Book a demo
AI Customer Service

Customer Service Automation: What to Automate First

Gaurav Mukherjee
Gaurav Mukherjee · Co-founder & CTO, Flowcall
·19 min read
Customer Service Automation: What to Automate First

Customer service automation is software doing customer service work without a human agent: answering questions, routing requests and, at its most useful, completing them. For a consumer brand, completing means the order is actually cancelled, the address actually changed or the return actually booked in your systems, under your policies, with a clean handover to a human agent whenever a rule says no.

Most automation stops earlier. A bot that replies with the cancellation policy has answered. A system that tags the message "cancellation" and queues it for an agent has routed. Neither has done what the customer asked. This guide is about that difference, and how to choose which requests to automate first.

In short

  • Automation works at three depths: it answers, routes or completes a request. Much of what is sold as automation stops at the first two.
  • A request is completed only when the change is made in your systems, under your rules, and the customer doesn't have to come back about it.
  • Automate first the requests that are frequent, follow a rule you can write down, sit in a system software can change, and are cheap to undo if something goes wrong.
  • Measure completions, repeat contacts and reversed actions for each request type. Deflection alone hides the customers who gave up.

Answer, route or complete: three depths of automation

Answer Route Complete
What the customer gets Information: a policy, an FAQ, a tracking link A ticket in the right queue, sometimes with an acknowledgement What they asked for, done: order cancelled, address changed, return booked
What it needs A knowledge base Intent detection and routing rules The customer's order, your rules, and access to change the order, delivery or payment system
What's left for your team Anything beyond information All of the work, just sooner and in the right queue Only the cases your rules send to a human agent
"Please change my delivery address" Replies with the address-change policy Tags it as an address change and queues it for the orders team Checks the order hasn't shipped, collects the new address, updates the order and confirms

All three have a place. Answering suits questions that are only questions, such as "do you ship to Canada?" Routing suits requests a human agent has to decide. Completing is what takes work off your team, because the request never becomes a ticket.

Customer service automation is also different from help desk automation: the triggers, SLA timers and assignment rules that move tickets inside a helpdesk. Those make your team faster at the work. They don't do the customer's request.

Flowcall's AI Agent works in the third column, because a customer who asks for a new delivery address wants the address changed, not the policy. It turns your written procedures into workflows it follows on WhatsApp, Instagram, Live Chat and email. It checks the order's current state in your systems and makes the change only when your required checks, eligibility conditions and approvals pass.

What "completed" has to mean

Many vendors now say their automation "resolves" rather than "deflects". It helps to be exact, because a conversation can end without anything being done. Count a request as completed by automation only when all five of these are true:

  1. The right customer and order. The customer was matched to the order by phone number, email or order ID. When the match is uncertain, the automation asks instead of guessing; acting on someone else's order is worse than asking a question.
  2. Your rule was checked on current data. Eligibility was decided on the order's live state (packed, shipped, delivered, return window open), not on what the customer said or what the data showed yesterday.
  3. The change was made where the record lives. The order was cancelled in the store or order-management system, the address updated on the order, the pickup booked with the courier, the refund started with the payment provider.
  4. The customer knows what happens next. They got a confirmation with a reference and the timeline your policy or payment provider gives.
  5. It stayed done. The customer didn't contact you again about that order within seven days, on any channel, and no human agent had to reverse the action.

One boundary keeps the reporting honest: a change in your system isn't the courier delivering or the bank crediting. Automation can start a refund; the money arrives on the bank's schedule. Report "refund started" as the automation's result, and track the rest as its own step.

One address change, three ways

A customer placed a prepaid order last night. This morning she messages on WhatsApp: "I've moved. Can you deliver to my new flat instead?" The warehouse hands orders to the courier around midday.

Answer only. The bot replies with the policy: addresses can be changed before dispatch, so email support with your order number. She emails at 10am. The email team picks it up after lunch, by which time the parcel has left with the old address. It comes back undelivered, and she asks for a refund.

Route only. The bot recognises an address change, creates a ticket and marks it urgent. At 11:30 an agent asks for the order number and new address, waits for her reply, opens the store admin, checks the order is still unpacked and edits the address. It works, with two agent touches and one repeated question, and only because the queue was short that morning.

Complete. The automation finds her order from her WhatsApp number and sees it hasn't gone to the courier. It sends her a short form for the new address, checks the courier delivers to the new postcode, updates the order and confirms. No ticket is created. Had the order already shipped, the rule would say so: the automation would tell her the options and hand the case to the logistics team with the order, the new address and the shipment reference.

The customer and the channel are the same all three times. What changed is whether the automation could see the order, apply the rule and change the record.

With Flowcall's Shopify integration, the AI Agent can make this change itself: it finds her order, checks it hasn't shipped, collects the new address in a form inside the WhatsApp chat and updates the address on the order. If the order has already shipped, it creates a ticket for the logistics team with the conversation, the order and what's left to do.

Six consumer-brand requests, answered vs completed

In a consumer brand's ecommerce customer service queue, the same few order requests come up again and again. This is what completing each one involves, and when a human agent should take over.

Request Answered Completed The rule it needs Hand over when
Where is my order? "Orders arrive in 3–7 days", or a tracking link Finds the order, reads the courier's latest scan, tells the customer where it is and when to expect it How long a parcel can sit at one scan before it counts as stuck It's stuck past your limit, or marked delivered but not received
Cancel my order The cancellation policy Checks the order hasn't been packed or shipped, cancels it, starts the refund if it was prepaid, confirms The last stage at which an order can be cancelled; how prepaid and COD orders differ It has already shipped and the customer won't take the return route your policy offers
Change my delivery address "Contact us before dispatch" Checks the order hasn't shipped and the new postcode is served, collects the address, updates the order Until when an address can change; which changes need a check (another city, another name) The parcel is already with the courier
Return or exchange The returns policy page Checks the return window and whether the product can be returned, collects the reason and photos, books the pickup or creates the exchange order Return window, excluded products, condition, which reasons need photos Outside the window, an excluded product, or photos that don't match the claim
Refund "Refunds take 5–7 working days" Checks the return arrived or the order was cancelled, starts the refund to the original payment method or as store credit, sends the reference When a refund may start, to which method, and the largest amount it may start without approval Above your limit, a payment dispute, or order and payment records that don't match
Confirm a cash-on-delivery (COD) order Nothing: the customer never asked Messages the customer on WhatsApp after the order, records confirm or cancel on the order, can offer a switch to prepaid Which orders need confirmation; what to do when there's no reply No reply after your follow-ups (the rule decides: hold, cancel or call)

Every "Completed" cell needs two things the "Answered" cell doesn't: access to the system that holds the order, and a rule precise enough to act on.

Five signs your automation answers but doesn't complete

Pull a sample of last month's automated conversations and look for these:

  1. The same customer comes back about the same order within a few days, often on another channel.
  2. Tickets the bot tagged still need an agent to open the store admin or courier portal and do the work.
  3. The bot asks for an order number that the customer's phone number or email already identifies.
  4. Your automation rate counts conversations that simply ended. A customer who gave up looks the same as one who was helped.
  5. Handovers arrive without the order or what's been tried, so the agent starts from the beginning.

Two or more of these usually means the automation answers and routes well, but the work itself still lands on your team.

What to automate first

"Start with high-volume, simple tasks" is the usual advice, and it isn't enough. A simple request can still be impossible to complete, because nothing lets software change the record, or expensive to get wrong. Take your 10 to 15 biggest request types from last month's tickets and ask four questions of each:

  1. How often does it come in? Its share of all contacts.
  2. Can you write the rule down? Could a new agent decide every case from a written rule, without asking a team lead? "Cancel if not yet packed" passes. "Use judgement on damaged items" doesn't.
  3. Can software reach the system? Can the automation read the order and make the change through your store, order-management, courier or payment system? Read access lets it answer with live data; completing needs the ability to make the change.
  4. What does a wrong action cost? An address changed before dispatch is easy to put back. A refund paid out or a replacement shipped is not.
If the request… Start by…
Is frequent, has a written rule, sits in a system software can change, and is cheap to undo Completing it end to end
Is frequent and has a rule, but the system can only be read Answering with live data, and routing the rest with the order attached
Has no rule you can write down Collecting what the agent needs (order, photos, reason) and handing over. Don't automate the decision.
Moves money or goods, such as refunds and replacements Completing within limits (an amount cap, approval for anything outside the rule) once the reversible requests are running cleanly

Applied to order requests, the four questions usually put order status, cancellations before dispatch, address changes and COD confirmation first: each has a clear rule, and a mistake is cheap to undo. Order status needs only read access to courier data, and order tracking automation covers where that data comes from. Cancellation needs write access to the store or order-management system and a clear cut-off, such as "until packed". COD confirmation automation runs before the customer asks anything, so it prevents contacts rather than answering them. Returns and exchanges follow once pickups can be booked through your courier, and refunds come after that, with limits (see returns management, which covers both).

Flowcall's AI Agent comes with ready-made workflows for returns, refunds, exchanges, cancellations, COD and damaged products, which you adapt to your own rules rather than writing them from scratch. For Shopify stores, COD confirmation runs on WhatsApp: the customer accepts or cancels the order in the chat and can be offered an incentive to switch to prepaid.

What to keep with human agents

Automation that completes requests also has to know when to stop (see our guide to escalation management). Keep these with human agents, and let automation prepare them by collecting the details, summarising the conversation and attaching the order:

  • Exceptions and goodwill. A refund outside the window for a long-standing customer is a decision, not a rule.
  • Angry customers, and customers threatening to go public. Route them quickly, with the history.
  • Suspected abuse, such as repeated "not received" claims on orders the courier shows as delivered.
  • Orders above the value limit you set for automatic refunds or replacements.
  • Safety issues, such as an appliance that sparked or a product that caused a reaction.
  • Anything your rules can't describe yet. Those cases are also your list of rules to write next.

Flowcall's AI Agent prepares these handovers: it creates a ticket with the conversation, a summary, the order and what's left to do, and routes it to the right team. In Flowcall's AI Helpdesk, the human agent can then process a refund or replacement from the ticket. Every incoming message is also checked for frustration, aggression, threats to escalate and requests for a person, so customers at risk of escalating skip the queue.

How customer service automation works

The building blocks are the same whichever software you use:

  • Understanding the request, including when the customer changes topic, and picking up order numbers, dates and addresses from free text.
  • Finding the customer and order from a phone number, email or order ID, cautiously.
  • A workflow for each request type: your procedure for cancellations or returns, written as steps.
  • Rules that must pass before any action: eligibility checks, allowed outcomes and approvals.
  • Actions in the systems of record (store, order management, courier, payment), with a log of each action and its result.
  • Follow-ups when it's waiting on the customer for a photo or an address.
  • Handover with context: the conversation, a summary, the order and what's left to do, routed to the right team.
  • Testing before launch and after every change, then review of what it actually did.

Most of this now runs on AI, which is what lets automation read free text and follow a customer who changes topic; our guide to AI customer service covers that side. The rules should be the same on WhatsApp, Instagram, email and chat. Otherwise a customer's outcome depends on where they asked, which is the opposite of omnichannel customer service.

How to measure customer service automation

Deflection, the share of conversations that ended without a human agent, is easy to count and easy to misread, because it includes customers who gave up. Keep it as context. Measure these customer service metrics for each request type instead:

Measure How to count it What it tells you
Completion rate Requests where the change was made in the system and there was no repeat contact about that order within seven days, divided by all requests of that type Whether automation is doing the work, not just ending conversations
Repeat contact Customers who contacted you again about the same order within seven days, on any channel, divided by requests handled Answers that didn't settle the issue
Reversed actions Automated actions a human agent had to undo or correct, such as the wrong order cancelled or the wrong refund amount Whether your rules and data are safe to act on. Review every one.
Asked to repeat Handovers where the human agent asked for something the customer had already given, divided by all handovers Handover quality
Time to completion From the customer's first message to the change being made The speed of the outcome, not of the first reply
CSAT by path Ratings split three ways: completed by automation, handed over, handled by a human agent throughout Whether customers accept automation for this request

A completed request that stays completed is also resolved at first contact, so your first contact resolution should rise with the completion rate. To count any of this, every conversation needs a request type, the order ID and a log of actions and their results. If your tools can't give you the order ID for each conversation, fix that first.

Numbers won't show a refund promised that your policy doesn't allow. Reading the automated conversations themselves, which is what AI quality assurance does at scale, catches those.

In Flowcall, AI Quality Assurance checks the AI Agent's conversations for wrong or made-up answers, loops, late handoffs, and promises or actions it wasn't allowed to make, and quotes the conversation in each finding. For the last row of the table, Flowcall's reports show CSAT for the AI Agent and human agents separately, by channel and team.

How to roll it out

  1. Pick two or three request types using the four questions above.
  2. Write the rules as your best agent applies them, including the exceptions and who approves what. If a rule only lives in someone's head, this step takes longest.
  3. Connect the systems: read access first, so the automation can answer with live data, then the ability to make changes.
  4. Test on real past conversations, including the awkward ones: a partly shipped order, two orders in one message, a customer writing from a different number.
  5. Launch with handover switched on, on one channel or a share of the traffic.
  6. Review every week: every reversed action, and a sample of repeat contacts and handovers. Tighten the rules, then add the next request type.

What to look for in customer service automation software

In a demo, skip the feature tour and ask the vendor to handle one of your real requests. Then ask:

  • Can it make the change, or only reply? Have it cancel an unshipped test order, then show you the order in your store.
  • Where do the rules live, and who can change them? You should be able to read the rule that allowed an action.
  • What happens when a rule fails or a system doesn't respond? It should tell the customer something true and hand over with the order attached.
  • Can you test it on your own past conversations before customers see it, and again after every change?
  • Can you see what it did in each conversation: which workflow ran, which data it read, which action it tried and the result?

The last two are built into Flowcall's AI Agent. You can run simulated customer conversations against a workflow before it goes live and after every change, and turn real failures into tests. In any real conversation, you can see which workflow ran, what data it used and which action it tried.

How Flowcall automates customer service

Flowcall puts automation and your human agents on the same requests. The AI Agent completes what your rules allow, in Shopify natively or in your other order and internal systems through their APIs, and hands the rest to the AI Helpdesk with the work done so far. AI Quality Assurance then groups the requests it couldn't resolve into the workflows, knowledge or integrations worth adding next. Flowcall's team sets up the first workflows and connections with you, and an engagement can start with one workflow.

To see a request completed from the customer's message to the change in the store, book a demo. Qualifying brands get a free proof of concept on the service problem they most need solved, built with their team and tested before anything goes live.

Frequently Asked Questions

What is customer service automation?

Customer service automation, also called automated customer service, is software that does customer service work without a human agent. It answers questions, routes requests to the right team and, most usefully, completes requests such as cancelling an order or booking a return, in your systems and under your rules. When a rule says it can't, it hands over to a human agent.

What are examples of customer service automation?

Five common examples for a consumer brand: giving live order status from the courier's data, cancelling an order before it ships, changing a delivery address, booking a return pickup, and confirming cash-on-delivery orders on WhatsApp. Ticket routing and automatic order updates are automation too, but they leave the work itself to your team.

What is the difference between customer service automation and CRM automation?

Customer service automation handles customers' requests: answering, routing and completing them. CRM automation handles records and internal processes, such as updating contact fields or assigning leads. The two meet when the automation writes the outcome of a request back to the customer's record.

Share:
Gaurav Mukherjee
Gaurav Mukherjee
Co-founder & CTO, Flowcall

Co-founder and CTO of Flowcall, where he leads product and engineering for its AI customer service platform. Previously Principal Software Engineer at F5 and Atlassian.

Connect on LinkedIn

See How 80%+ Automation Actually Works

We'll show you real conversations—AI handling refunds, tracking orders, switching languages—all fully automated, end to end.