AI customer experience platform: 5 layers to evaluate

An AI customer experience platform can check all the boxes, impress in a demo, and still frustrate customers after launch. A chatbot might resolve a test billing question, then fail when a customer asks about a live order or changes channels to reach support. If the system can’t see the latest interaction or pass its context to a person, the customer starts again. That creates repeat effort for customers and support teams, even when the model’s answers look good in isolation.

Enterprise teams need more than AI that can write a plausible reply. The system needs current customer context and a way to choose the next action. It also needs to reach customers on their channels and route them to human help. Measures must show whether the system resolved a need, not just sent a response.

This article uses five layers as an evaluation framework for AI customer experience platforms. The layers follow a customer request from the first data point to the result. The article also looks closely at channels, since they affect what customer context is available and who handles delivery failures.

Three gaps can leave a successful pilot short of a complete customer experience.

• The AI can’t reach customers where they are. A chatbot may work on a website but not on the messaging channel a customer uses. Even an accurate answer is useless if it can’t reach the customer.
• The handoff drops context. When a customer asks for a person, the agent needs the conversation history and relevant details. Whether the handoff stays on the same channel or thread depends on configuration.
• The model answers from old data. If the AI can’t access a recent order status, conversation, consent, or subscription change, it may give outdated guidance. A stale response is a data and model layer problem.

All three failures point to the same test. Does the platform connect the systems and data needed to serve a customer, or does each tool work on its own?

An AI customer experience platform connects five things: customer data, journey orchestration, AI reasoning, channels, and measurement. These layers share context, so one interaction can move from first contact to resolution. That includes handing the customer to a person when automation should stop.

A point tool solves one job, such as building a chatbot or collecting survey responses. A suite groups related products, but they may not share customer data or interaction context. A platform connects its parts, so data, state, and orchestration can carry across the journey.

Don’t rely on the label. Ask what information carries over between components. Then check whether it reaches the next system when a customer switches channels or moves to a human agent. The five layers below give you a way to check.

A buyer can test one layer at a time, then check how the parts work together. The layers trace a request from customer data through a measured outcome.

The data layer gives AI and orchestration a current view of the customer. That view can include profile details, order status, consent, past conversations, and activity on your website or app. Identity resolution links records that belong to the same person, even when they come from different devices.

Data must be up to date and include the details a task needs. A customer may change an address or receive a refund moments before asking for help. An AI using an old profile may contradict that update. A unified profile doesn’t need to store every detail in one database, but the systems do need a reliable way to fetch relevant context during an interaction. Keep access scoped so the AI retrieves only the fields its task needs.

Buyers should ask which data sources feed the profile, how quickly changes appear, how consent is enforced, and what the AI can use. That context sets up the next step.

The orchestration layer coordinates the next step across a journey. A campaign calendar sends messages on a planned schedule. Real-time journey orchestration can respond to current signals. These may include a purchase, failed payment, support request, or customer reply.

Orchestration can branch a journey based on what the customer does. It can also check eligibility and consent and stop messages from clashing. It can also connect marketing journeys with service conversations. A support issue can then change what happens next instead of leaving a promotional sequence untouched. For example, a new service case can pause a promotional journey until it closes, when policy allows.

Ask whether journeys react to live events or to scheduled segments. Find out how teams handle exceptions, and how a journey stops when a customer needs a person. Those decisions define the task the reasoning layer must handle.

The reasoning layer interprets a request and chooses or recommends a response. A rules-based chatbot follows designed paths. Retrieval-augmented generation can find relevant information and use it to ground an answer. Agentic AI can plan and take bounded actions through connected systems, such as checking a status or submitting a permitted request.

These approaches aren’t interchangeable. A generative model can phrase an answer, but it shouldn’t be allowed to take every action. Each AI task needs clear limits, trusted data, the right permissions, and ongoing testing. There should also be a route to a person when the AI isn’t sure, or policy says so. Match each agent’s permissions to its task and log the actions it takes. A copilot can support an employee without speaking directly to the customer, while an AI agent may handle a defined request autonomously.

A field study of 5,179 support agents at one company found that those using a generative AI assistant resolved about 14% more issues per hour on average. Newer and lower-skilled agents saw the biggest gains. The study shows that AI can help less-experienced agents resolve more issues, not just automate set tasks.

Once the system has decided what to do, the channel layer determines how the response reaches the customer.

The channel layer carries interactions through messaging, voice, email, web, or app experiences. Channel-layer services may include routing, sender or number provisioning, delivery status, channel-specific rules, and fallback paths. A platform channel list is only one part of the evaluation. Buyers also need to know which services the provider operates, and which rely on partners.

A channel connection should let the rest of the platform use the right customer context and journey state. If the channel sits outside that flow, teams may have to reconcile events and outcomes across systems. Delivery receipts and failures should map back to the original interaction. The practical questions are simple. Who controls routing? Who finds the cause when delivery fails? And who’s responsible when a channel changes its rules?

Because delivery alone doesn’t prove success, the final layer measures what happened after the interaction.

The measurement layer connects activity to customer and business outcomes. Useful metrics include automated resolution, CSAT for AI-handled conversations, first-contact resolution, escalation handle time, cost per resolution, and customer effort. Define a measurement window and compare like-for-like journeys before drawing conclusions.

A chatbot that ends a conversation hasn’t necessarily resolved the issue. Containment can include customers who gave up, switched channels, or contacted support again. Pair automated resolution with CSAT or another direct measure of quality. Also check repeat contacts and escalations. NPS and voice-of-the-customer analysis give a wider view, but they can’t replace evidence from the conversations themselves. Compare results by journey, channel, and request type, so a strong average doesn’t hide a weak path.

The channel layer carries the platform’s response or action to the customer. If delivery fails, the rest of the stack can’t complete the interaction.

An AI can write a strong answer and still fail the customer if the platform can’t deliver it through the channel they use. The channel decides what the AI knows, how the conversation carries on, and who fixes delivery problems. Instead of it being a list of integrations, this is ultimately an operating decision.

Ask each provider what it operates directly and what relies on another service. Confirm who manages routing, sender IDs, and delivery monitoring, and who handles incidents and policy changes. Then ask for an architecture diagram and trace one message through a real journey, including fallbacks and any third parties.

Infobip’s global messaging infrastructure runs on 800+ direct operator connections across 190+ countries, with 99.95% uptime. Ask every provider for the same details. Which routes does it run directly? Which depend on partners? What uptime applies to your market?

The right channel mix also changes with the customer and market.

There’s no single channel set for every audience. Preferences vary by market, language, use case, accessibility needs, and local rules.

When you build an AI platform for omnichannel customer experience, start with the channels each journey needs. SMS reaches almost any phone, which suits one-time passwords and delivery alerts. RCSadds branded, interactive messages on supported devices. Viberhas a strong reach in parts of Eastern Europe and Asia-Pacific. Voice suits urgent or complex issues, while in-app messaging and push notifications reach customers who already use your app.

Channel choice also shapes what happens when self-service stops, and a person takes over.

A good handoff gives the agent the conversation history, customer details, and session context. That way, the customer doesn’t have to explain the issue again. Make clear what the AI has already tried. Without that context, the agent starts cold.

Some handoffs also move the customer to a new channel, such as from a Viber chat to a phone call. Ask the provider to show what the agent sees when that happens, and whether the earlier conversation comes with it.

Use the same handoff test when you assess AgentOS’s components. The next section maps them to the five-layer framework.

AgentOS is Infobip’s platform for running AI-led customer experiences. It brings together customer data, journey orchestration, AI agents, a cloud contact center, and analytics. All of them run on Infobip’s own messaging infrastructure. Each capability is a separate module, so a business can start where its biggest gap is and connect the rest to systems it already uses. The sections below follow the five layers in order, from the data that feeds AgentOS to the measurement that shows what it achieved.

The data layer makes customer profiles and interaction history available to AI agents and journey orchestration. The Conversational CDP pulls in web and app activity, CRM records, transactions, chatbot conversations, and contact center calls. It then merges them into one live profile per customer. Because the profile updates in real time, an AI agent can see a refund or address change from minutes before the conversation starts.

That data should feed the orchestration layer.

The Customer Engagement Platform runs journeys across channels and links marketing with service. Its orchestration layer picks the next step based on live behavior, CDP profile data, and predictive scores. That means a failed payment or new support case can change what the customer gets next. Teams build journeys in a drag-and-drop builder and connect their CRM or data warehouse through REST APIs and webhooks.

The next layer interprets the journey and acts on it.

The reasoning layer combines AI Agents, which handle defined tasks, with the AI Chatbot Builder, where teams design conversations. Teams can build rule-based or generative chatbots without code or write their own in Python. Retrieval-augmented generation grounds answers in company content. When a request goes beyond what a chatbot should handle, it passes from chatbot to AI agent to human agent, with the transcript and customer profile attached.

The next layer is human support when AI reaches its limits.

The Cloud Contact Center covers the human part of the journey. At handoff, the agent sees the customer’s purchase history, channel preferences, and segment data. An AI summary also shows what the chatbot already covered. Voice, email, social, and messaging conversations arrive in one inbox and are routed by agent availability and workload.

If a Viber chat needs a phone call, the full conversation history follows the customer. Agents can also start a voice call inside a messaging app, so the customer doesn’t have to switch. That works because the contact center runs on the same channel layer as the rest of AgentOS.

The channel layer underneath all of it

The channel layer isn’t a separate module. It’s the delivery foundation that supports every other layer of AgentOS. It runs on Infobip’s 800+ direct operator connections across 190+ countries, with 99.95% uptime. Profiles, journeys, and AI responses all reach customers through that network, and delivery results flow back into the platform.

The final mapping is measurement across the journey.

Measurement should cover the whole journey, from first message to resolution. That includes channel switches and handoffs to people. Compare automated resolution, CSAT, repeat contacts, and handling time, because containment alone doesn’t tell you enough.

AgentOS Insights and Analytics pulls data from journeys, chatbots, AI agents, and human agents into one view. It spots recurring issues, traces their root causes, and ranks them by how many interactions they affect. That way, teams know what to fix first. Separate dashboards track chatbot sessions, AI agent tool usage, journey conversion, and contact center performance.

Customer examples show what this looks like in practice.

Each example below shows one layer, or a handoff between layers, at work in a live deployment.

Bolt moved driver registration to WhatsApp, where a chatbot guides sign-up, agents check documents in the background, and reminders bring back drop-offs. With fewer applicants lost along the way, driver conversion rose 40%.

Farm Superstores, a Saudi retailer, answers common questions 24/7 with a WhatsApp chatbot, and agents take over with the full conversation history. Automating routine questions cut operational costs by 60%, and revenue grew 4% in the first three months.

LAQO Insurance, Croatia’s first fully digital insurer, runs a generative AI assistant in Croatian and English that passes anything complex to a live agent with the history and customer profile attached. As a result, the assistant handles 30% of customer queries on its own, and 90% of queries are handled within three to five messages.

Mukuru lets customers open accounts, send money, and make payments through a WhatsApp chatbot in 10 languages. Serving customers in their own language on a channel they already use grew active customers more than 30x, lifted read rates to 92%, halved drop-off rates, and cut costs by 15%.

Evaluate the links between customer data, orchestration, reasoning, channels, and measurement. Don’t treat them as disconnected feature demos, and put the same questions to every provider.

Use these questions to test architecture, governance, and customer outcomes across the stack.

• Does AI use live customer data or a scheduled sync?
• How does identity resolution respect consent and access limits?
• Can live signals branch journeys across marketing and service?
• Which tasks can agents perform, and when is approval required?
• Which channels are direct, integrated, and supported in each market?
• Who owns routing, deliverability, and channel incidents?
• What transcript, profile, and session context reaches a human agent?
• Can you report resolution, CSAT, repeat contacts, and handle time together?
• How are AI outputs tested, audited, and corrected?
• Can APIs extend the current stack, or is replacement required?
• What changes when you add a market, language, or channel?
• Which certifications, residency options, channel rules, and SLAs apply?

A provider’s answers show whether components share state or exchange records at system boundaries. Those answers also help you judge whether consolidation would solve a real operating problem.

Consolidation helps when separate tools create duplicate profiles, repeated integrations, unclear ownership, or lost context. It isn’t always the right choice, though. Keep a specialized tool when it meets a distinct need, integrates reliably, and fits your governance requirements.

Weigh that against the cost of switching. That includes migration work, channel dependencies, and the compliance reviews a new provider brings. Move one journey from first contact through resolution before a broad migration. A modular platform can also add the capabilities you need through APIs, without replacing the whole stack. If you go that route, ask how the provider keeps context in sync and handles system changes.

Judge a platform by its connections. Does current data inform the next action? Are AI tasks bounded? Can the channel reach the customer? Does measurement go beyond containment?