AI for Customer Support: When a Custom Agent Beats a Chatbot

AI for Customer Support: When a Custom Agent Beats a Chatbot

Every customer-service vendor now sells an "AI agent." Intercom's is called Fin, Tidio's is Lyro, and in 2026 Zendesk repositioned its entire platform around resolution rather than replies [2][4]. The word "chatbot," meanwhile, has become a liability — for buyers it signals rigid, scripted flows and the frustrating loops that make customers mash "agent, agent, agent" [1]. But the branding war hides the decision that actually matters. The real question for a support leader in the consideration stage is not chatbot versus agent as marketing labels; it is when a custom, action-taking AI agent genuinely outperforms a scripted chatbot — and, just as importantly, when it does not. This guide draws the line with the numbers, the architecture, and the single most instructive deployment in the field, so you can judge which one your support operation actually needs.

The distinction that matters: deflect versus resolve

Strip away the vocabulary and there is one functional difference. A chatbot answers; an agent acts. Picture a customer who writes, "I received the wrong color, I want a refund." A chatbot replies with the returns-policy link and a form. A custom AI agent reads the order, verifies the item, checks the refund policy, processes the refund in the store's backend, sends the confirmation, and closes the ticket [2]. That last step is the whole distinction: the chatbot's role ends at information; the agent's role ends at the outcome. A chatbot deflects. An agent resolves [2].

The industry's own definitions now hinge on that acting part. AWS defines agentic AI as an autonomous system that can act independently to achieve predetermined goals, and G2's AI Customer Support Agents category requires products to execute tasks on the customer's behalf — refunds, subscription renewals, appointment scheduling — via function calling [2]. This is why the "chatbot" label has become too narrow for these systems; calling one a chatbot is like calling a smartphone a cellphone — technically true, but missing the point of what it now does [4]. The difference is architectural, not cosmetic: an agent maintains conversation context across multiple steps, connects to CRMs and payment platforms, and takes direct action, where a chatbot provides static, FAQ-style responses with limited functionality [8].

The metric that exposes the gap: resolution, not deflection

For years, support teams measured chatbots by deflection — how many contacts they avoided. That metric is now understood to be misleading, because deflection measures what the AI did, not what the customer achieved [7]. A chatbot that deflects 60% of conversations but leaves customers unsatisfied has not reduced your workload; it has hidden it [3]. The consumer data is damning: in UJET's research, 80% of consumers said chatbots increased their frustration, and 78% still had to connect with a human afterward [2]. Deflection, in other words, too often becomes a euphemism for abandonment [1].

The metric that matters is resolution — how many issues were actually finished — and it should always be read alongside customer satisfaction, never in isolation [3]. Here the architecture sets a hard ceiling before anyone touches configuration. Basic chatbots max out around 20–40% automated resolution by handling FAQs; standard AI assistants with embedded business logic reach 40–60%; and true agentic platforms routinely hit 70–85% precisely because they connect directly to backend systems and execute real actions [7]. First-contact resolution tells the same story: well-implemented AI achieves 70–90% on autonomous interactions, while scores below 60% indicate a system that responds without actually resolving anything [7]. The shift from deflection to resolution is the single most important reframe in support AI — and it is exactly the shift a scripted chatbot cannot make, because it can answer but not act.

When a chatbot is genuinely enough

The honest answer is that not every support problem needs a custom agent. Chatbots are a reliable, cost-effective solution for high-volume, low-complexity queries where the answer is always the same: FAQs, store hours, basic policy information, password resets, and initial contact triage all fit this profile [6]. A chatbot also works well as the first filter in a support flow — collecting information, confirming identity, and routing to the right team — as long as it is not expected to go further than that [6]. Roughly 80% of routine, high-frequency interactions (order-status checks, basic FAQs) are effectively solved problems for even simple automation [13].

The failure point is specific and avoidable: it arrives when organizations deploy chatbots against complex workflows and then measure the resulting deflection numbers as though deflection and resolution were equivalent [6]. A password reset handled by a chatbot is invisible to your brand; a billing dispute where a customer spends 45 minutes trapped in a bot loop is the interaction everyone remembers [13]. Matching the tool to the ticket is the whole game.

When a custom agent wins

A custom AI agent earns its cost when three conditions hold. First, when customers need things done, not just answered — when resolution requires taking an action in a system rather than surfacing an article [3]. Second, when the work is complex, multi-step, or ambiguous — the action-shaped tickets that scripted flows cannot navigate [2]. Third, and this is where "custom" specifically beats off-the-shelf, when resolution depends on your systems, your policies, and your brand: authenticated context about the customer's account and history, actions executed through your internal APIs, answers grounded in your documentation, and guardrails tuned to your compliance reality.

That last point deserves emphasis because it is where the "custom AI chatbot" question is really decided. A generic, bolted-on chatbot has no authenticated context and no ability to act; a custom agent is embedded in the product flow with the customer's purchases, payment status, and history available before the first message [11]. The gap between what a deeply integrated custom system ships and what an off-the-shelf vendor ships is real [11]. The trade-off is speed versus depth: off-the-shelf platforms deploy in days and cover common patterns well, while a custom agent takes longer but reaches resolution depth, brand fidelity, and compliance control a template cannot.

The proof — and the caution: Klarna

No deployment illustrates both sides of this better than Klarna's, the most-cited AI customer-support case study in the world. In February 2024, Klarna launched an OpenAI-powered assistant custom-integrated into its app and web flow. Within its first month it handled 2.3 million conversations — two-thirds of all customer chats — doing the equivalent work of 700 full-time agents, resolving issues in under 2 minutes versus 11 minutes previously, driving a 25% drop in repeat inquiries, operating across 23 markets and 35-plus languages, and on track to add roughly $40 million in profit [9]. By its Q3 2025 update, Klarna reported the assistant doing the work of 853 agents and about $60 million in annual savings, with response times 82% faster than before [10]. This was not a chatbot answering FAQs; it read account data, processed refunds, and rescheduled payments through internal APIs without a human in the loop for the majority of cases [10].

That is the existence proof that a custom agent beats a chatbot at scale. But Klarna's second chapter is the one worth studying. In May 2025, CEO Sebastian Siemiatkowski publicly conceded the company had "gone too far" — "we focused too much on cost; the result was lower quality" — and began rehiring human agents, committing that a customer can "always [have] a human if you want" [10][12]. The reasons are precise and generalizable. Hallucinations on edge cases produced confident-but-wrong answers about fees and payment terms, and in fintech a wrong answer about money is a compliance problem, not merely a CSAT one [10]. Customer satisfaction dropped on complex, emotional, dispute, fraud, and hardship tickets even when the AI was correct [10]. And the "replaced 700 agents" framing was itself misleading — Klarna had avoided hiring during a growth phase, not laid off 700 people [10].

The lesson is not that AI failed; it is the boundary. AI handles the volume tier; humans handle the value tier; and the line between them is a design decision, not a static rule [11]. The 80% of interactions AI handles well are invisible to your brand, while the 20% it handles badly are the only ones anyone remembers [13]. Klarna did not fail by using AI — it over-rotated by using it instead of humans rather than alongside them, and the eventual model is an explicit hybrid: AI on the bulk of routine volume, humans on the complex, emotional, and high-risk cases, equipped with AI assist [12]. Klarna's AI remained the front line for most chats throughout [11]. Salesforce's own Agentforce deployment reinforces the upside of getting this right: on Salesforce's support site it reached 45,000 conversations a week at an 85% resolution rate, with AI-resolved CSAT of 4.6 out of 5, edging the 4.4 average for human-only interactions [5].

The architecture of a custom support agent

Klarna's success and its correction both come down to architecture, which for a production support agent has three layers plus a protocol [15]. The knowledge layer grounds every answer in clean, current documentation via retrieval-augmented generation — Klarna invested one to two weeks cleaning its help center before launch, because without consistent policy docs the AI would hallucinate or give conflicting answers [10][15]. The action layer is the middleware that lets the agent safely trigger API calls to read customer data and execute transactions [15]. The orchestration layer is the router that decides, for each conversation, whether the AI can handle it or whether it requires human judgment, and routes with full context preserved [15][12].

Guardrails wrap all three. The patterns that made Klarna's system safe are the ones any custom deployment should adopt: grounding answers against a whitelist of approved topics and handing to a human when the model is unsure rather than guessing; pre-approved responses for sensitive matters like charges, fees, refunds, and disputes; automatic escalation when legal-complaint language appears; and complete conversation logging with timestamps for audit [14][12]. Modern platforms formalize this into dozens of guardrails — PII detection, sensitive-topic redirection, hallucination prevention — paired with human-in-the-loop QA and weekly hallucination monitoring on sampled conversations [16][10]. Guardrails are not an afterthought bolted onto a chatbot; they are what make an autonomous agent safe to point at real customers and real money.

How to decide

For a consideration-stage decision, the sequence is concrete. Start by analyzing your recent tickets and categorizing them by intent [15]. Then ask the three diagnostic questions. Do your customers need things done or just answered? If actions are required, prioritize a custom agent over a traditional chatbot [3]. Which tier are you automating — the high-volume, low-complexity requests where a chatbot suffices, or the complex, action-shaped, high-value cases where only an integrated agent resolves [6][2]? And can you measure the outcome honestly — resolution rate read together with CSAT, not deflection in isolation [3][7]?

From there, three non-negotiables separate a system that improves CX from one that quietly compounds risk: backend integration with write access (an agent that can only read can explain a fix but not perform it) [3]; guardrails, controlled escalation, and audit logging designed in from the start, not retrofitted [14]; and a hybrid model by design, with the AI on the volume tier and humans on the value tier [12]. The build-versus-buy nuance follows the same logic as the rest of enterprise AI: buy or configure an off-the-shelf agent for common, well-bounded patterns; invest in a custom, deeply integrated agent when resolution depth, brand voice, and compliance control are the competitive difference — and Klarna's arc is the reminder that even a best-in-class custom agent needs its scope drawn deliberately.

Where Etheon stands

The pattern across every data point is consistent: the win does not come from a smarter chatbot but from an orchestrated system that resolves rather than deflects — grounded in your knowledge, able to act through your systems, wrapped in guardrails, and routed to a human wherever judgment, emotion, or risk demands it. A custom AI customer support agent is not a scripted bot with better copy; it is a governed, observable, action-taking system with humans designed in alongside it, not replaced by it. That is the premise Etheon builds on: support automation as an orchestrated, auditable multi-agent system on infrastructure the enterprise controls, where the boundary between AI and human is an explicit design decision. The chatbot ends the conversation with a link. The agent ends it with the problem solved — and knows when to hand you to a person.

FAQ

What is the difference between an AI customer support agent and a chatbot?
A chatbot answers questions with scripted or FAQ-style replies and deflects volume; an AI customer support agent resolves issues by taking action — reading account data and executing tasks like refunds, renewals, or scheduling through connected systems via function calling. A chatbot's role ends at information; an agent's ends at the outcome [2][8].

When does a custom AI agent beat a chatbot for customer support?
When customers need things done rather than just answered, when the work is complex or multi-step, and when resolution depends on your systems, policies, and brand — authenticated account context, actions through your APIs, answers grounded in your docs, and compliance-tuned guardrails. For high-volume, always-same-answer queries like FAQs and password resets, a chatbot is enough [3][6][11].

What resolution rate can an AI support agent achieve?
Architecture sets the ceiling: basic chatbots resolve about 20–40% (FAQs), standard AI assistants 40–60%, and true agentic platforms with backend access 70–85%, with first-contact resolution of 70–90% for well-implemented systems. Klarna's custom agent handled two-thirds of chats; Salesforce's Agentforce reached 85% on its own support site [7][9][5].

Is it safe to let an AI agent handle customer support autonomously?
Yes, within limits and with guardrails. The reliable pattern grounds answers in approved sources, uses pre-approved responses for sensitive topics, escalates to humans on complex or high-risk cases, and logs everything for audit. Klarna's 2025 course-correction showed the boundary: AI on the volume tier, humans on disputes, fraud, and hardship — alongside AI, not replaced by it [14][12][10].

What did the Klarna deployment teach about support AI?
That a custom agent can resolve at massive scale — 2.3 million chats, two-thirds of volume, resolution time cut from 11 minutes to under 2 — but that the 20% of complex, emotional, and compliance-sensitive cases determine brand trust. The takeaway is a hybrid by design: automate the volume tier, keep humans on the value tier, and draw the boundary deliberately [9][10][11].

References

1. Assembled — Best AI Chat Agents for Customer Support (2026) (resolution vs deflection; "chatbot" a loaded term; hidden CX debt — loops, blocked escalation, lost context). https://www.assembled.com/blog/ai-chat-agents-customer-support

2. Engaige — AI Chatbot vs AI Agent: What Is the Difference? (2026) (AWS/G2 definitions; refund example; UJET 80% frustration / 78% still reach a human; Salesforce 30%→50%; deflection → resolution). https://letsengaige.com/blog/ai-chatbot-vs-ai-agent/

3. Fin.ai (Intercom) — 11 Best AI Chatbots for Customer Support in 2026 (agents resolve and act; 60% deflection but unsatisfied = hidden workload; "answers vs actions" first question; resolution rate with CSAT). https://fin.ai/learn/best-ai-chatbots-customer-support

4. Chatbase — AI Chatbot vs AI Agent: Customer Support in 2026 (Zendesk Resolution Bot; Einstein GPT actions; Intercom "Fin AI Agent"; 85% of enterprises using agents in 2026). https://www.chatbase.co/blog/ai-chatbot-vs-ai-agent

5. CoSupport — We Tested Autonomous AI Agents in Customer Support for 90 Days (Salesforce Agentforce own support: 45,000/week, 85% resolution, CSAT 4.6 vs 4.4; coaching over rigid rules). https://cosupport.ai/articles/autonomous-ai-agents-customer-support-test-results

6. Notch — AI Chatbot vs. AI Agent: Customer Support Guide (chatbots for high-volume/low-complexity FAQs, password resets, triage; failure = complex workflows measured by deflection). https://www.notch.cx/post/ai-chatbot-vs-ai-agent-customer-support-guide

7. Notch — AI Customer Service Metrics That Matter in 2026 (deflection measures what AI did, not customer outcome; resolution by architecture 20–40% / 40–60% / 70–85%; FCR 70–90%). https://www.notch.cx/post/customer-service-ai-metrics

8. ChatSpark — How AI Customer Service Agents Are Replacing Traditional Chatbots (agents: multi-step, context, autonomous; CRM/payment actions; up to 85% resolution; API costs down ~90% since 2023). https://chatspark.io/blog/how-ai-customer-service-agents-replacing-traditional-chatbots

9. Klarna / OpenAI — Klarna AI Assistant Handles Two-Thirds of Customer Service Chats in Its First Month (2.3M chats; 700-agent equivalent; <2 min vs 11 min; 25% fewer repeats; on-par CSAT; 23 markets, 35+ languages; ~$40M profit). https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/

10. Twig — What Klarna's AI Did in 30 Days — And What Broke (architecture: OpenAI + custom prompts + internal APIs + RAG grounding; beta human-review; Q3 2025 853 agents / ~$60M / 82% faster; walk-back reasons). https://www.twig.so/blog/how-klarna-is-revolutionizing-customer-support-with-ai

11. Perspective AI — Klarna AI Customer Service: Replacing 700 Agents — A 2026 Case Study (off-the-shelf gap is real; authenticated context before first message; embedded not bolted-on; volume tier vs value tier; scope correction). https://getperspective.ai/blog/klarna-ai-customer-service-replacing-700-agents-conversational-ai-case-study

12. Asisteclick — Klarna Laid Off 700 Humans for AI. Now It's Rehiring ("we went too far"; hybrid/Uber-style; guardrails — pre-approved responses, escalation on legal keywords, audit logging; "instead of" vs "alongside"). https://asisteclick.com/en/blog/klarna-error-ia-customer-service-leccion/

13. Medium (Veriprajna) — Klarna Replaced 700 People with AI (the 80/20 lesson — the 20% is brand and liability; the invisible-vs-remembered interaction). https://medium.com/@ashutosh_veriprajna/klarna-replaced-700-people-with-ai-0ff99fe4ada7

14. ChatAid — How Klarna's AI Assistant Replaced 700 Customer Service Agents (guardrails: whitelist of approved topics; hands to human when unsure; L1 tickets; fintech wrong answer = money). https://blog.chataid.com/blog/klarna_ai_customer_service/

15. ModelGate — Klarna's AI Agent Results (three-layer architecture: knowledge/RAG, action/middleware, orchestration/router; outdated wikis → hallucination; humans to Tier 3). https://modelgate.ai/blogs/ai-automation-insights/klarna-ai-agent-results-bpo-extinction-2025

16. Fini — Klarna Automates Two-Thirds of Customer Service with AI Assistant (40+ guardrails: PII detection, sensitive-topic redirection, hallucination prevention; human-in-the-loop QA; hybrid model). https://www.usefini.com/blog/klarna-automates-two-thirds-of-customer-service-with-ai-assistant