. But the savings were an illusion. Gartner research found that only 1 in 7 customer service queries are fully resolved through self-service. Even for issues described as "very simple," only 36% were handled without a human. The rest bounced back into the queue, often angrier than when they started.

The distinction between deflection and resolution is the single most important concept for second-time buyers to internalize. Deflection asks: did we avoid a human interaction? Resolution asks: did the customer's problem get solved?

When you evaluate your next platform, open the reporting dashboard before anything else. If you cannot see resolution rate and deflection rate as separate, distinct metrics, walk away. The vendor has already told you what they optimize for.

The Klarna warning.

Klarna became the poster child for AI-first customer support in 2024, announcing its AI could replace 700 agents and handle 2.3 million conversations per month. By mid-2025, the company was rehiring human agents after customer satisfaction cratered.

CEO Sebastian Siemiatkowski told Bloomberg the company had gone "too far in the wrong direction." His diagnosis: "Cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality."

Klarna's mistake was not adopting AI. It was adopting AI that could only do one thing: answer questions from a knowledge base. When customers needed refunds processed, account details changed, or billing disputes investigated, the bot could not act. It could only talk. Customers received what CNBC described as "generic, repetitive, and insufficiently nuanced replies" and flooded complaint channels.

The lesson is structural, not cautionary. First-gen tools were built to summarize FAQs and route tickets. That architecture has a ceiling, and most companies hit it within twelve months.

What broke.

Across conversations with dozens of CX leaders who are on their second or third AI vendor, the failure modes cluster into five patterns.

No memory across conversations. Each ticket is treated as a standalone event. The AI does not know the customer asked about the same issue last week, was promised a callback, or has contacted support four times in ten days. 74% of customers expect bots to remember past interactions, but only 28% say this matches their experience.

No ability to take action. The bot can tell a customer what the refund policy is. It cannot process the refund. It can explain how to update billing information. It cannot update it. For anything requiring a write operation to a backend system, the conversation dead-ends with "let me transfer you to an agent."

Broken handoffs. When the AI does escalate, the human agent receives a ticket with minimal or no context. The customer repeats everything. The agent has no visibility into what the AI already tried. The experience feels worse than no AI at all, because the customer already invested time explaining the issue once.

Brittle at the edges. The demo worked on the ten most common queries. But real support traffic includes billing disputes (where Gartner found only 17% chatbot resolution), multi-step workflows, regulatory edge cases, and emotional customers. First-gen tools handle the easy middle and collapse on everything else.

Impossible to maintain. One prospect we spoke with described their custom OpenAI integration as "finicky," requiring constant maintenance every time documentation changed or a workflow needed updating. Another called their legacy chatbot setup "a house of cards." When the underlying model updates, the behavior shifts unpredictably, and there is no testing infrastructure to catch regressions before they hit customers.

The evaluation playbook.

Second-time buyers have earned a more sophisticated evaluation framework. Here is what the best ones are doing.

Test with your worst tickets, not your best. Pull 200 real tickets from your queue, including the messy ones, the multi-turn conversations, the policy-ambiguous cases, and the edge cases that trip up your human agents. Any vendor confident in their platform will welcome this. If they insist on using their own demo data, that tells you everything.

Trigger an escalation during the demo. Then switch to the agent interface. Did the human agent get full context? Can they see the entire conversation thread? Do they know what the AI already tried and what the customer already said? The handoff is where most platforms disintegrate.

Ask to see a write operation. Not a knowledge base lookup. Not a canned response. Ask the AI to process a refund, update an account, cancel and rebook a reservation. If the vendor hesitates or says "that is on our roadmap," you are looking at another deflection engine.

Check the analytics for resolution, not containment. Can the platform tell you what percentage of conversations ended with the customer's problem actually solved, verified by outcome data rather than assumed by conversation closure? Can it show you which intents are rising, where resolution fails, and which workflows cause repeat contacts?

Ask about the knowledge management burden. 61% of support leaders report backlogs in editing outdated knowledge articles, and that directly limits AI effectiveness. How does the platform handle knowledge that is stale, conflicting, or missing? Does it fail silently, or does it surface the gap?

Resolution means action.

The fundamental architectural difference between first-gen chatbots and actual AI agents is the ability to take action in the real world.

A chatbot reads from your systems. An AI agent reads and writes. It processes the refund, updates the address, escalates with full context to an internal team, and follows up to confirm the issue was resolved. The AI agent market is growing at 46.3% CAGR, nearly double the chatbot market's 23% growth rate, because buyers have learned that answering questions is table stakes and solving problems is the actual product.

This is not a marginal upgrade. It is a category shift. When your AI can take action, the economics change. The customer gets a resolved issue in minutes instead of hours. The human agent team handles genuinely complex cases instead of re-doing work the bot could not finish. And the cost per resolution drops because you are not paying twice for the same interaction.

Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues by 2029. The companies already operating this way are not waiting for 2029. They are building the muscle now.

What Lorikeet gets right.

We built Lorikeet for this exact buyer: the CX leader who has already been burned and is not interested in another deflection engine.

Lorikeet is an AI customer support platform that resolves tickets end-to-end across chat, email, and voice. That means processing refunds, updating accounts, coordinating with third parties, and handling complex multi-step workflows. Not summarizing FAQs. Not containing conversations. Resolving them.

Three things matter to second-time buyers, and they are all core to how we built the product.

Resolution over deflection. Lorikeet measures and optimizes for end-to-end resolution. The agent does not succeed by avoiding a human interaction. It succeeds when the customer's problem is actually solved. That is a fundamentally different design philosophy than containment-first platforms, and it shows up in how the product behaves with edge cases, complex queries, and multi-turn conversations.

Action-oriented architecture. Lorikeet agents take action: calling APIs, processing transactions, updating records, and coordinating with internal teams. Our Team of Agents capability lets a primary agent spawn secondary agents that contact third parties, investigate issues, and report back to resolve the customer's problem entirely. A customer reports a missing delivery. Lorikeet contacts the logistics provider, gets the status, and reports back. Minutes, not days.

Context that survives the handoff. When Lorikeet does escalate to a human, the agent gets the full conversation, every action the AI took, and a summary of what was tried and what remains. No customer repetition. No blind handoff. The human picks up where the AI left off, with complete context.

We work with companies in fintech, healthcare, and other regulated industries where getting the answer wrong has real consequences. That is a forcing function for quality that FAQ-based systems never face.

The second purchase is smarter.

The first time a company buys AI for customer support, they buy the vision: fewer tickets, lower costs, happier customers. The second time, they buy the architecture: what can it actually do, how does it handle failure, and how does it prove it worked?

64% of customers say they would prefer companies did not use AI for customer service. That number is not a rejection of AI. It is a rejection of bad AI. It is the accumulated frustration of millions of people who were deflected, misunderstood, and forced to repeat themselves by first-gen tools that measured success by how many conversations they avoided.

The companies that win the next phase of AI customer service will not be the ones with the most sophisticated language models or the largest knowledge bases. They will be the ones whose AI can actually solve problems. Take action. Remember context. Prove resolution.

If you are shopping for your second platform, you already know this. Trust what you learned. Demand more than deflection. And test everything with your hardest tickets, not your easiest ones.

See how Lorikeet handles the tickets your last platform could not.