Mikkel Kjaer
Dansk
Customer service quality · tools built with AI

I find the manual steps and automate them away with AI.

I'm Mikkel, and I work on quality in customer service at CNX (Concentrix). Since August 2026, alongside my full-time job, I have built three internal tools with AI agents. The customer service agents and management use them every day. Here you can read about me and see how the tools work.

Portrait of Mikkel Kjaer
Mikkel Kjaer
QE (quality) in customer service at CNX. Previously SME, customer service agent, and game presenter for Evolution in Malta.
1 hr 30 min → 5–10 minto build the tables for the weekly business review (WBR)
3internal tools that agents and management use every day
1,200+automated tests that run whenever the code changes
10 weeksfrom the first version in August to today
About me

From hosting on camera in Malta to quality in customer service

I work in customer service at CNX (Concentrix) on a Danish client account. I started as an agent in 2024, became an SME (subject-matter expert), and today I work in quality (QE). Before that, I was a game presenter for Evolution in Malta.

Everything I know about IT and AI is self-taught, mostly by building things. Today I build with AI agents. They write the code. I make sure they solve the right problem and that my colleagues can use the result.

Lives in
Store Merløse, Denmark
Languages
Danish (native) and English (fluent)
Free time
I play PC games and develop games with AI. I have also built other personal projects with AI.
  1. 2026–present

    QE (quality)

    CNX · customer service

    Quality assessments, calibrations, and follow-up on customer satisfaction scores (tCSAT) and dissatisfied ratings (DSAT) for a team of about 50 agents.

  2. 2025–2026

    SME

    CNX · customer service

    Sounding board for the agents, helping with chats, emails, and escalations.

  3. 2024–2025

    Customer service agent

    CNX · customer service

    Chat and email in Zendesk. I was one of the team's top performers and helped train new colleagues on the product and tools.

  4. Earlier

    Game presenter

    Evolution · Malta

    Live host on camera for players from all over the world, about 48 hours a week, with English as the working language.

How I work

I'm not a trained developer, and I don't write the code myself. The AI agents do that, mainly Claude Code. My work is everything around it: understanding the problem, steering the AI, and making sure the result works for the people who will use it.

  1. Spot the problem

    All three tools started in my own workday: templates that had to be fixed by hand, tables for every weekly meeting, and numbers the team couldn't see.

  2. Let the AI ask

    Before the follow-up plans were built, the AI got a screenshot of the team's table. It came back with 56 questions, and my answers became the spec.

  3. Build in small rounds

    Each round ends with a change that I review and approve. The new version of the templates tool was built in two rounds in about a day.

  4. Insist on tests and review

    No change goes in without automated tests and a review round where the AI attacks its own work.

  5. Test with colleagues

    An AI-proposed redesign of the WBR tables looked good but was harder to read. It was dropped the same day, and we went back to the layout the team knew.

  6. Run and improve

    The leaderboard started as an open web page. When it became part of everyday work, I moved it behind a login so names and numbers aren't visible to outsiders.

The tools

Three cases from everyday work

The tools are internal, so all names, templates, and numbers in the demos are made up. The mechanics are the same as in the real tools.

The quality platform

The quality team works from two export files: the customers' satisfaction scores and the team's assessments of responses from dissatisfied customers. The platform merges them and produces what used to be built by hand for every weekly meeting.

Before

  • Two export files that had to be matched and totaled by hand
  • WBR tables built by hand for every weekly meeting
  • Status per agent typed into a table of follow-up plans

After

  • Both files are imported with validation and data cleaning.
  • The WBR tables are built from the data and copied as an image straight into slides.
  • Calibrations happen online, with automatic scoring against an answer key.
  • Follow-up plans per agent are calculated live.
5–10 minfor the WBR tables, down from 1 hr 30 min
~50agents in the data
~390automated tests
WBR tables · demo
On target (9.00+) 8.50–8.99 8.00–8.49 Below 8.00
Before1 hr 30 min
Now5–10 min
Made-up agents and numbers. In the real platform, the table is copied as an image with one click.
InTwo export files
→
ImportValidation and data cleaning
→
PlatformWBR, calibration, follow-up plans
→
OutImage for slides

What I learnedTest with the real file. The made-up test file worked, but the real export spelled out the months and used a different file format. The import was adjusted and got new tests.

Templates that insert themselves

Agents answer customers on chat and email with set templates, 276 text variants in all. The first version was a single HTML file that got sent around every time a template was corrected. Now the tool has been rebuilt as a web app with a login, a shared database, and a Chrome extension for Zendesk. The new version is built and tested.

Before

  • Find the template, copy it, switch windows, and paste
  • Fix the name and fields like XXXX yourself before sending
  • No way to see which templates were being used

After

  • Type something like /intro in Zendesk and pick a template. The text is inserted with your name filled in.
  • If it still says XXXX or [AMOUNT], the fields have to be filled in before the text can be used. That way a “Hi XXXX” never reaches the customer.
  • In the new version, a template is corrected in one place instead of in a file that gets sent around.
276text variants
~565automated tests
2 roundsin about a day
Reply to customer · demo
Type / to pull up a template
    Try:
    Made-up templates from a generic customer service desk. The real ones are internal.

    What I learnedNot every "XXXX" is a placeholder. The copy check had to tell real fields apart from words like "XXL" and links in square brackets. Otherwise it would stop agents for no reason.

    Bonus system on the big screen

    The team's bonus system runs on TV screens in the office. Agents can see who is performing best, compete against each other, and see what bonus they can get by hitting the right KPIs. Each agent has their own rocket that rises with customer satisfaction. The first version went into use the same day it was built.

    Before

    • The scores sat in an export with duplicates
    • Emails had to be matched to agents and averages worked out
    • Bonuses and streaks were tracked by hand

    After

    • Drop in today's export. Duplicates are removed, emails are matched to agents, and an import report shows what happened.
    • Bonus tiers and streaks are calculated from the same data, so everyone can see what it takes to reach the next tier.
    • The screens sit behind a login.
    ~6,700surveys in the data set
    Day 1in use the same day it was built
    ~320automated tests
    TV view · demo
    Made-up agents (A–H) and scores on a scale from 0 to 10. The target is 9.00 (dashed line), and the track starts at 8.00. The real bonus amounts are left out.
    InToday's survey export
    →
    ImportDuplicates, email to agent, report
    →
    CloudflareLogin and shared data
    →
    OutThe office TV screens

    What I learnedThe AI can misread data. In one review round it assumed the export's last day was incomplete and changed the rule. I checked the real file, and the change was rolled back. Otherwise yesterday's scores would have been counted twice.

    Timeline

    Ten weeks, three tools

    Code changes per week from August to October 2026, counted from each tool's code history. The bonus system's automatic data updates aren't included.

    Bonus system (32)Quality platform (34)Templates (16)

    Swipe sideways to see all weeks →

    1. Aug 4Bonus system built and put to use the same day
    2. Aug 29First version of the quality platform
    3. Sep 1Bonus system moved behind a login on Cloudflare
    4. Sep 15Follow-up plans: 56 questions, then a spec, then code
    5. Oct 7WBR tables that copy straight into slides
    6. Oct 10–11Templates tool rebuilt with a Chrome extension
    Tools and skills

    What the tools taught me

    Each point has an example from the cases above. The tags show which ones.

    What I use today

    • Claude and Claude Code
    • ChatGPT and Codex
    • Figma
    • GitHub and GitHub Actions
    • Cloudflare Workers and D1
    • Zendesk
    • Excel and Google Sheets