There’s a wide gap between a chatbot that can answer “what are your opening hours” and an AI agent that can actually resolve a customer issue end to end. Most businesses have deployed the former and called it AI customer service. The more interesting question is what the latter looks like in practice, and which situations it’s actually suited to.
Why Classic Chatbots Hit a Ceiling
Rule-based chatbots and even early generative AI chat interfaces share a fundamental limitation: they can retrieve information and generate text, but they can’t take actions. They can tell a customer their refund status. They can’t initiate the refund.
An AI agent changes this. An agent connected to your order management system, your helpdesk, and your email platform can look up an order, verify the return policy, create the return label, and send the confirmation email — without any human involvement. That’s the practical gap between a chatbot and an agent.
What’s Available Off the Shelf
Intercom Fin is the most deployed example of a purpose-built customer support agent. Fin connects to your knowledge base and can resolve common queries autonomously, with a clear escalation path to a human agent when it can’t. Intercom has reported resolution rates above 50% for teams with well-maintained knowledge bases — meaning more than half of incoming conversations never need a human to touch them. That’s not a marginal gain; for a team handling thousands of tickets a month, it’s a material reduction in support costs.
Zendesk AI takes a similar approach. Its Intelligent Triage feature classifies incoming tickets automatically — intent, language, sentiment — and routes them or resolves them based on configured automation. The agent layer handles order status queries, account questions, and information requests without human intervention, and hands off with full context when a human is needed.
Freshdesk’s Freddy AI agent offers comparable functionality and tends to be the better fit for smaller businesses on tighter budgets. It handles FAQs, creates tickets from conversations, and escalates with structured summaries when a human needs to take over.
Building a Custom Agent for Complex Workflows
Off-the-shelf tools cover the common patterns well. For businesses with more complex requirements — technical support where the agent needs to query diagnostic APIs, financial services where the agent needs to verify identity before acting — a custom build using a framework like LangGraph gives you the flexibility the packaged tools don’t.
The basic architecture: a primary reasoning agent receives incoming queries, decides which tools to call (ticketing system, knowledge base, product API), processes the results, and either resolves the issue or hands off to a human with a structured summary. LangGraph’s stateful graph model makes the escalation logic explicit and testable, which matters when you’re deploying this in a regulated environment or one where the consequences of a wrong action are meaningful.
The integration work is where the real effort goes. Connecting an agent to your CRM, ticketing system, and order platform takes several weeks of development. The reasoning and tool-calling layer is relatively straightforward; the data plumbing is the hard part.
What Agents Handle Well and What They Don’t
Agents are genuinely good at high-volume, repeatable tasks where the right action is deterministic given the customer’s situation. Order status, refund requests within policy, subscription changes, password resets, appointment bookings, account updates. Anything where you could write a flowchart and a human follows it every time.
Agents are not good at situations requiring genuine judgement, relationship management, or novel problem-solving. A long-standing customer who’s furious about a series of failures needs a human who can make a discretionary call. An agent will follow policy. A good support rep might offer something outside of policy and save the relationship.
Fair enough — that’s not the agent’s job. The agent’s job is to handle the high volume of solvable queries so human agents can focus on the situations where human judgement actually matters.
Getting Escalation Right
The escalation design matters as much as the agent itself. An agent that escalates with a structured summary — what the customer asked, what was attempted, what information was gathered, what the customer’s account status is — enables a human agent to pick up immediately without asking the customer to repeat themselves. That’s the experience difference between AI support that customers tolerate and AI support that customers actually prefer.
Build escalation triggers based on: confidence threshold (agent doesn’t have a high-confidence answer), query type (anything touching refunds above a certain value, anything mentioning legal or complaint), and customer tier (if you have SLAs for certain customers, don’t let the agent delay them).
GDPR Considerations
Deploying an AI agent in your customer support flow means personal data — names, order histories, account details, potentially health or financial information — flowing through your AI pipeline. Your data processing agreement needs to cover this explicitly. If you’re using a third-party agent platform, check whether your customer data is used to train their models. Under GDPR, using personal data for model training requires a lawful basis, and legitimate interest arguments for this use are fragile.
The ICO has published guidance on AI and data protection that’s worth reviewing before going live. For UK businesses in regulated sectors — financial services, healthcare, legal — get legal advice on the specific use case rather than relying on a generic privacy policy update.
Starting Right
Resolution rates above 50% are achievable, but they depend on well-maintained knowledge bases and genuinely automatable workflows. Start with your highest-volume query category, build the agent for that single use case, measure it for four to six weeks, and then expand. Trying to automate everything at once is how you end up with an agent that handles nothing well.