First-level support (AI-powered)
First-level support (AI-powered): definition, how it works and use in customer service
Definition · First-level support (AI-powered)
AI-powered first-level support is the first contact tier in customer service, where incoming requests are handled entirely or mostly by an AI agent – without a human actively stepping into the conversation. The agent understands the request in natural language, asks for missing information when needed, retrieves data from connected systems and formulates an answer. Only when the agent can't resolve a request or wants to perform a sensitive action is there a handover to a human (human handover) or an approval step (Assist).
When a company wants to automate its customer service but every second request still ends up with an employee, it's almost always due to the wrong boundaries: the bot takes on tasks it can't solve on its own, or it escalates requests it could easily have answered itself. AI-powered first-level support solves this problem structurally – not with more rules but with a different approach to automation.
What first-level support is – and what it isn't
The term first-level support describes a process tier, not a technology. It's about who handles a request first and which requests can be fully resolved there. The Gartner glossary on customer service & support provides an industry-wide definition of support tiers.
Three distinctions are especially important. Second-level support is the escalation tier: a human takes over when the first level couldn't resolve the request – for example with complex goodwill decisions, legal questions or special cases without a clear rule set. The AI agent is the technical system that delivers first-level support – it isn't a synonym for the process tier itself but the tool used to automate it. Finally, self-service means the customer finds information on their own without an agent actively doing anything – for example via an FAQ article or a status portal. In AI-powered first-level support, by contrast, the agent acts: it reads, asks follow-up questions, retrieves data and formulates an answer.
How does AI-powered first-level support work technically?
A modern AI agent doesn't follow a rigid decision tree. It analyzes the incoming message, chooses the right strategy and uses connected systems – steered by a natural-language behavior description instead of hard-wired node flows. IBM Research explains in "What are AI agents?" what fundamentally sets this approach apart technically from rule-based systems. melibo applies exactly this principle in its AI agents platform: users describe the agent's behavior in natural language instead of building flows.
The process starts with the agent recognizing the intent of the message – understanding what it's about without relying on keyword matching. It then checks whether it has enough information to act. If the order number or email is missing, it asks a targeted follow-up question, always one at a time. Once it has everything, it uses the connected integration – for example, it fetches the order data via the Shopify connection or retrieves the current shipping status via the DHL connection. These connections to external systems aren't technical nodes in a flow but tools the agent uses depending on the situation. Based on the returned data, it formulates a clear, context-aware answer in natural language.
The key difference from a classic chatbot: there's no fixed, modeled path the agent has to follow. It decides based on the message, context and available tools – and can therefore also handle requests that were never explicitly built into a flow builder.
Why do classic chatbots fail at first level?
Classic rule-based chatbots work on the principle: if the message contains X, answer with Y. That works well as long as customers use exactly the phrasing the bot was trained on. As soon as a request deviates slightly, a special case comes up or several topics are mixed in one message, the flow breaks down. According to the Salesforce State of Service report, a large share of all service contacts are repetitive standard requests – but even these come in so many variants that rule-based flows quickly reach their limits.
AI agents solve this problem structurally. Follow-up questions are dynamic instead of hard-wired, system connections are described directly in the behavior text, and behavior changes mean: adjust the behavior text, save, done – no flow rebuild, no newly modeled nodes. Andrew Ng describes in "The Batch" how exactly these agentic design patterns fundamentally improve the performance of language models in practice.
First-level support in e-commerce: typical use cases
AI-powered first-level support is suited wherever requests follow a clear structure and the necessary data can be retrieved from connected systems. In e-commerce, this concerns three core areas.
The order status agent is the clearest example of fully autonomous first-level support. When a customer asks "Where is my order?", the agent first asks for the order number or email, then fetches the order data via the Shopify connection and checks the current shipping status via the DHL connection. Both accesses are read-only, no human approval is needed, and the agent answers on its own. This is made possible by melibo's direct Shopify integration, through which the agent can access order data in real time.
The returns agent works in a mixed mode: creating the return and generating a return label run autonomously. But if a refund is to be triggered, Assist kicks in – the employee sees the suggestion, checks it and approves or rejects it. The complaints agent has the highest need for Assist of the three core scenarios. The agent captures the complaint, asks for photos or receipts and checks the order on its own. As soon as it proposes a solution – whether a voucher, replacement delivery or refund – approval is required for every financial step. Here the handover to the employee is built in structurally, not the exception.
Limits of autonomy: Assist and human handover
AI-powered first-level support doesn't mean the AI decides everything on its own. Anthropic explains in "Building Effective Agents" why exactly these control mechanisms are indispensable for productive agent architectures. Two mechanisms define where autonomy ends.
Assist is the controlled approval of individual actions. When an agent wants to perform a write or money-related action – such as creating a voucher or triggering a refund – that action can be set to require Assist. The agent prepares the action, justifies its suggestion with the proposed input values and waits for an employee's explicit approval. Only then is the action executed, while the conversation with the customer continues. Read-only access – such as retrieving order data or checking a shipping status – typically runs without approval. Write and money-related actions should require Assist by default.
Human handover is the complete transfer of the conversation to an employee in live chat. It applies when a request is too complex, the customer explicitly asks for a human or the agent can't solve the problem. The employee sees the full conversation history and can take over seamlessly. Intercom's overview of AI in customer service gives a good summary of best practices for the handover point between agent and human.
What to consider when rolling it out
AI-powered first-level support stands or falls with the right boundary between autonomous actions and those requiring approval. The most common mistake when rolling it out: too many actions are set to require Assist because it feels safer – the result is an agent that waits for approval on every second request and wipes out the efficiency gain. The rule of thumb: read access without approval, write and money-related actions with Assist by default.
The second critical point is the agent's behavior description. It is the central configuration tool and completely replaces the classic flow builder. If you phrase it too vaguely, you get an agent that answers but doesn't reliably trigger the right system access at the right time. Concrete instructions such as "If the order number or email is missing, ask for it politely – always one question at a time" work better than general phrasing such as "Be helpful and ask for missing information".
The third point concerns system connections. An AI agent is only as good as the data it can access. If you don't connect Shopify, DHL or other systems directly, you force the agent to answer from general knowledge – with all the risks that brings.
Conclusion
AI-powered first-level support isn't a marketing phrase but a precise description of an automation tier: the first contact tier in customer service is taken over by an AI agent that handles standard requests autonomously, obtains human approval for sensitive steps (Assist) and hands genuine exceptions over to an employee (human handover). Natural language instead of node flows, tools instead of nodes, behavior instead of rules – that's the core of the approach. melibo shows what else AI can do in customer service in its own overview of AI agents.
