Customer support automation

This page is for UK businesses where a small support team spends much of the day answering the same questions and sorting messages into the right pile. I set up the workflow automation behind your support inbox: it sorts each incoming message, answers the routine ones from replies you have approved, sends everything else to the right person and makes sure nothing sits unseen.

The aim is not fewer people talking to customers. It is your people spending their time on the conversations that need judgement, while questions such as “where is my order?” or “how do I change my booking?” get a correct answer without waiting in the queue.

Who this suits, and who it does not

It suits teams of roughly one to ten people handling support by email, web form, live chat or a helpdesk such as Zendesk, Freshdesk, Help Scout or Gorgias, where a large share of messages are repeats. Online shops, subscription businesses, software companies, training providers and booking-based services tend to fit well, because their repeat questions follow a pattern you can see in the inbox.

It is a poor fit if you get a handful of messages a week; saved replies and a shared inbox will serve you better. It is also not the page for building a conversational bot on your website or WhatsApp, which is a separate chatbot service. This page covers what happens to a message once it arrives, whichever channel it came through: how it is sorted, answered, routed and escalated.

What gets automated

What I deliberately leave with people

Some messages should always reach a person. I build these as fixed rules, not settings that someone can switch off by accident:

• Complaints, and anything mentioning a solicitor, an ombudsman, Trading Standards or a chargeback.
• Messages that suggest a customer may be vulnerable: bereavement, illness, money trouble or distress. If you are regulated by the FCA, the Consumer Duty and its guidance on vulnerable customers make this a compliance matter as well as a decent one.
• Refunds, cancellations or account changes above a value you choose.
• Data protection requests, such as a customer asking for a copy of their data or for it to be erased. These usually have a one-month deadline under UK GDPR, and a named person needs to own each one.
• Anything the system is not confident about.

UK GDPR also places extra conditions on decisions made solely by automated means that have legal or similarly significant effects on a person. So I do not build automations that refuse a refund or close an account without a person reviewing the case. That is the human in the loop principle, applied to support.

How a build runs

1. Read the inbox. I review a sample of recent tickets, with personal details masked where your helpdesk allows, and group them by question type. This shows how many are genuine repeats and which only look simple.
2. Agree the rules. Together we decide which categories get an automatic reply, which get a suggested draft, which go straight to a person, and what your response targets are.
3. Approve the answers. You write or sign off every automatic reply. If your help articles are out of date, they get fixed first, because automation repeats whatever it is given.
4. Build in your accounts. I connect your helpdesk, shop or booking system and CRM using their own automation features where those are good enough, and Zapier, Make or n8n where they are not. Everything lives in accounts you own.
5. Run it quietly first. For an agreed trial period the system tags, routes and drafts but sends nothing to customers, so we can compare its decisions with what your team actually did.
6. Switch on in stages. Automatic replies go live one category at a time, starting with the lowest risk, and I review the early tickets with you before adding the next.

Problems I see most often

• Acknowledgements that answer nothing. A “we have received your message” email adds one more message to every thread and tells the customer nothing new.
• No way out. Customers caught in a loop of automated answers with no visible route to a person, which turns a simple question into a complaint.
• Stale answers. The returns policy changed in spring, but the old wording still goes out automatically every day.
• AI replies without limits. A model writing free-form answers can state a refund rule or delivery date that does not exist, known as hallucination. I keep generated text as a draft for a person unless it is grounded in content you have approved.
• Customer data sent where it is not needed. Whole messages, with names, addresses and occasionally card numbers, passed to tools that only needed the topic.

Customer data and your obligations

Support messages are full of personal data, so every tool in the chain matters. You remain the data controller; the helpdesk, the automation platform and any AI provider act as processors and need data processing terms in place. I pass each tool only the fields it needs, prefer providers that let you turn off use of your data for training, and set deletion steps to match your retention policy. I am not a lawyer, so the contracts and your privacy notice should be checked by your adviser.

What you receive

• A summary of your ticket sample: question types, how often each appears and which I recommend automating.
• The routing, escalation and reply rules, written in plain English and agreed before anything is built.
• The automations, built and tested in your own accounts.
• A reply library your team can edit without touching the automation.
• A record of which tools process which customer data, to support your privacy notice.
• Documentation for each workflow (what starts it, what it does, what to do when it fails) and a recorded walkthrough for your team.
• Any support after handover, agreed in writing before work starts.

Pricing

The build cost depends on how many channels and systems are connected and how many question types are automated. Each quote is written after a free first call, in GBP, and agreed in writing before work starts. Running costs are separate and paid directly to each provider: your helpdesk plan, any automation platform subscription and, if AI classification is used, the model usage, which is billed by volume.