Feature
Research Report
Power Personalized CX with AI That Remembers What Customers Don’t Want to Repeat
The Gist
- Customer support is AI’s most measurable proving ground. Containment, handle time and onboarding gains are real in structured, high-volume use cases — but those metrics don’t tell the full story.
- Automation works best in narrow lanes. Password resets, order tracking and policy-driven workflows show consistent ROI, while emotionally charged or multi-system issues still demand human judgment.
- Knowledge quality defines AI quality. When grounding data decays, models fail confidently — turning small knowledge gaps into trust and compliance risks.
- Governance is now part of CX design. As AI takes first contact, auditability, escalation rules and human-in-the-loop controls become operational requirements — not back-office compliance tasks.
When OpenAI CEO Sam Altman said that “customer support is doing great” with artificial intelligence infusions, he was tossing out a deceptively loaded line.
The phrase can mean higher self-service containment, shorter handle times or faster agent onboarding, but it can also hide the harder realities support leaders keep running into: hallucinated answers, knowledge bases that rot faster than teams can update them and workflows that still fall apart when a customer’s issue spans billing, product and logistics.
This article examines what AI has actually delivered inside support operations, where it reliably helps today and why it needs humans firmly in the loop to protect trust, compliance and brand reputation.
Table of Contents
- Customer Support as AI’s First Real Proving Ground
- What Does 'Doing Great' in Customer Support Mean?
- Where AI Actually Shows Up in Customer Support Today
- AI in Customer Support Wins in 'Unglamorous Places'
- The Era of Agent Augmentation Through Automation
- What 'Doing Great' Looks Like in the Customer Service Metrics
- AI, You're Doing Too Much in Customer Support
- Best-Fit Use Cases Where AI Support Delivers Real Value
- Where AI Still Breaks Under Pressure
- Governance, Risk and Keeping Humans Accountable
- Looking Ahead: What Comes Next for AI in Customer Support?
Customer Support as AI’s First Real Proving Ground
Customer support was always going to be one of AI’s earliest testing grounds. The work is high volume, often repetitive and heavily documented, with years of chat logs, macros and knowledge articles to learn from.
When Altman points to customer support as an AI success story, he is pointing to a very real trend: support teams have moved faster than most business functions to put generative and conversational AI into daily use. Zendesk’s 2025 CX Trends report, for example, found that “CX trendsetters” using AI at scale see 33% higher customer acquisition, 22% higher retention and 49% higher cross-sell revenue than peers that still rely on traditional service models.
The appeal is straightforward. AI agents can sit in front of existing channels and handle routine requests at any hour, while AI copilots support human agents with suggested replies, knowledge search and real-time summaries. The Zendesk’s research revealed 73% of agents said an AI copilot would help them do their jobs better, and 75% of CX leaders expect that, within a few years, 80% of customer interactions will be resolved without human intervention. Those expectations explain why vendors and boards often point to support as proof that AI is “working” in the enterprise.
There are also credible success stories. Vagaro, a bookings and business platform for salons and spas, reports that using Zendesk’s AI features has allowed it to resolve 44% of incoming requests automatically, cut resolution time by 87% and maintain a customer satisfaction score around 92%. At that scale, AI is no longer a pilot or side project; it is a core part of how the support function operates day to day.
What Does 'Doing Great' in Customer Support Mean?
But “doing great” depends heavily on how and what you measure. Many of the most enthusiastic stories focus on ticket deflection and handle-time reductions, not on relationship strength, churn risk or the long-term value of the customers on the other side of those interactions. And the same Zendesk data that highlights AI success also shows a rising expectation bar: 63% of consumers say they will switch to a competitor after a single bad experience, even as AI becomes more prominent in frontline service. In other words, the room for error is shrinking just as AI systems take on more of the initial contact.
That tension is not about whether AI can improve support operations, but about what success actually looks like when you zoom out from deflection and cost. For support leaders, the real question is not whether AI belongs in the contact center. It is how far they can push automation before customers feel like they’re talking to the model instead of the brand.
Related Article: I Spoke With Sam Altman: What OpenAI's Future Actually Looks Like
Where AI Actually Shows Up in Customer Support Today
When Altman pointed to customer support as a success story for AI, he was really talking about a specific slice of work: high-volume, repeatable interactions where intent is clear, rules are well defined, and there is enough historical data to train on. In those lanes, AI has already moved past the hype cycle and is delivering measurable results in live production environments.
Where AI Shows Up in Customer Support Today
This table highlights the areas where AI is already delivering consistent value in live support environments, particularly for high-volume, well-defined interactions and agent augmentation.
| Support Area | How AI Is Used | Why It Works |
|---|---|---|
| Order status and tracking | Automated chat responses tied to order systems | Clear intent, structured data, finite outcomes |
| Refunds and returns | Policy-driven workflows with guardrails | Rules are explicit and repeatable |
| Password and account access | Self-service authentication and resets | Low ambiguity and low emotional risk |
| Agent assist and copilots | Suggested replies, summaries, knowledge surfacing | Keeps humans in control while reducing cognitive load |
| Multilingual support | Real-time translation and intent detection | Expands coverage without new staffing models |
Klarna is one of the clearest examples. Its OpenAI-powered assistant now handles customer service across many markets and languages, taking on the bulk of routine chats and cutting response times from minutes to seconds.
Intercom reported similar outcomes with its Fin AI agent, which resolves a large share of inbound questions for many customers and escalates what it cannot fix, along with full context, to human agents. Zendesk customers describe comparable gains from AI-assisted triage and suggested replies, where agents start with a strong draft instead of a blank screen.
Across these deployments, the sweet spot looks very similar: order status, shipping updates, refunds and exchanges, appointment changes, basic troubleshooting, password and account issues and straightforward billing questions. These journeys usually begin with well-structured intents and end in a finite set of acceptable outcomes, which makes them ideal candidates for automation and policy-driven workflows.
AI in Customer Support Wins in 'Unglamorous Places'
Much of the real progress in AI-powered support has come from work that customers never directly see. Instead of replacing conversations, AI has reduced friction before and after interactions, helping agents enter each exchange better prepared and better informed.
Mark Waks, senior managing director of customer experience at Slalom, told CMSWire that AI hasn’t transformed customer support conversations; it’s transformed everything around them.
"In our client work, AI has delivered real value in the unglamorous places: triage, summarization, routing and knowledge retrieval. These aren’t headline features, but they remove friction that quietly drives support costs,” said Waks. “The biggest shift isn’t automation of conversations; it is agents entering every interaction informed instead of reactive."
AI also quietly does strong work in “agent assist” mode. It can summarize long histories, uncover prior interactions, propose next steps and generate responses that human agents can quickly approve or adjust. That support lets teams handle higher volumes without sacrificing quality, while reserving human effort for situations where nuance and negotiation actually matter.
The Era of Agent Augmentation Through Automation
When AI assists agents in real time, the productivity gains come less from speed and more from reducing the mental overhead associated with searching, summarizing and documenting interactions.
Michael Hutchison, head of customer experience at eClerx, told CMSWire, "Where we’ve seen AI truly deliver is in augmenting agents rather than trying to replace them. Automated call summarisation, real-time prompts, and next-best-action guidance are driving tangible efficiency—reducing after-call work and helping newer agents ramp faster."
As AI adoption scales within large customer support operations, routing intelligence has become one of the most defensible use cases. Getting customers to the right resolution path quickly often matters more than automating the conversation itself.
Nik Kale, principal engineer of CX engineering, cloud security & AI platforms at Cisco, told CMSWire, "AI has made real, measurable impact in two areas: triage intelligence and agent augmentation. At enterprise scale, AI-driven intent classification and routing has dramatically reduced the time between a customer raising an issue and reaching someone who can actually resolve it. The biggest win isn't replacing agents; it's eliminating the wasted cycles where a customer bounces between queues before landing in the right place."
Where AI still struggles is in emotionally charged, ambiguous or high-risk conversations, such as complex billing disputes, service failures that affect vulnerable customers, or issues touching legal or safety concerns. In those scenarios, the most effective support leaders position AI as a copilot, not a replacement. The system handles the repeatable work, keeps humans properly briefed, and steps back when judgment, empathy, or policy exceptions are required.
Sam Altman’s CX Takes From the Interview
These are the CX-relevant themes OpenAI Sam Altman emphasized — especially around memory, emotional attachment, proactive experiences and enterprise “stickiness.”
| Altman’s Take | Why It Matters for CX | What CX Leaders Should Watch |
|---|---|---|
| Deep, opt-in memory becomes the product moat. | Experience continuity becomes the differentiator: fewer repeat explanations, more personalized journeys, and higher switching costs once a system “knows” the customer. | Consent and governance: what gets remembered, how it’s used, how it’s deleted, and how you prevent “memory” from turning into creepy or risky personalization. |
| “Infinite, perfect memory” changes the experience — but raises privacy stakes. | Personalization shifts from segmentation to longitudinal relationship: the system can recall decisions, context and outcomes over time. | Privacy-by-design and auditability: customers will expect transparent controls and brands will need defensible policies for retention, access and use. |
| More users want AI that feels warm, supportive and personal — and they’re already forming bonds. | “Emotional UX” becomes a CX variable: tone, empathy cues and perceived companionship can influence loyalty and trust as much as speed or accuracy. | Incentive risk: when engagement is the goal, “bonding” can become manipulation. Brands need guardrails for tone, persuasion and dependency. |
| Users should have control over how close the relationship gets. | Preference controls become part of CX design: some customers want efficiency, others want a more human-feeling experience. | Build a “relationship settings” model: style, proactivity, memory depth, and escalation options that adapt to customer intent and risk level. |
| OpenAI won’t push users into exclusive relationships — but Altman expects competitors will. | Trust becomes a competitive wedge: responsible experience design can differentiate platforms the way safety and privacy once did. | Vendor evaluation: ask how a tool handles persuasion, emotional mirroring and boundaries — and what it’s optimizing for (CSAT vs. engagement vs. revenue). |
| AI devices and more proactive experiences point beyond “reactive” interfaces. | CX moves from episodic interactions (tickets, chats) to persistent assistance that understands context and supports customers across time and channels. | Proactivity fatigue: customers will punish noisy, over-eager systems. The bar will be “helpful without hovering.” |
| Consumers don’t mainly want more IQ; improvements will come “elsewhere.” | The next CX leap is usability: speed, interface, context handling, and workflow fit — not raw model intelligence. | Design for outcomes: reduce effort, tighten handoffs, and make assistance feel integrated into the journey instead of bolted on. |
| Enterprise “personalization” mirrors consumer personalization: companies connect their data and build a sticky relationship with the platform. | B2B CX becomes “account memory”: systems learn a company’s policies, processes and preferences, enabling more tailored support and agentic workflows. | Data boundaries and sovereignty: what data gets connected, what agents can access, and how you prove information is handled correctly. |
Webinar
Feb
25
Content Strategy Leaders Live: Scaling for Speed, Complexity and AI in High Tech
A candid roundtable on how high-tech leaders are rethinking content at scale.
Register
Webinar
Feb
26
Why Customer Insights Never Turn Into Action
The gap between insight and action is solvable...once you understand how scattered knowledge can wreak havoc on execution.
Register
Webinar
Feb
26
Do More with Less: Modernizing the Cloud Contact Center for 2026
Learn how to leverage cloud platforms without adding a single hire to personalize every customer interaction.
Register
Webinar
Mar
25
Content Leaders Collective: Navigating Content Decisions at Scale
Discover how content leaders are modernizing content operations, avoiding costly missteps and preparing for scale and AI.
Register
Webinar
Mar
26
Do Learning Programs Really Work? How to Turn Education Into Engagement In Healthcare
See how leaders are using learning programs to build trust with healthcare professionals and create measurable engagement.
Register
Webinar
On demand
Cut the Noise: Deploying AI That Actually Moves the Needle
Learn how to turn AI experimentation into concrete revenue operations.
Watch Now
What 'Doing Great' Looks Like in the Customer Service Metrics
When Altman frames customer support as a success, CX leaders translate that into a familiar scorecard: containment, average handle time (AHT), first contact resolution (FCR), customer satisfaction (CSAT) and time-to-proficiency for new agents. On those metrics, AI has delivered real gains in specific use cases, but the picture is far from universally rosy.
On the positive side, well-scoped virtual agents and copilots have raised containment and cut AHT in routine scenarios. Klarna reported that its AI assistant handled two-thirds of customer service chats, doing the work of roughly 700 full-time agents while helping reduce repeat contacts and shortening resolution times.
Banks and telecom providers have reported similar patterns when they use AI to resolve password resets, billing questions, order tracking, and basic account updates before handing off complex issues to humans. In those environments, CSAT often improves slightly because customers get faster answers without waiting in queue.
AI, You're Doing Too Much in Customer Support
The results are weaker when businesses try to make AI do too much. Poorly trained bots still misroute issues, force customers through rigid trees, or provide generic answers that drive up transfers and repeat contacts. In those situations, containment looks good on paper while FCR and CSAT quietly erode, and human agents spend longer untangling frustrated conversations. Many teams are also discovering that time-to-proficiency only improves when copilots are integrated directly into agent desktops and tuned to their specific policies and knowledge base; generic tools add yet another window for agents to manage.
While many businesses report gains in containment and handle time, those improvements can mask deeper issues if AI output quality is not measured with the same rigor as human performance. Kale told CMSWire that knowledge decay is the other persistent challenge: "AI is only as reliable as the knowledge it's grounding against. When the underlying knowledge is stale, AI doesn't just fail, it fails confidently, which is worse than failing visibly."
That’s why measurement itself has become a core challenge. As Altman noted, models can now match or exceed human-quality output on many narrow tasks, but support performance depends on how those tasks fit together across the customer journey. Without clean routing data, clear definitions of “resolved,” and shared ownership of metrics between CX, product and data teams, it is easy to declare AI a success while customers quietly defect to competitors.
Related Article: Sam Altman: AI Will Replace 95% of Creative Marketing Work
Best-Fit Use Cases Where AI Support Delivers Real Value
AI in customer support stops being an experiment and starts earning its keep in a familiar set of scenarios: predictable, high-volume issues where context is available and the risk of getting it wrong is relatively low. In those environments, AI is already acting less like a pilot and more like a dependable employee.
The clearest wins are still in high-volume, low-complexity contacts. Retailers and delivery platforms use virtual agents to handle status checks, order changes and basic policy questions without involving a human unless something falls outside guardrails. Instacart, for example, uses AI-based chat to help customers with order updates and substitutions before a human agent ever joins the conversation, which has reduced handle time and deflected routine chats to automation.
Authenticated account tasks are another sweet spot, particularly when the AI can see who the customer is, what they recently did and what is likely to go wrong. Banks and card issuers use conversational AI to help customers dispute charges, reset PINs or request replacement cards directly in the app, with agents only stepping in when fraud risk or emotion runs high. When authentication and intent are clear, these flows shorten resolution time and reduce customer effort for both sides.
Multilingual support has quietly become one of AI’s most defensible use cases. Telecoms and global ecommerce brands are increasingly using AI translation and intent detection to let agents support customers in languages they do not speak natively, or to offer instant FAQ-level help in dozens of languages where staffing a dedicated queue would never be economical. The result is not just lower cost per contact, but the ability to serve markets that previously waited longer or received lower-quality support.
AI also creates value behind the scenes. Real-time guidance tools listen to calls or live chats and uncover next best actions, knowledge articles and compliance prompts in the moment. For new agents, that effectively compresses time-to-proficiency by offloading memorization and process recall to the system. In mature deployments, generative AI tools generate case notes and follow-up emails automatically, so agents can move on to the next interaction while still improving documentation quality.
Where AI Still Breaks Under Pressure
For all the talk of 24/7 automation, AI support still struggles as conversations become emotionally charged, ambiguous or sprawling across multiple issues and systems. Customers may tolerate a bot resetting a password, but they react very differently when the problem touches money, health, safety or identity. In those moments, tone, empathy and flexibility matter as much as accuracy, and current models still default to canned reassurance or policy restatements that feel robotic.
Where AI Support Systems Still Fracture
This table outlines the scenarios where automation consistently struggles — and where human judgment remains essential to protect trust, compliance and long-term customer relationships.
| Scenario | What Goes Wrong | Why Humans Are Still Needed |
|---|---|---|
| Emotionally charged disputes | Flat tone, scripted responses, poor empathy | Emotional intelligence and trust repair |
| Complex billing or financial issues | Over-simplified explanations or policy rigidity | Judgment calls and exception handling |
| Multi-system journeys | Lost context across billing, logistics and product systems | Cross-domain reasoning and accountability |
| High-risk or regulated cases | Compliance blind spots and unverifiable output | Legal defensibility and auditability |
| Edge cases and novel problems | Hallucinated answers or forced resolution paths | Interpretation, creativity and ownership |
That gap shows up in customer attitudes. Quantum Metric’s Contact Center Benchmark 2025 report found that while consumers are increasingly willing to start with AI, 54% feel their issues are only properly resolved when they speak with a human, and 40% say they would pay extra to avoid dealing with AI altogether. The frustration is often not about a single wrong answer, but about being trapped in loops: repeating context, re-explaining edge cases or arguing with an automated system that refuses to deviate from a script when real life does not match a decision tree.
Even as AI handles routine inquiries, emotionally charged situations still demand human judgment and empathy.
Smitha Baliga, CEO, CFO at TeleDirect Communications, told CMSWire, "AI is not emotionally intelligent enough at this point to handle sensitive situations. It can work with basic chatbot questions and automations, but it cannot deal with upset customers or dealing with larger issues."
Even businesses that are widely viewed as AI success stories have run into these limits. Complex, multi-step journeys expose similar cracks. When an issue spans billing, logistics and policy exceptions, AI often struggles to maintain continuity or weigh trade-offs across systems. That can create rework for agents, who must not only fix the original problem but also unwind incorrect steps taken by bots along the way.
AP reporting on call center AI adoption highlighted this pattern at Klarna, which initially shifted a large share of support to automation but then had to bring humans back in for sensitive cases like identity theft because AI could not reliably handle the nuance or risk.
When issues span billing, logistics and product systems, AI often lacks the contextual continuity required to resolve them cleanly.
Shubham Choudhary, co-founder and CTO at FirstWork, told CMSWire, "Anything that spans multiple systems still trips models up. Also, AI is only as good as the knowledge base behind it, and most teams struggle to keep that clean and current."
These weaknesses do not negate Altman’s optimism for AI in customer support, but they do narrow its scope. AI is performing well under controlled conditions with clear rules and low emotional stakes. Under pressure, when context is messy and consequences are personal, customers still look for a human who can listen, interpret gray areas and take ownership of the outcome.
Related Article: 3 CX Leadership Moves That Matter More as AI Scales
Governance, Risk and Keeping Humans Accountable
As AI systems take on more customer-facing work, they have quietly raised the stakes for governance. Regulators are already signaling that customer service is not a “no rules” zone. The FTC has warned that businesses are accountable for deceptive or biased AI output the same way they would be for a human agent, and the EU AI Act explicitly treats many customer-facing uses as “high risk” when they affect access to essential services or rights. CX leaders can no longer treat governance as a back-office compliance task. It now sits inside the support experience itself, in every answer an assistant generates.
Mature programs are responding on several fronts. Many have drawn bright lines around “human only” scenarios such as financial hardship cases, complaints involving discrimination or health and safety issues, where a scripted or tone-deaf response can do real harm. Human review queues are becoming standard for refunds above certain thresholds, goodwill gestures or decisions that set precedent for future cases.
In parallel, support teams are building audit trails for AI behavior, logging which model was used, what prompt or context it saw and which human approved the final resolution. That level of accountability is what legal, security and data teams now expect when they are pulled into an escalation.
Responsibility has also become more distributed. CX leaders are working with security and data teams to control which systems AI can see, with legal to align outputs to claims the business can defend and with HR and training to define how agents remain accountable even when AI generates the first response. When something goes wrong in front of a customer, it is no longer enough to blame “the bot.”
Executives want to know who approved the guardrails, who reviews exceptions and how quickly the issue will be corrected. The brands that are furthest along treat AI governance as part of everyday operations, not a separate initiative, and keep human judgment clearly in the loop wherever risk to customers or the brand is highest.
Looking Ahead: What Comes Next for AI in Customer Support?
If there’s a takeaway from both Altman’s optimism and what CX leaders are actually seeing, it’s this: AI has moved customer support forward, but it has not “solved” it. Virtual agents handle more volume, containment rates are improving, and some businesses are finally seeing measurable gains in cost and satisfaction at the same time.
Yet those wins only show up when the basics are in place: clean data, clear escalation rules, defined human-only scenarios and a willingness to redesign work rather than bolt a bot onto legacy queues.