Customer Service AI Assistants
Answering repeat questions from the business’s own approved facts and routing the rest to the right person — while complaints, refunds and exceptions stay with a human.
The real load in customer service is not writing an answer, it is writing the same answer twenty times a day. Opening hours, price ranges, the address, the cancellation terms and delivery times account for most inbound messages, and all of them are facts already written down somewhere.
The assistant’s job is to take on that repetition and the team’s job is everything left over. So the first step of a build is not writing a bot but collecting the business’s correct facts in one place: the service list, the hours, the terms and the exceptions. The assistant speaks only from that source.
We do not allow free generation. When a question has no match in the source, the assistant’s default behaviour is to hand the conversation to a person rather than invent an answer. A support line that gives wrong information costs more than one that answers late.
What this service covers
Scope is bounded by the channels you run and the facts you have written down: the chat window on the website, WhatsApp, email and any helpdesk software. If the facts are not written down, that comes first; the longest step of a build is usually not the technical connection but deciding who signs off on the correct answer.
Source of facts and an approved answer set
Services, hours, price ranges, cancellation and delivery terms are collected into one source and every answer gets an approved owner. The assistant does not step outside that source; a question with no match drops straight into the handover rule.
Channel setup and one conversation history
The website, WhatsApp and email are connected into a single flow. The same customer writing from two channels no longer creates two separate conversations, and the team does not have to answer without seeing the history.
Classification and routing
Inbound enquiries are separated by subject: information request, booking, technical issue, complaint. Each class is tied to a team and to a response time expectation, and who the next enquiry is waiting on stays visible.
Handover rules and out-of-hours behaviour
Which subjects stop the assistant, and who it hands the conversation to, are defined in writing. An out-of-hours enquiry is not lost: it is recorded, the customer is given a realistic response time and it enters the queue the next day.
Logging, review and updates
Every conversation is logged, and which answer came from which source can be inspected. Wrong and incomplete answers are reviewed regularly and written back into the source; a source that is never refreshed starts producing wrong answers.
How we work through it
Timing depends on the number of channels and on how ready the business’s written facts already are. The schedule below describes a typical start.
- Stage 1 — first week
Question inventory
Recent months’ messages are separated into subjects and we work out which questions repeat and how often. The scope is chosen from that list, starting with the headings that repeat most.
- Stage 2 — weeks 2-3
Writing the source of facts
Answers are written, approved and given an owner. Skip this stage and the assistant speaks from the scattered text on the site and gives two different answers to the same question.
- Stage 3 — weeks 3-5
Pilot on one channel
The assistant goes live on a single channel and a limited set of subjects first. The aim of the pilot is not volume but seeing which question it gets wrong.
- Stage 4 — ongoing
Expansion and upkeep
Findings are written into the rules and the scope opens to new channels and subjects. If changed prices, hours and terms are not updated, the assistant starts giving wrong information over time.
Where human approval stays
On the items below the assistant informs or prepares a draft and a person decides. This section is a build rule, not a setting that can be switched off:
- Refunds, cancellation fees and discounts: the assistant reads the terms and shows the amount, it does not approve the transaction. Approval stays with a person and nothing completes without it.
- Complaints and negative feedback: the conversation goes to a person after the first message. The assistant never offers an apology, compensation or a goodwill gesture on its own.
- Judgement in health, legal and financial matters: the assistant relays written facts and routes; diagnosis, advice and interpretation stay with the qualified person. That boundary is not our preference, it is the regulation.
- Individual pricing and quotes: no figure outside the written range is produced and no negotiation happens; a conversation asking for a quote is handed to sales.
- Sensitive data such as identity, payment and health details is never requested inside the chat. Any step needing it moves to a secure channel and to a person.
What we measure
Measurement is not built on “how many messages were answered” but on whether the customer’s question actually closed:
- The share of conversations the assistant closed on its own, and the direction that share moves over months.
- First response time per channel, read separately for in-hours and out-of-hours, because the two do not carry the same expectation.
- The subject spread of handed-over conversations. The same heading being handed over repeatedly points at a gap in the source of facts.
- How often a customer writes a second time about the same thing; a second message exposes a conversation that looked closed but was not.
- The number of cases where a support conversation turned into a booking, an order or a quote.
Frequently Asked Questions
Let us work out which questions you answer most often
We separate your inbound messages into subjects and mark which ones a written answer can close and which have to stay with the team.
Get in touch