Customer support automation, from first message to resolved case
Most inbound support volume is a small number of questions asked many times, mixed with a few that need judgement. Automation is worth doing when it separates those streams reliably and hands the second to a person with everything they need. This page describes that workflow, where it breaks, and how it is scoped.
Submit a project See the related serviceReviewed by David (CEO) · Updated 23 Sep 2026
Who is in this workflow
- The customer arrives on whichever channel is nearest: a messaging app, email, a web form, a phone call. They will not repeat themselves twice without irritation.
- Front-line agents answer, look things up in two or three systems, and copy the result somewhere else. They are the people who know which answers are wrong.
- The team lead owns queue health, quality review and the escalation rules nobody has written down.
- Subject owners outside support hold the real answer: billing, logistics, engineering, legal. A case waits on them more often than on the agent.
- Automated actors already exist: the autoresponder, the routing rule, the notification job, the form that opens a ticket nobody reads.
How the workflow runs today
A message arrives in a shared mailbox or messaging inbox. Someone reads it, decides what it is about, looks up the customer in another system, and either answers from memory or asks a colleague. The answer is typed fresh each time. If it needs another team it moves by internal chat and comes back, or does not. What was agreed lives in the thread, not in the customer's file.
That works while volume is low and the same three people handle everything. It fails when volume rises, when a second channel opens, or when the people who held the answers in their heads leave.
Where the workflow breaks
-
The same question is answered from scratch
There is no approved answer, so two agents give two versions and neither is wrong enough to correct.
-
Nobody knows what the volume is made of
Without a classified record of intents, a decision about what to automate is a guess dressed as a plan.
-
Routing is tribal knowledge
New agents learn who to ask by asking. When that person is on leave, the case stops.
-
Context does not travel
An escalated case arrives as a forwarded thread, so the specialist repeats the diagnosis the agent already did.
-
Channels diverge
The same customer gets one answer on chat and another by email, because the queues share no record.
-
Knowledge decays silently
A policy changes; the article does not. The error surfaces weeks later in a complaint.
-
Automation is bolted on without an exit
A bot that cannot say "I do not know" and cannot reach a person produces angrier customers than no bot at all.
The target flow
-
Arrival and identification
The message enters one queue whatever the channel, and the customer is matched to an existing record rather than retyped.
-
Intent classification
The request is classified against a defined intent list, with a confidence score and the original text preserved.
-
Automated answer
High-confidence, low-risk intents are answered from approved content, with the source of the answer recorded.
-
Containment check
Anything below the threshold, or in a category marked sensitive, is never answered automatically. The threshold is a written decision, not a default.
-
Escalation with context
The case reaches a person with the classified intent, customer record, conversation and any automated attempt already made.
-
Resolution and record
The outcome is written to the customer record, not only into the thread, so the next conversation starts informed.
-
Feedback
Wrong classifications, poor answers and new intents feed a review cycle that updates the intent list and the approved content.
System boundary
The solution owns intake, identification, intent classification, automated answers from approved content, the escalation rule, the record written back at the end, and the audit trail of what was answered automatically and on what basis.
It does not own the answers themselves, the policy behind them, staffing, or the decision about which categories may never be automated. Those stay with the support organisation. The automation proposes; your policy and your people decide. Keeping that boundary explicit is what makes the workflow auditable when a customer disputes an answer.
Orders, invoices and stock questions resolve in the systems behind ERP and back-office operations. Questions about a specific listing and its seller belong in marketplace and classifieds operations.
Data and integrations
The objects to design are the conversation, the classified intent, the customer record, the approved answer with its version, the escalation event and the resolution. Versioning the answer is what lets you explain, months later, what the system told someone.
Typical connections are the messaging or email channel, a CRM or helpdesk, the systems holding order and account data, an identity check and analytics. Each needs an owner, a sandbox and defined behaviour when it returns an error or an uncertain result: a half-written case is worse than none.
Netbase works with commercial and open-source AI models chosen per project and implies no vendor partnership. Conversation data is personal data: decide residency, retention, consent and what may be sent to an external service before the architecture. Netbase security practice includes secure code review and version control, role-based access control, multi-factor authentication for admin dashboards, contributors under NDA, and NDAs and data processing agreements on request.
Delivery modules
- Intake and channel connection: one queue, identification, message normalisation.
- Intent model: the intent list, classification, confidence thresholds and the evaluation set.
- Approved answer library: content, versioning, ownership and review dates.
- Escalation and routing: rules, context hand-off and the queue a person actually works from.
- System integration: CRM or helpdesk write-back, account and order lookups.
- Quality and audit: sampling, review of automated answers, the audit trail.
- Reporting: intent mix, containment, escalation reasons and reopened cases.
They are built as AI workflow automation milestones with written acceptance criteria.
Rollout
Stage one is measurement: classify several weeks of real conversations by hand, and you learn both the intent mix and which answers your team disagrees about. Skipping this step is why automation lands on the wrong intents.
Stage two is a narrow pilot: two or three high-volume, low-risk intents on one channel, with every automated answer sampled by a person. Stage three widens intent by intent, and only after the sampling shows the previous batch behaving.
Netbase delivers remote-first from Hanoi in Agile increments with weekly reviews, using AI-assisted engineering under human review. Teams draw on business analysis, project management, solution architecture, development, QA and UI/UX roles. Milestone acceptance is described on the project delivery page, and engagement shapes on engagement models.
Risks
-
A confident wrong answer
A generated reply that sounds right and is not is harder to catch than silence. Sample automated answers from day one.
-
No escape hatch
If a customer cannot reach a person, the automation becomes the complaint.
-
Sensitive categories
Complaints, cancellations, safety, payment disputes and anything with a legal consequence belong with a person by policy, written down before launch.
-
Unowned content
An approved answer with no owner and no review date is a future error with a timestamp.
-
Personal data in prompts
Sending conversation content to an external service without a decision on consent and retention is expensive to unwind after launch.
-
Measuring deflection only
Containment that pushes people to a second channel is a moved problem reported as a win. Track reopened and repeat contacts beside it.
-
Agent trust
If agents find the classification wrong and cannot correct it, they work around the system and your data describes a workflow nobody follows.
How success is measured
Baseline first, then track: intent mix and how much of it is repetitive; the share answered automatically and the share escalated, with reasons; classification accuracy on a sampled set; time to first useful reply; reopened and repeat-contact rate; how often an escalated case arrives with complete context; and the age of the approved answer library.
These are your numbers, on your workflow. Netbase publishes no performance figures from past projects, and this page claims none.
Which industries run this workflow
The workflow holds wherever a business answers many similar inbound requests about accounts, orders, bookings, deliveries or service status: retail and e-commerce, logistics, travel and hospitality, education, subscription software, utilities, and professional services with a client desk.
It does not fit regulated advice, where the answer must come from a licensed person, and it does not fit clinical or safety-critical triage. It does not fit a support function whose volume is genuinely varied, where every case is different and there is nothing repetitive to separate out. If the bottleneck is a broken back-office process rather than the conversation about it, automating the conversation only makes the queue polite.
Proof from delivery
For a client who is not named, Netbase built a WhatsApp Business AI chatbot with intent and conversation-flow handling, a large language model API, and CRM synchronisation for lead capture, customer data and workflow automation. GPT-4 appears in that project's recorded stack; naming it describes the stack and implies no partnership. The work was delivered in milestones from design and prototype to documentation, knowledge transfer and 30 days of support, over four to eight weeks.
That record describes what one project contained. It is not a standing offer, not a package, and not a commitment to the same scope, stack or schedule. The documents do not describe a hand-over from the chatbot to a person, so no such hand-over is claimed, and no volume, sector or result is reported. The full record with its evidence limits is the WhatsApp AI chatbot and CRM integration page; how evidence is labelled is explained on the methodology page.
Netbase designed and built an AI chatbot on WhatsApp Business that handles intents and conversation flows and keeps lead and customer data in step with the client's CRM, delivered in milestones from prototype to knowledge transfer.
Keep ReadingCommon questions
No. It should remove repetition from a named set of intents and make every escalation arrive with context. Headcount decisions are yours and follow evidence, not a launch.
Sampling catches it, the audit trail shows what was said and why, and the intent moves back to human handling until the approved content is fixed.
For custom development the client owns the intellectual property created for it. Netbase productized modules and products are licensed rather than transferred, and any a project uses are named in the proposal.
Netbase JSC's head office is in Hanoi, Vietnam, and it is the company's only office. Delivery is remote-first, with onsite work scoped when needed.
Services behind this solution
AI workflow automation with evaluation and human control
AI Workflow Automation starts with one operating workflow, one accountable owner and one agreed way to judge the result. The goal is a workflow that handles the routine cases correctly on representative test cases, routes uncertain or high-impact cases to a person, and can be monitored and changed after handover. It is not a promise that every process can or should be automated.
Learn More
Scope this workflow
Bring a few weeks of real conversations, the channels you run, the categories you will never automate, and the systems that hold the answers. The commercial route is AI workflow automation; other engagements are listed under services. When ready, submit a project brief. OutsourcingVN is operated by Netbase JSC and is Netbase's own outsourcing-services platform.