Skip to content

HomeAI chatAI customer service

An assistant that knows what it is not allowed to answer.

For AI in customer service, CoderConda’s assistant sorts every enquiry on arrival, before a ticket exists: what it closes, what it prepares and what goes straight to a person. That way forty identical delivery-status questions do not push the one complaint to the back.

Channels
Mail, chat, phone, WhatsApp
Sorted
Before the ticket
Measured
Solved without a return

02Not how much a machine takes away, but what it should not touch.

01 · THE PROBLEM

It is not the hard cases that cost the time. It is the identical ones.

  1. Monday

    The same question, four channels

    Delivery status gets asked in the inbox, in the chat, over WhatsApp and on the answering machine. Each reply takes two minutes and has to be written separately.

  2. Tuesday

    The complaint gets the quick answer

    In between sits a case that needed a good one. It gets a short one, because eighteen more messages are waiting behind it.

  3. Wednesday

    Two cases, one request

    A customer writes a second time because the first answer got lost. Now two cases for the same matter sit in two channels.

  4. Friday

    The pile is empty, the case is not

    Everything short has been handled. Tuesday’s complaint is still open and starts next week at the top, only angrier.

02 · THE FOUR PILES

One decision per request, and the fourth pile is the one that matters.

  1. Pile 01

    The assistant closes it

    Delivery status, opening hours, dimensions, warranty periods, part numbers. Answered from your own sources, with provenance, without anybody looking at it.

  2. Pile 02

    It prepares it

    Category, summary, customer history and a draft reply are ready. A person reads, changes a sentence and sends.

  3. Pile 03

    It hands it over at once

    Complaints, goodwill, exceptions — anything with money or annoyance attached. The assistant records and forwards without forming a view of its own.

  4. Pile 04

    It never touches it

    Cancellation rights, terminations, warranty, legal questions, complaints about individuals, anything health-related. That list lives in the contract, not in an instruction to the model.

Solved without a return

03 · WHAT GETS BUILT

Six parts, and one of them is a list of prohibitions.

  1. 01Mail · chat · phone · WhatsApp

    Sorting before the ticket

    Every incoming message gets a category, an urgency and a pile. What gets closed never becomes a ticket and appears in no queue.

  2. 02A source per answer

    Answers with provenance

    Answers come from your own sources: service pages, price list, delivery data, contract terms. For every answer it is traceable where it came from.

  3. 03On request · at once · with the thread

    The handover point

    One sentence is enough to reach a person — no loop, and no need to tell the whole story again.

  4. 04Cancellation · goodwill · law · health

    The register of prohibitions

    Subjects the assistant does not answer even where it could. They stand in the contract and get checked, not merely mentioned in an instruction.

  5. 05The existing system, not a new one

    Connected to what is there

    Ticketing, inbox, inventory, CRM. People keep working where they are used to working — only fewer repeats arrive.

  6. 06Returns · gaps · sources

    A monthly report with the gaps

    What was closed, what came back and which questions went unanswered. The gap list is the real value: it says which page should be written next.

04 · TWO WAYS TO MEASURE

The same data, two measures, opposite incentives.

The same data, two measures, opposite incentives.
The same data, two measures, opposite incentives.Deflection rateSolved without a return
What it countsEnding without a person.Solved and not returning.
An abandonmentCounts as a success.Does not count at all.
The same case againTwo successes.One gets deducted.
The incentiveMake handover harder.Hand over early where it sticks.
CollectedInstantly, from the log.After 14 days, against a check.
What it tells youThat it got cheaper.Which answer is still missing.

05 · DUTIES AND LIMITS

Anybody who wants a person gets one, without repeating the request three times.

  • The EU AI Act requires that people are told when they are interacting with an AI system; that duty has applied since August 2026. Independently of it, the fact stands in the first sentence, because a customer who works it out later loses trust exactly once.
  • Deadlines, amounts, warranty and goodwill are decided by nobody automatically. The assistant reads, classifies and drafts; a person makes the commitment.
  • Wrong answers get reported, corrected and fixed in the source, not overwritten case by case. How many there were stands in the monthly report.

PLAIN ANSWERS

How much can a machine actually close for me?
That depends on how much of your own answers is written down anywhere. Where services, deadlines and procedures are documented, an assistant closes most of the recurring questions. Where the knowledge sits in people’s heads it closes little and prepares instead. A dependable figure exists after a real volume has been recorded; before that it would be a guess.
Does this replace people?
It takes the recurring questions away and prepares the hard ones. What remains is on average more demanding: complaints, exceptions, goodwill. Headcount therefore falls less than many offers suggest, while waiting time falls a lot. Anybody budgeting for "half the team" is budgeting wrong.
Does it work with my ticketing system?
Yes, and that is usually the better route. Triage runs before the ticket: what the assistant closes never becomes one. What it prepares arrives with category, summary and history in the system people already know. A second tool beside the first is rarely the answer.
How do I measure whether it works?
By requests solved without a return within 14 days, not by the deflection rate. Alongside that: the time to the first usable answer, and the list of questions that went unanswered. Those three stand in the monthly report, each with its method beside it.
What if the assistant says something wrong?
It gets reported, corrected and fixed at the source, so the same question does not go wrong again for the next customer. For the subjects where a wrong answer gets expensive — cancellation rights, warranty, goodwill, health — it does not answer at all but hands over.

THE NEXT STEP

One web address is enough to start.

I check which questions the existing pages can already answer today and which content is missing for the rest. You get that result even if no project comes of it.

WHAT LIES ALONGSIDE

Anybody who searched for one channel, or for the whole volume, finds it here.

← Back to the AI chat