Customer service automation does not start with a chatbot on your website. It starts in the back office: qualifying enquiries, drafting replies for approval and updating systems — where AI shortens handling time before it ever speaks to a customer.
The chatbot is the most visible element of service automation — which is why it is often bought first. It is also the riskiest: it talks to customers unsupervised, and every invented or off-target answer spends trust that took years to build. Companies that have been burned by “automated” service were not burned by automation — they were burned by the wrong order.
Meanwhile, most of the time in customer service is not spent on the conversation itself. It goes on everything around it: working out what the enquiry is about and who should handle it, finding the order and the customer’s history, checking the procedure, writing the reply, copying the outcome into the CRM. Those stages automate safely, because an agent stands between the system and the customer and approves the output — and the customer gets an answer from a person, only faster.
When enquiries require not just an answer but actions across several systems — checking stock, issuing a correction, changing an order — that is the line beyond which the work belongs to an AI agent. When an agent beats ordinary automation, and how its permissions are set, is covered on our AI agent implementation page.
Customer service is, by definition, the processing of personal data — names, addresses, purchase history, sometimes health or financial details customers volunteer without being asked. Before any enquiry reaches a model, you have to settle which data leaves the company at all, where it goes and on what basis — from the choice of provider and processing region to limiting the data to the minimum a given step needs. How we approach it, and what to ask any supplier, is set out on our security and GDPR page.
We start by counting the distribution of topics in your real enquiries from recent months — it says what to automate first and what result can honestly be expected. Then one process, the narrowest possible pilot in approval mode, and decisions to loosen oversight made on pilot data. The full course is set out in our implementation methodology, and the full scope of the service — from process audit through to upkeep — on our AI implementation page. The general conditions a process must meet to qualify for automation are set out on our AI automation page — customer service meets them more often than most processes in a business, because enquiries are text, they recur, and they leave a trail in systems.
Cost is driven by the number of systems to connect, the state of the knowledge base, and the level of certainty the answers require. The biggest hidden line is often the knowledge base itself: if the answers live in the agents’ heads, writing them down is the first stage of the project — and it stays with the company whatever happens to the automation. All the components, upkeep included, are broken down in our note on what an AI implementation costs.
No — and confusing the two is expensive. A chatbot is the last, most visible and riskiest element: it talks to the customer unsupervised. Most of the value sits earlier and is invisible to the customer: qualifying and routing enquiries, drafting replies for an agent to approve, filling in the CRM, surfacing answers from the knowledge base. Those stages can be automated without the risk of a customer receiving an invented answer — because a person stands between the system and the customer.
The honest answer: nobody knows until the distribution of topics in your real enquiries has been counted — and that is where we start. In a typical support inbox, a handful of topics accounts for most of the volume: order status, returns, invoices, passwords. Those automate best. The rest is a long tail where we automate the preparation of the answer rather than the answer itself. Declaring a percentage before counting the distribution is a promise without cover.
Customer enquiries are personal data, and sometimes special-category data too — when a customer volunteers health information, for example. So this has to be settled before the integration, not after it. You can control which data reaches the model, where it goes and on what legal basis: from the choice of provider and processing region to limiting the data to the minimum a given step needs. The principles, and the questions to ask any supplier, are set out on our security and GDPR page.
Cost is driven by the number of systems to connect — mail, CRM, shop, order system — the state of the knowledge base, and the level of certainty the answers require. A company with a tidy knowledge base and one CRM will pay less than one where the answers live in the agents’ heads. Upkeep is a separate line: per-operation fees and keeping the knowledge base current, because the offer changes. We break the components down in our note on what an AI implementation costs.
They are put off by a bad machine — one that does not understand the question, cannot hand the conversation to a person, or pretends to be one. That is why we recommend the reverse of the usual order: first automate the back office, where the customer talks to a human faster than before, and only then — on topics where the system has proven its quality — direct contact, clearly labelled as automated, with a handover to an agent always available.
Tell us where they come from and what takes the most time. We will say which stage we would start with — including when the answer is that it is not worth it yet.
Book a free consultation