Why AI in customer service rarely fails because of the model

Most customer service assistants today work on the same principle. When an enquiry comes in, the system first searches the stored documents for relevant passages. Only then does the language model turn them into an answer. Specialists call this retrieval-augmented generation, or RAG: the model does not answer from its own knowledge but from what it is given.

This leads to an awkward chain of consequences. If the search finds nothing suitable, the model fills the gap with a plausible-sounding guess. If it finds two contradictory passages, it picks one or blends both. If it finds an outdated page, it passes on the old information politely and with conviction. The customer cannot tell the difference, because right and wrong answers sound equally confident.

A stronger model changes little here. It writes more elegantly and handles convoluted questions better. It still cannot know a rule that is written down nowhere. Anyone who wants better answers therefore has to start with the sources.

What a knowledge base needs to contain

A knowledge base is the collection of texts the assistant draws its answers from. In small businesses it rarely sits in one place. Part of it is on the website, part in old emails, part only in the head of the person who has handled the inbox for years. Reliable use requires at least four kinds of content.

Facts with a validity date. Delivery times, shipping costs, returns periods and payment methods form the core. Each of these needs a date from which it applies and a person who has confirmed it. Without a date, nobody can later tell which of two versions is correct.

Procedures rather than intentions. A sentence such as “We handle complaints quickly” helps no customer. What helps is what they should actually do: which details to send, to which address and what happens next.

Exceptions. These are almost always missing. What applies to custom orders, international deliveries or reduced items is rarely written down anywhere. Yet these are exactly the cases that reach the inbox most often, because the standard answer on the website does not cover them.

Limits. The knowledge base also needs statements about what the assistant should not answer. These include questions about health, legal claims or individual terms. A sentence such as “I cannot give a binding answer on that, I will pass you on” is the right response here.

Why good help pages are not enough

The obvious move is to upload the existing FAQ page. It works less well than expected. Help pages are written for people who skim a whole page and work out the context themselves. An assistant, by contrast, only sees individual sections, lifted out of their surroundings.

An example shows the mechanism. On the shipping page, under the heading about international orders, there is the sentence “14 days apply here”. A reader knows this refers to returns from abroad. If the sentence reaches the model without its heading, it can be read as a general deadline. Every section should therefore make sense on its own and name its subject itself.

On top of that, readable copy often stays deliberately vague. Phrases such as “as a rule” or “usually within a few days” are acceptable on a website. An assistant turns them into a commitment, and the customer takes it as a promise.

For the knowledge base, it pays to replace such passages with clear values and clear conditions. This runs against many teams’ instincts: the texts become drier, not more elegant. That is the point, because the knowledge base is working material for a machine, not marketing copy.

Who maintains the answers is an organisational question

A knowledge base goes out of date faster than most teams expect. A new shipping option, a changed deadline, a discontinued product: each of these changes also affects the assistant’s answers. If nobody updates them there, it answers questions with yesterday’s information.

One owner per subject area. Shipping, returns, product information and payments each belong to a named person. That person decides which version applies and deletes the old one. When everyone is responsible for everything, nobody resolves the contradiction between two pages.

Changes at the source, not in the assistant. The knowledge base should be fed from the same data the website and shop use. Otherwise two versions of the truth soon exist. For product details, maintenance therefore starts with the product data itself, as described in the article on product data management.

A fixed review date. Even with clear ownership, things slip. A regular pass through the knowledge base, roughly once a quarter, catches them. One checkable sign: every page carries the date it was last confirmed. Pages without a date are suspect.

Escalation rules: when the AI has to hand over

The trickiest part is not the answer but the moment when the assistant should not give one. Without a clear rule, it tries anyway. Escalation here means passing an enquiry to a person, together with the conversation so far.

Content triggers. Complaints about damage, mentions of skin reactions, legal demands and requests for goodwill gestures belong here. A person decides, because the answer has consequences someone has to stand behind.

Situational triggers. A customer asks the same thing a second time, writes in visible frustration or explicitly asks for a person. In such cases every further automated reply does more harm than good.

A destination for the handover. The rule is only as good as the place where the enquiry lands. It needs an inbox or ticket view someone actually reads, and a commitment to when the customer will hear back. An assistant that says “I will pass you on”, followed by three days of silence, damages trust more than no assistant at all.

When AI in customer service is not worth it

Not every business benefits from an assistant. Whether it pays off depends less on the technology than on the kind of enquiries and the people available.

Few enquiries. A business answering a handful of messages a week gains hardly any time. Building and maintaining the knowledge base costs more than the assistant saves. A good help page and a few saved replies do the same job.

Mostly one-off cases. Some enquiries barely resemble each other: individual advice, custom orders, complex arrangements with business customers. Where hardly any answer fits twice, the assistant rarely finds anything useful and escalates almost everything.

Nobody to maintain it. If no one has time to keep the knowledge base current, the assistant gets worse every month. It does not break down; it quietly gives wrong information. That is harder to notice than an outage.

A high cost of error. Where a wrong answer can have legal or health consequences, there is little room for manoeuvre. An assistant can sort enquiries and prepare information there. The answer itself should come from a person.

How to check whether the groundwork is sufficient

A robust test needs no new tool, only real enquiries from the past. Collect fifty to a hundred messages from recent months, across all topics. Deliberately include the awkward cases.

For each enquiry, the team answers one simple question: is the correct answer clearly in the knowledge base, yes or no. Where it is missing, contradictory or exists only in someone’s head, the knowledge base is not ready yet. This check needs no AI at all. It is often more revealing than any trial run.

Only then is a test with the assistant itself worthwhile, internally at first. The team puts the collected questions to it, compares the answers with the knowledge base and notes where the assistant guesses instead of quoting. Only once the errors are confined to a few known gaps does it go in front of customers.

After launch, the maintenance begins

A customer service assistant is never a finished project. Every enquiry it cannot answer or has to pass on reveals a gap in the knowledge base. Reviewing these cases regularly improves the sources step by step, and with them the answers.

How we look after existing AI systems and their knowledge bases in day-to-day operation is described under AI Operations Care.