AI in Customer Experience: 9 Real-World Playbooks for 2026

The frustrating hold music, the "please describe your issue again" loop, the three-day email reply - most people reading this still remember support that worked like that. By 2026, the floor has moved. Customers expect resolutions in seconds, in their language, on the channel they already use, and they expect the agent on the other end to know who they are. AI is the reason that floor moved.

This post walks through nine companies - large enterprises and lean teams - that are using AI agents to deliver the kind of customer experience their users now consider table stakes. Each one made a different design choice: some built in-house, some bought, some routed AI to specific lanes of work and kept humans on others. The patterns repeat enough that you can borrow them directly.

What AI is actually doing for customer experience right now

Before diving into examples, it helps to anchor in what changed. Three things make 2026's AI agents qualitatively different from the chatbots most companies bolted onto their sites a couple of years ago.

Frontier models that can actually reason. OpenAI's GPT-5.5 and GPT-5.5 Pro (with parallel reasoning) shipped in April 2026. Anthropic's Claude Opus 4.7 leads SWE-bench Pro at 64.3% and Claude Opus 4.6 / Sonnet 4.6 ship with a 1M-token context window at no surcharge. Google's Gemini 3.1 Ultra runs on a 2M-token context and is natively multimodal across text, image, audio, and video. For a support agent, that means the model can hold an entire conversation history, the customer's full account record, your policies, and your knowledge base in one prompt - and reason over them all without hallucinating its way out of trouble.

Open-weight models that collapsed inference cost. DeepSeek V4 Flash (April 2026) is priced at $0.14 per million input tokens and $0.28 per million output tokens. MiniMax M2 / M2.7 runs at roughly 8% the price of Claude Sonnet at twice the speed. Z.ai's GLM-5.1, Moonshot's Kimi K2.6, Alibaba's Qwen 3.6 family, and Xiaomi's MiMo-V2-Pro all ship as open weights with MIT or Apache licenses. The practical consequence: a Berrydesk deployment can route routine questions to DeepSeek V4 Flash or MiniMax M2 at fractions of a cent per resolution, and reserve Claude Opus 4.7, GPT-5.5, or Gemini 3.1 Ultra for the hard escalations.

Agentic tool-use that's stopped being demoware. Kimi K2.6 sustains 12-hour autonomous coding sessions with up to 300 sub-agents. GLM-5.1 runs an 8-hour plan-execute-test-fix loop. Claude Opus 4.7 and Qwen 3.6 are reliable enough at chained tool-use that "AI Actions" - booking a meeting, issuing a refund, looking up an order, charging a card, swapping a SIM - are now production-grade, not screen-recorded demos.

That's the substrate. Here's what nine companies are doing on top of it.

1. Unity - a focused chatbot that hands off cleanly

Unity, the game engine that powers a meaningful chunk of mobile and indie titles, runs a customer-facing assistant called UB-1 on its support portal. UB-1 is deliberately scoped: it answers the high-volume, low-complexity questions Unity's support team gets every day - license activation, Asset Store purchases, account access, billing changes. When a question goes beyond what UB-1 can handle, it converts the conversation into a ticket and hands it to a human on Unity's Customer Experience team with full context.

Why it works. Unity didn't try to make UB-1 omniscient. They identified the categories of question where a self-serve answer would actually be faster and more accurate than a human reply, scoped the bot to those, and made the handoff to humans frictionless. Customers get instant answers for the boring stuff; the human team focuses on the harder cases - engine bugs, custom-license negotiations, escalations.

The workflow customers see.

• A live chat widget surfaces UB-1, with quick-reply buttons for common issues like "license isn't activating."
• Free-form questions are parsed and answered against Unity's documentation.
• For Unity Gaming Services issues, UB-1 routes users to the dashboard sign-in.
• Sales inquiries go to a separate channel.
• Anything UB-1 can't close becomes a ticket - with the full chat history attached, so the human agent doesn't ask the customer to repeat themselves.

The pattern here - narrow scope, clean handoff, no pretense - is the one that scales. Trying to make a single agent answer everything is how you get the bad chatbots customers learned to hate.

2. Deutsche Telekom - a digital assistant that grew into a self-service portal

Deutsche Telekom's "Frag Magenta" ("Ask Magenta") is one of the more ambitious deployments in European telecom. It started as a deflection bot - handling FAQ traffic - and evolved into something closer to a self-service portal wearing a chat interface. Today it sits on telekom.de and handles questions about phone, internet, and TV issues; bills, contracts, and orders; moves and address changes; Wi-Fi, email, and SIM card configuration.

What makes Frag Magenta worth studying is the depth of integration. The assistant has secure access to customer account data, which means it can:

• Pull up mobile phone bills going back 18 months and let users download them on the spot.
• Tell a customer exactly when they're eligible to upgrade their contract.
• Update bank details on file for direct debit.
• Carry a longer conversation across multiple turns without losing context - a place where 1M-token context windows on Claude Sonnet 4.6 or DeepSeek V4 genuinely matter.

When Frag Magenta hits the limit of what it can do, it routes the customer to the right human expert with all the conversational context intact. The result is a noticeable reduction in call-center load and a self-service experience that actually feels like service rather than a wall.

This is the workflow Berrydesk users build with AI Actions: a chat agent that can authenticate a user, hit your billing or order system, and complete a transaction without ever touching a human queue.

3. Camping World - Arvee, where AI replaced the after-hours void

Camping World, the U.S. RV retailer, deployed an AI virtual assistant named Arvee (yes, R.V.) on its website and via SMS. Arvee handles common product and account questions, integrates with Camping World's Oracle and Salesforce systems to fetch customer-specific data, and proactively engages site visitors who look like they need help.

The numbers from Camping World's deployment are the kind that get an AI project funded:

• 40% increase in customer engagement after Arvee went live.
• Wait times dropped to 33 seconds on average.
• Human agents handled 33% more conversations, because Arvee absorbed the easy ones first.
• Significant lead capture after hours, where the human team simply wasn't available before.

The after-hours story is worth pulling out. Most support teams have a daily rhythm - heavy traffic during business hours, a thin night shift or nothing at all, and a backlog every Monday morning. AI doesn't sleep. A meaningful share of the leads Arvee captures arrive between midnight and 6 AM, when prospects are browsing on their phones in bed. Without an AI, those leads either bounce or sit in a contact form until the morning. With one, they get answered, scoped, and queued for the right human follow-up.

Arvee also handles a smart flow specific to phone callers: if a customer is on hold, the assistant offers to switch them to SMS so they can keep doing whatever they were doing. That's a small UX choice that translates directly into a higher contact-resolution rate.

4. DNB (De Nederlandsche Bank) - internal AI as a force multiplier

Not every AI customer experience improvement is customer-facing. DNB, the central bank of the Netherlands, built an internal tool called ChatDNB that serves the bank's own staff - particularly the supervisors who answer questions from commercial banks and other regulated institutions.

ChatDNB was trained on more than 10,000 pages from DNB's "Open Book Supervision" knowledge base - the laws, regulations, and supervisory guidance the bank's staff need to reference daily. Crucially, it's restricted to that corpus. It doesn't reach out to the open internet. When it answers a question, it cites the specific document and section it pulled the answer from, the way a research paper cites a source.

Why this is a customer experience play, not just an internal tool. The customer here is the regulated institution calling DNB to ask whether a particular reporting practice is compliant. Before ChatDNB, the answer depended on which DNB analyst picked up the phone and how recently they'd reviewed the relevant rule. After ChatDNB, every analyst is answering from the same authoritative corpus, with citations they can quote back. The external customer experience improves because the internal one became consistent.

This is also a useful case study for regulated industries deciding between closed and open-weight models. If you want a ChatDNB-style internal agent on-prem, in an air-gapped environment, with auditable training data and weights you control, the open frontier matters. GLM-5.1 (MIT license, 754B-param MoE), Qwen 3.6-27B (Apache 2.0, dense), and MiMo-V2 (MIT, >1T params) all make that deployment realistic in a way it wasn't even a year ago.

5. Windtre - automating ten thousand tickets a month

Windtre, one of Italy's major telecom operators, partnered with IBM to build an "Intelligent Automation" pipeline that handles customer complaints end-to-end. A complaint comes in; the system classifies it, gathers the relevant data from internal systems, and - for a meaningful share of cases - solves the problem completely autonomously. For the rest, it pre-processes the ticket so a human agent can resolve it in a fraction of the time.

The architecture has two components:

• A Dispatcher that classifies and queues incoming tickets.
• A Performer that does the work, calling out to RPA bots and back-office systems to actually fix things.

Tickets fall into one of three buckets after the dispatcher takes a look: fully automated resolution, partially automated (AI gathers context and hands a packaged ticket to a human), or routed to the right team based on classification.

The reported outcomes:

• Over 10,000 complaints per month handled automatically.
• Response times 10x faster than the previous human-only workflow.
• Lower error rates, because the system pulls data from sources of truth instead of relying on agents copy-pasting between systems.
• Linear scalability, because adding capacity is a matter of provisioning more cloud compute, not hiring more agents.

The Windtre case is worth contrasting with Unity's. Unity scoped narrowly. Windtre went deep on a high-volume use case - telecom ticket processing - and built end-to-end automation for it. Both work. The wrong move is in between: a vague "AI-powered support" deployment that doesn't commit to either narrow scope or deep automation.

6. eye-oo - AI as a sales channel, not just a deflection bot

Eye-oo, an Italian online retailer selling designer eyewear, deployed an AI chatbot as the first line of support on its website. The bot is trained on the full product catalog, eye-oo's policies on shipping, returns, and prescriptions, and the order management system.

What makes eye-oo notable is that they treated AI as a sales surface, not just a cost-deflection lever. The bot recommends products, answers prescription questions, walks customers through the lens-fitting process, helps them check order status, and engages cart abandoners with targeted nudges to complete a purchase.

The reported impact:

• +25% sales after rollout.
• 5x conversion uplift on visitors who engaged the chat.
• €177,000 in attributable revenue generated through the AI's recommendations and recovery flows.
• 86% reduction in wait time - from 5 minutes down to 30 seconds.
• Over a thousand leads captured that would otherwise have bounced.

The general lesson: support AI and sales AI are increasingly the same agent. A customer asking "do these glasses come in tortoiseshell?" is mid-purchase. The fastest path to revenue is letting the agent answer, surface the SKU, and complete the transaction inline - which is exactly what AI Actions are for.

7. Zalando - personalization at the scale of 25 markets

Zalando, Europe's largest online fashion retailer, runs two AI surfaces that together define what "personalized at scale" means in 2026.

Zalando Assistant is a conversational shopping advisor inside the Zalando app and website. You can ask it something like "what should I wear to an outdoor wedding in Lisbon in October?" and it understands the occasion, the climate, the formality, and your size and style preferences. It then recommends specific products from the catalog. It runs in all 25 of Zalando's markets in the local language.

Trend Spotter is the inverse: instead of you asking the AI what to wear, it tells you what's trending right now in your city - and across 10 European hubs from Amsterdam to Warsaw - with explanations of why something is hot ("global tailoring revival" vs. "Berlin streetwear micro-trend").

The reported impact: +23% on product clicks, +40% on wishlist growth.

The Zalando setup is a good example of where the multimodal frontier matters. Gemini 3.1 Ultra's native vision and Claude Opus 4.7's strong visual reasoning let an assistant look at a product image, a runway photo, or a customer-uploaded outfit and reason about it the way a human stylist would. For any retail or media company, that capability changes what an AI agent can be - from a text-based FAQ machine into something closer to a knowledgeable shopping companion.

8. A SaaS support agent built on Berrydesk

The platform pattern - companies building their own branded AI agent on top of an off-the-shelf platform - is the largest cohort of AI customer experience deployments by count. It's also the easiest to underestimate. The eye-oos and Camping Worlds of the world have engineering budget to integrate IBM watsonx or build custom RAG. The long tail of SMB and mid-market companies need a faster path.

This is where Berrydesk fits. A typical Berrydesk deployment looks like:

• Pick a model - or several. Route routine ticket triage to DeepSeek V4 Flash or MiniMax M2 at fractions of a cent per resolution. Send anything that needs careful judgment (refund disputes, account escalations, complex multi-policy questions) to Claude Opus 4.7 or GPT-5.5. Use Gemini 3.1 Ultra when the customer attached a photo of a damaged product or a screenshot of an error.
• Train on your data. Point Berrydesk at your help docs, your website, your Notion workspace, your Google Drive folders, and your YouTube tutorials. The agent ingests everything and stays in sync as you update it.
• Brand the widget. The chat surface uses your colors, your typography, your name. Customers don't see "powered by Berrydesk" - they see your support team.
• Add AI Actions. Hook the agent into your billing system, your booking calendar, your order database, your refund flow. The agent can now actually do things, not just point customers at the right form.
• Deploy everywhere. Web widget, Slack, Discord, WhatsApp, SMS, email. One agent, one knowledge base, every channel your customers use.

The pattern that emerges from real Berrydesk deployments: companies typically see resolution rates north of 70% on Tier 1 traffic within the first month, with first-response times measured in seconds rather than hours. The harder cases still go to humans - but they arrive pre-qualified, pre-contextualized, and with the easy parts already handled.

9. The lean operator - one agent, many channels

The last example is less a single company and more a class of deployment we keep seeing: small-team operators - indie SaaS founders, course creators, agencies, niche e-commerce stores - who use a single AI agent as a force multiplier across their entire customer-facing surface.

A representative version of this looks like:

• A founder runs a paid community plus a course platform, with maybe two human team members.
• They train one AI agent on the course content, the community FAQ, the product roadmap, and their support email history.
• The agent lives on the marketing site (capturing leads and answering pre-sales questions), inside the community (answering "how do I…" questions about the platform), in a Slack workspace (handling internal team queries about product status), and on WhatsApp (where their customers actually message them).
• AI Actions let it onboard new members, issue refunds inside the company's policy limits, schedule calls with the founder, and update billing details - without a human in the loop for any of it.

For this kind of operator, AI isn't a 33% efficiency improvement. It's the difference between being able to run the business at all and drowning in DMs. The combination of cheap open-weight inference (DeepSeek V4 Flash, MiniMax M2) for the high-volume work and a frontier model (Claude Opus 4.7, GPT-5.5) for the few escalations a day makes the unit economics work even at small scale.

What to actually take away

A few patterns repeat across these nine deployments, and they're worth naming explicitly.

Scope beats ambition. The teams that won - Unity, Camping World, eye-oo - picked specific, high-volume use cases and got them right. The teams that tried to launch a "general-purpose AI assistant for everything" tend to ship something mediocre everyone ignores.

Handoff is part of the product. Every successful deployment has a clear path from AI to human, with full conversation context preserved. If your handoff makes the customer repeat themselves, you've lost the trust the AI just built.

Personalization compounds. Frag Magenta and the Zalando Assistant outperform generic chatbots not because their NLU is better, but because they have authenticated access to customer data and use it to ground every response in the customer's actual situation.

Cost routing is a real lever. With open-weight frontier models priced at fractions of a cent per resolution, sending every query to Claude Opus 4.7 is leaving money on the table. Route routine traffic to DeepSeek V4 Flash, GLM-5.1, or MiniMax M2; reserve the expensive frontier for the escalations.

AI Actions are where the real ROI lives. A bot that answers questions saves you read time. An agent that books the meeting, issues the refund, swaps the SIM, or completes the purchase saves you the entire downstream workflow.

Long context changes what's possible. With 1M to 2M-token context windows on Claude Sonnet 4.6, DeepSeek V4, and Gemini 3.1 Ultra, you can fit your entire knowledge base, the customer's full history, and your policies into a single prompt. RAG becomes a tuning lever rather than a hard requirement, and a class of "the agent forgot what we discussed three messages ago" failures simply disappears.

Common pitfalls to avoid

Worth flagging the ones we see most often:

• Launching without a fallback. If the AI doesn't have a clean path to a human for the cases it can't solve, it becomes the wall customers used to hate.
• Training on stale docs. An AI agent confidently quoting a policy you changed six months ago is worse than no agent. Set up automated re-ingestion when your knowledge sources update.
• Skipping observability. If you can't see which conversations the agent handled well and which it dropped, you can't improve it. Every deployment needs analytics on resolution rate, escalation rate, and customer satisfaction by topic.
• Assuming one model fits everything. The teams getting the best ROI are routing - fast, cheap models for triage and FAQ; frontier models for nuance.
• Treating AI as set-and-forget. The good deployments treat their AI agent like a new team member: they review transcripts, refine instructions, expand its tool access as it earns trust.

Where this is going

The next 12 months are going to make 2025-era chatbots look prehistoric. With agentic tool-use models like Kimi K2.6 and GLM-5.1 already running multi-hour autonomous workflows, the next frontier is AI agents that don't just answer customer questions but proactively manage relationships - flagging accounts at risk of churn, reaching out before a customer has to ask, resolving issues the customer hasn't noticed yet.

If your support team is still measured in tickets-per-agent-per-day, that metric is on its way out. The new metric is something closer to: how much customer outcome can your team produce, with the AI doing the bulk of the throughput and humans focused on the moments that genuinely require human judgment.

The teams that started shipping in 2025 are already a year ahead. The good news is that the gap is closable in weeks, not quarters - the platforms now exist to do it.

If you want to see what your support stack looks like with a branded AI agent doing the heavy lifting, start building on Berrydesk. Pick the model, point it at your docs, brand the widget, wire up the actions you care about, and ship.