AI Customer Service
Your support inbox, sorted and half answered before your team opens it.
What is AI customer service?
AI customer service is a layer over your existing inbox or helpdesk that classifies incoming messages by topic and urgency, drafts replies from your previous answers, and closes the straightforward cases outright once you permit it to.
Support teams lose most of their day to two things: reading each message to work out what it concerns, and typing an answer they have already typed nine times that week.
We add a layer to whatever you already use. It sorts by topic and urgency, drafts replies from your own past responses, and can close the simple cases without a person touching them.
Your team keeps control throughout. Drafts wait for approval until you trust them, and anything unusual goes straight to a person rather than being attempted.
Why we start in draft only
Duara implements support automation from Nairobi for teams whose ticket volume has outgrown the number of people answering it.
We start every engagement in draft-only mode, where nothing sends without a person approving it. Clients sometimes want to skip that phase to see savings sooner. We hold the line on it, because the fortnight of supervised drafting is what tells us which categories are genuinely safe to automate.
Who this is for
- The same questions arrive every day
- Replies are slow during busy periods
- You want faster support without hiring
- Messages get missed or answered twice
Industries we serve
- eCommerce and retail
- Telecommunications
- Financial services
- Utilities
- Travel and hospitality
- Software and technology
Problems this solves
Support automation is usually considered when hiring another agent stops looking affordable.
- Sorting consumes the morning. Somebody reads every message simply to work out what it is about and who should handle it, before any actual work happens.
- The same reply, endlessly retyped. A handful of question types account for most volume, and each one gets answered from scratch every time.
- Queues that collapse under load. Response times are acceptable until a busy week arrives, and then everything slips at once.
- Messages answered twice or not at all. In a shared mailbox with no ownership, cases fall through gaps or get duplicate replies from two people.
What we deliver
A layer over your existing helpdesk that classifies, prioritises, drafts and, where you allow it, resolves, with a person retaining approval over every category until it has earned trust.
Messages sorted by topic and urgency automatically
Draft replies written from your own previous answers
Simple cases closed outright once you permit it
Anything unusual routed straight to a person
Reporting on volume, category and resolution rate
Category-by-category control over what may send automatically
Features
- Classification before generation. Messages are sorted and routed first, which alone removes a large part of the daily overhead.
- Drafts from your own history. Replies are built from answers your team has already sent, so the tone is yours rather than generic.
- Per-category permissions. You decide which message types may send automatically, and expand that list as confidence grows.
- Escalation on sentiment. Angry or unusual messages bypass automation entirely and go to a person immediately.
- Works inside your existing tool. No migration. Your team keeps using the helpdesk they already know.
Benefits
- Faster first response. Customers get an answer sooner, which is the single measure most correlated with satisfaction in support.
- Capacity without headcount. The routine volume is absorbed, so the same team handles more without the queue collapsing.
- Consistent answers. The same question receives the same answer regardless of who is on shift.
- Visibility into what customers ask. Category reporting shows where volume actually comes from, which often surprises the business.
Technologies we use
- Helpdesk integration. Direct connection to the system you already use so nothing moves and no second inbox appears.
- Classification models. Topic and urgency assignment trained on your own historical tickets rather than generic categories.
- Retrieval from past replies. Drafts assembled from answers your team actually sent, which keeps the voice consistent.
- Sentiment detection. Identifying frustration or escalation risk so those cases skip automation entirely.
- Permission rules. Per-category configuration of what may send automatically and what must be approved.
The integration matters more than the model. A system that lives outside your helpdesk creates a second place to check, which nobody does.
How the process works, step by step
Support automation reaches supervised drafting within three weeks, with automatic resolution enabled gradually after that.
-
Step 1
Free ticket analysis
We examine your historical tickets to establish what proportion is genuinely repetitive.
-
Step 2
Category design
We agree the topics and urgency levels that match how your team actually works, not a standard taxonomy.
-
Step 3
Draft-only phase
Every reply is drafted and approved by a person, for at least two weeks, with no exceptions.
-
Step 4
Selective automation
Categories that performed reliably in drafting are enabled to send automatically, one at a time.
-
Step 5
Ongoing review
Monthly review of what was sent automatically and whether any category should be pulled back.
Quality assurance and how we work
What we commit to when automation touches your customers directly.
- Draft-only comes first, always. A minimum two week supervised phase before anything sends automatically. We do not shorten this on request.
- Categories are enabled one at a time. Automation expands gradually, so a problem in one category never becomes a problem across all of them.
- Angry messages never get automated. Sentiment escalation routes those cases to a person before any draft is generated.
- Every automatic reply is logged. You can see exactly what was sent without human review and to whom, at any time.
- Pulling a category back is instant. If a category starts performing badly you can switch it to draft-only immediately without calling us.
Frequently asked questions about AI customer service
Do we have to change helpdesk software?
No. We work inside what you already use, because a system living outside your helpdesk becomes a second place to check that nobody checks.
Can it reply without us checking?
Only for categories you explicitly approve, and only after a supervised draft-only phase of at least two weeks.
Why insist on the draft-only phase?
Because it is what reveals which categories are genuinely safe to automate. Clients sometimes ask us to skip it and we decline.
How does it deal with somebody who is upset?
Sentiment detection routes those messages straight to a person. Nothing is drafted or sent automatically for them.
How does it know what to reply?
It draws on answers your team has already sent, so responses match your existing tone rather than sounding generic.
Which helpdesks does it work with?
The common ones, and most others through their API. We confirm yours before any work is agreed.
How long until it is useful?
Supervised drafting typically within three weeks. Automatic resolution follows gradually, category by category.
What proportion of tickets can it close?
It depends entirely on how repetitive your volume is. The free ticket analysis gives you the number before you commit.
Can we turn a category off?
Yes, instantly and without calling us. If a category starts performing badly you switch it back to draft-only yourself.
Does it work on email and chat both?
Yes, and on messaging channels where your helpdesk already receives them.
How do we know what it sent?
Everything sent without human review is logged and visible, including to whom and when.
Will customers know it is automated?
We recommend disclosure on automated replies. What form that takes is your decision, but hiding it entirely tends to surface awkwardly.
Can service conversations run in Swahili?
Yes, including messages mixing English and Swahili, which is common in customer support here.
What about attachments and images?
Messages with attachments can be routed to a person by default, since those usually need judgement.
Can it update order status automatically?
Where we connect it to your order system, yes. That is scoped separately because it depends on your systems.
What is deflection?
A case resolved without a person. It is the main measure of whether the automation is earning its cost.
Does our team lose visibility?
No. Everything remains in your helpdesk where your team already works, including the automated replies.
What if it drafts something wrong?
During draft-only nothing sends unapproved, which is precisely why that phase exists before anything is automated.
Do we need someone managing it?
Someone reviews the monthly category report. Day to day it needs no attention once categories have settled.
What are the running costs for service automation?
Message volume drives it. You get the projected monthly figure, based on your own ticket counts, before anything is agreed.
Glossary of terms
Support automation terms as we use them.
- Triage
- Sorting incoming messages by topic and urgency so they reach the right person in the right order.
- First response time
- How long a customer waits for any reply, which is the measure most closely tied to satisfaction.
- Deflection
- A case resolved without a person, either by the automation or by the customer finding the answer themselves.
- Draft-only mode
- The system writes replies but a person approves each one before it sends.
- Escalation
- Moving a case out of automation to a person, usually because of sentiment or complexity.
- Resolution rate
- The proportion of cases closed without human involvement, measured per category.
Related services in AI Implementation
Talk to us about AI customer service
Export three months of your tickets and send them over. The analysis tells you what proportion is automatable before you commit to anything.
Start your project