AI customer service assistant
The assistant answers only from your own texts; when it cannot find the answer there, or the topic is sensitive, it hands the conversation over to someone on your team. Before it goes live, it is tested with real customer questions.
What is an AI customer service assistant?
An AI customer service assistant is a chat assistant that answers website or WhatsApp questions from the company’s own documents: FAQs, returns and delivery terms, product information. Its language model looks for answers only there; when it finds none, or the topic is a payment, complaint or personal data, it hands over to staff. No assistant is free of mistakes.
What it covers
Knowledge base
FAQs, returns and delivery terms and product information are gathered; texts that contradict each other or are out of date are weeded out. Every text gets an owner and a way of being updated.
Answer rules
The assistant answers only with what is in the source; when the source has no answer, it says so and hands over. Its language and tone are set to your brand voice.
Handover to a person
We write down which topics always go to a member of staff, who the handover goes to and through which channel, and what the customer is told outside working hours.
Data map
One page shows which data passes through which tool, and with which provider and in which country the model processes it. If a step does not need personal data, none is sent to the model.
Testing with real questions
A test set is built from questions customers have asked in the past. Every answer is marked as correct, incomplete, wrong or “should have been handed over”.
Logs and review
Conversation logs are read; the questions the assistant could not answer are listed, and the gaps in the knowledge base are filled from that list.
How it runs
- Question and document inventory.
In the first week, recent questions are counted by topic and each is matched to the document that holds its answer. The data map is drawn in the same week.
- Knowledge base and rules.
Missing answers are written and contradictions resolved. The answer rules and the topics that always go to a person are set.
- Testing.
Before going live, the assistant is tested on the real questions in the test set. We trace each wrong answer to its source; most often it is the document that needs correcting.
- Pilot.
The assistant opens on one channel and a limited set of topics. At first, draft answers are approved by an agent; we track the share of drafts approved without changes.
- Written decision and monitoring.
At the end of the pilot, the topics the assistant will answer directly are agreed in writing. The conversation logs continue to be read, and a summary is in the Monday report.
What you hold at the end
- A working assistant and knowledge base
- Written answer and handover rules
- A data map
- The test set and test results
- A list of the questions it could not answer
- A guide to updating the knowledge base
When this is not the right thing to buy
- If you want an assistant that answers your customers with no human oversight at all, that is not something we build.
- If the answers currently live only in your staff’s heads, or your documents disagree with each other, those texts have to be written first; an assistant cannot find an answer that does not exist.
- If few questions come in and most of them call for a case-by-case decision, a good FAQ page and a set of ready-made reply templates will do the job.
Questions
The questions we hear most about this work. If yours is not here, write to us and we will answer within one working day.
How long does it take to set up an AI customer service assistant?
The state of your documents decides most of it. The first week goes on the question and document inventory; if most answers are already written down, the knowledge base comes together quickly, and if not, those texts are written first. Testing and a pilot on one channel follow. The pilot ends with a written decision on which topics the assistant will answer directly. More channels, and integrations such as an order system, add to the time.
What do you need from us for the assistant?
The documents the answers will rest on: FAQs, returns and delivery terms, product information. Where possible we also need real questions customers have sent recently; we build the test set from them and strip out personal details. You need to name the person or team the conversations will be handed to, and give access to the channel the assistant will work on. The AI provider’s account, like the other accounts, is opened in your name.
What determines the fee for an AI customer service assistant?
The number of channels and languages the assistant will work in, the volume of documents and how much tidying they need, integrations such as order status look-up, and the volume of conversations to be read during the pilot. The usage cost of the language model is a separate item; it is paid straight to the provider from your own account and grows with the number of conversations. The first meeting is free; 2 to 3 days later we send a proposal that lists each item.
What happens if the assistant gives a wrong answer?
We cannot promise a wrong answer will never appear; no assistant is free of mistakes. We reduce the risk in four ways: the assistant answers only from your documents, it hands over when it cannot find the answer, topics involving payments, complaints and personal data always go to a member of staff, and conversation logs are read throughout the pilot. When a wrong answer is spotted, we trace it to its source and correct the document or the rule.
Where is customer data processed?
At the start we draw a data map showing which data passes through which tool. We choose the AI provider together with you, looking at the country where the data is processed and the provider’s data usage terms. If a step does not need personal data to reach the model, we do not send it. Whether your privacy notice covers this use is a legal question; we recommend reviewing it with your own legal adviser.
Which channels can the assistant work on?
Most often two: the chat window on your site and WhatsApp. Running an assistant on WhatsApp requires a provider connected to the WhatsApp Business Platform. We recommend starting with one channel: once the knowledge base and the handover rules have settled there, adding a second is easier. We decide with you which channel to start with, by looking at where most questions come from in the inventory week.
How is the assistant updated when our documents change?
The assistant says whatever the knowledge base says; if your returns terms change and the text stays as it was, it gives the old answer. That is why every text has an owner, and why we write a guide to updating the knowledge base at handover. When a text changes, the questions on that topic in the test set are run again. For information that changes often, such as promotions and prices, we agree the update routine before bringing it into scope.
Further reading
- Business process automation examples: 13 workflowsFrom sales to accounts, from the warehouse to customer support: the workflows small and mid-sized companies build, what they save and where they break.
- How to choose an automation agency: what to ask firstAccount and workflow ownership, the process inventory, error alerts, handover documents, data protection and a pilot: what to settle with an automation agency before you sign.
- Where to start with automation: inventory, scoring, pilotStarting with the one process that takes the most time gets results faster than a big plan.