Customer support automation: what to automate, and what not to

The layers of support automation

Support automation is not one thing. It is a stack of layers, and most teams already use the first without thinking of it as automation. Each layer saves a different kind of time, and each carries different risk.

• Routing and rules. New conversations are tagged, prioritised and sent to the right queue or person, with alerts when a service target is about to be missed.
• Automatic answers. An AI assistant replies to common questions from your knowledge, or drafts a reply for an agent to check.
• Actions in your systems. The assistant looks up an order, opens a case or records a request in your own tools, so the customer gets a real answer.
• Multi-step workflows. A process with several steps, such as a cancellation that checks the account, waits for a manager's approval, and then records the result.

What to automate first

Start where the work is repetitive, the answer is clear, and a wrong answer is cheap to fix. Count last month's tickets, tag the five biggest reasons customers write in, and begin with the ones that repeat.

For most teams, the first wins are order or booking status, policy questions such as returns and shipping, opening hours and pricing, and simple how-to questions. These are common, they have one right answer, and they are tedious for a person to type again and again.

Start with one channel and a short list. It is better to automate five questions well, and watch the results, than to switch on everything and find the problems from customers.

What not to automate

Some work should stay with people, and deciding that early is part of doing automation well. The usual candidates are emotional conversations, such as a customer who is angry or upset, anything involving money above a set amount, disputes, and sensitive topics such as health, legal advice, payments or account security.

The aim is not to take people out of support. It is to keep them for the conversations where they make the difference, and to make sure those conversations reach them with the context already gathered.

Guardrails that keep automation safe

Automation goes wrong when it can do too much, too quietly. A few limits remove most of the risk, and they are worth setting before launch, not after the first mistake.

• A confidence line. On email and text the assistant answers alone only when it is sure enough, and otherwise drafts for your team or offers a person.
• An easy way to a person. A customer who asks for one gets one, and the handoff carries a summary.
• Limits on lookups. Anything that returns a customer's details should need a reference the customer provides, such as an order number, not a name or an email that anyone could type.
• Approvals for changes. A cancellation, refund or other change to data can wait for a manager to approve before it runs.
• A record. Every action the assistant takes is logged, so you can see what it did and why.
• A test first. Try it on real questions, on each channel, before customers do.

How LiyaSupport handles each layer

In LiyaSupport the rules layer is Automations: rules for tagging, routing and service-level alerts. The answer layer is the AI first responder. When it is unsure it says so and offers a person, and on email it drafts for an agent instead of replying.

Actions connect your own systems. You describe an endpoint the assistant may call, such as an order lookup, and mark who can use it. Lookups that return a person's details are protected by default, so they run on a reference the customer gives and refuse a name, email or phone number.

Workflows chain steps together, including an approval step. When the assistant starts a workflow that needs a person's decision, the conversation gets a note with a link to the approval, and the approver can see the request, where it came from, what has already happened and what approving will set off before they decide.

Measuring the result

Choose your measures before you switch anything on, and write down today's values. A small set is enough: first response time, the share of conversations the assistant handled without a person, customer satisfaction, and how many tickets your team handles per person.

Check the quality of what was automated, not only the volume. Each week, read a sample of the assistant's answers, especially the ones it was unsure about, and fix the knowledge behind any that were wrong. Each fix improves every later answer.