AI agent tooling

Customer service agent

A concept for one agent that answers chat and phone in Swedish, hands over to staff with context, and turns dictation into reviewed notes.

See the demo

When
Jul 2026
Status
Concept presentation; nothing deployed
Source
No source; concept only
Stack
  • Concept
  • Chat
  • Telephony
  • Dictation

Problem and idea

Support teams answer the same questions every day, write the same notes, and are closed when many customers need help.

The idea: one agent that uses the team’s approved knowledge across web chat, phone and staff dictation, with a clear limit where a person takes over and every note reviewed before it is saved.

How it works

  1. Customer asksWeb chat or phone, in Swedish
  2. Agent answersOnly from approved knowledge
  3. HandoverA person takes over with the background attached
  4. DictationStaff dictate; the agent drafts the case note
  5. ReviewStaff check and approve before anything is saved

Demo

The concept was presented as a 14-slide deck in seven parts. Each part in short:

  1. Customer valueCommon questions get an answer even when the support desk is closed, and many customers can be helped at once.
  2. Chat and knowledgeAnswers come only from the organisation’s approved knowledge, so the agent explains instead of guessing.
  3. Phone in SwedishA clear spoken answer, and a clear limit for what the agent will not handle.
  4. HandoverWhen a person is needed, the background follows the case so the customer does not repeat themselves.
  5. DictationStaff describe the case; the agent drafts the note; staff review and approve it.
  6. Team and controlThe organisation decides what the agent knows, what it can access and when a person takes over.
  7. The offerOne agent at roughly the cost of one support employee, with scope and price set in a quote.

The earlier video of this presentation carried the client’s branding, so it is not published here.

What I built

  • Turned the presentation into a chaptered, silent walkthrough video with Swedish captions and a readable transcript.
  • Wrote the recording script that captured each slide at fixed timings.

Stack

  • Node.js
  • Playwright
  • ffmpeg
  • WebVTT

Quality evidence

  • Chapters, captions and a full text transcript, so the walkthrough works without sound.

Current limits

  • This is a concept. The phone and case-system integrations are not built or verified.
  • Capacity and savings have not been measured.