Business Overview

Prototype · demo data Data sources: Voicebot · Chatbot · Live Agent · Phone & Meeting Summary
CHANNEL
WHEN

Conversations by topic iTopics are customised per company — each company gets its own topic list, with a management screen for editing it (to be added). Changing the list affects new conversations only; old ones keep their labels.

This month vs last month · every conversation is tagged automatically right after it ends

25 50 75 0 Product/Service — Feb: 72, Mar: 53 53 Customer Service — Feb: 40, Mar: 31 31 Campaign — Feb: 53, Mar: 34 34 Technical — Feb: 25, Mar: 17 17 Delivery — Feb: 61, Mar: 23 23 Billing — Feb: 19, Mar: 10 10 Others — Feb: 18, Mar: 9 9 Product/Svc Cust. Service Campaign Technical Delivery Billing Others
Mar 2026Feb 2026

How customers felt iThe feeling shown is the customer's FINAL sentiment when their journey ends. If one customer chats, then calls, then reaches an agent in the same day, that is one journey — scored by how it ended, not how it started.

All conversations vs negative only — spot where unhappiness concentrates

Product/Service: 9 Campaign: 12 Customer Service: 12 Billing: 4 Delivery: 8 Technical: 9 Others: 18 72 all Product/Service: 3 Campaign: 3 Customer Service: 2 Billing: 2 Delivery: 4 Technical: 5 Others: 8 27 negative
Product/ServiceCampaign Customer ServiceBilling DeliveryTechnical Others

Matching conversations 23 records

One line per conversation — click a row to open the full record

IDDateFeelingTopicWhat happened
3013-Feb 15:34NegativeDeliveryCustomer complained that Eggrolls have no stock in the TST outlet for a long time. Crispy Eggrolls Gift Box
3024-Feb 9:29MixedCampaignCustomer could not use a credit card promotion; resolved after a human agent stepped in.
3034-Feb 11:05NegativeProduct/ServiceCustomer received stale Almond Biscuits from Mong Kok outlet; asked for a refund. Almond Biscuits
3045-Feb 14:18PositiveProduct/ServiceCustomer praised the new Pineapple Shortcake packaging. Pineapple Shortcake
Uses tokens · usually takes a few minutes · result is cached ?For HKT to confirm: we assume a flat token cost per analysis job (for example, 10 tokens), charged to the company pool. If the price should instead depend on the number of records analysed, please define that rule.
SHOW
FEELING
SORT

Record #301 — transcript ?For HKT to confirm: speaker separation (who said what) depends on diarization support in the transcription pipeline, which has not been confirmed yet. If it is unavailable, the transcript shows plain text without the BOT/CALLER labels seen here.

Voicebot call · General Hotline · 3-Feb-26 15:34 · 4 m 12 s · glossary corrections shown inline

BOT
00:02Thank you for calling Kee Wah. How can I help you today?
CALLER
00:09I went to the tsim sha choy Tsim Sha Tsui shop three times this month and the crispy eco Crispy Eggrolls gift box is never in stock.
BOT
00:24I am sorry about that. Let me check the stock at the Tsim Sha Tsui branch for you.
CALLER
00:41It has been like this since Chinese New Year. I want to order for a company gift, at least twenty boxes.
BOT
00:55For corporate orders of twenty boxes or more, I can arrange delivery from our warehouse instead of the branch. Would you like me to do that?

Note: playback of the original audio recording is not supported in the portal.

AI summary AI iGenerated automatically the moment each conversation ends, as part of the standard pipeline (write down → summarise → label), using your company's glossary for names the machine usually mishears. Reading summaries costs no tokens.

What the customer wanted

  • Crispy Eggrolls Gift Box — repeatedly out of stock at TST branch since CNY.
  • Wants a corporate order of 20+ boxes.

What happened

  • Bot offered warehouse delivery for the corporate order.
  • Customer accepted; order handed to sales follow-up.

Tags

Negative Delivery Crispy Eggrolls Gift Box TST Branch

Deep case review AI iPlanned as the next goal after the core pages. Works like list analysis: uses tokens (flat cost per report, to be confirmed), usually takes a few minutes, and the result is cached — re-opening it is instant and free.

Generate a full report for this one case — background, actions taken, business recommendation

Uses tokens · usually a few minutes · cached

AI Analysis iFailed runs are retried automatically. If a run still fails after retries, its tokens are refunded to the company pool. Finished results are cached — re-opening them is instant and costs nothing.

Every AI analysis you asked for, with its token cost — runs usually finish in a few minutes (large record sets or busy AI servers take longer)

IDTypeFiltersRecordsDateTokensStatus
AI-2026-0512✦ SentimentNegative Mixed42328-May 14:3212,450Done
AI-2026-0511◈ AttributeProduct Quality1,28728-May 09:1028,760Done
AI-2026-0510✦ SentimentPositive Negative85627-May 18:0519,850Analyzing…
AI-2026-0507◈ AttributeService Tone98426-May 10:1522,150 refundedFailed

Result — AI-2026-0512 Done

From 423 filtered records · 12,450 tokens · filters: Negative, Mixed · exportable as a report

Executive summary AI

Delivery complaints concentrate on two products and one branch: eggroll and mooncake gift boxes at TST, driven by stock-outs during festive peaks. Campaign friction is payment-related, not offer-related. Sentiment recovers quickly when a human agent intervenes — but only 12% of negative journeys reach one.

Key findings AI

  • TST branch = 61% of stock-out complaints.
  • Credit-card promotion failures resolved only after human escalation.
  • Corporate orders drive the angriest calls — and the biggest revenue.

Recommended actions AI

  • Priority restock of gift boxes at TST before festivals.
  • Fix promotion validation in the chatbot payment flow.
  • Route corporate-order calls to an agent immediately.
Calls received today
412
via all SIP numbers
Transferred to agent
37
9.0% of received
System success rate
97.6%
setup-failed 6 · transfer-failed 4
Avg call length
3 m 48 s
this month

Calls per hour — today ?This page merges the unconfirmed "Quantitative Insights Portal" into the User Portal (raw counters from the phone system and apps). To confirm with HKT: is a separate portal required, and how long must raw call logs be kept?

Raw counts from the phone system — no AI, no tokens

40 80 09:00 — 22 calls09 10:00 — 41 calls10 11:00 — 58 calls11 12:00 — 34 calls12 13:00 — 29 calls13 14:00 — 47 calls14 15:00 — 66 calls15 16:00 — 72 calls7216 17:00 — 43 calls17

Raw call log

Straight from the phone system (FreeSwitch) — one row per call · export CSV

Call IDStartedNumberCallerDurationHandled byResult
C-8811216:42:103656 4221+852 9•••12343:12VoicebotCompleted
C-8811116:38:553656 4221+852 6•••88026:47Voicebot → AgentTransferred
C-8811016:31:023656 4223+852 9•••55110:00—Setup failed
C-8810916:24:473656 4221+852 5•••99302:05VoicebotCompleted
? For HKT to confirm: caller numbers are shown masked here as our default. Please confirm your privacy rule — masked, fully visible, or hidden — and whether it differs by user role.

This portal answers one question: "How well is our AI customer service doing, and what should we fix?" Every phone call, chat and meeting handled by the AI is automatically written down, summarised and labelled the moment it ends. This portal is where you read the results. Looking is free; asking the AI to produce a new analysis costs tokens (your company's prepaid credits).

How it works — two pictures

1. From a call to your dashboard (automatic, no tokens)

Everything below happens by itself, right after each conversation ends.

sequenceDiagram
    participant C as Customer
    participant B as AI apps (bot / notetaker)
    participant E as AI engine
    participant P as User Portal
    C->>B: talks (call, chat, meeting)
    B->>E: sends the recording when finished
    E->>E: write down, summarise, add labels
    E->>P: store the finished record
    Note over P: You browse records and charts - free

2. Asking the AI for a deep analysis (uses tokens, takes minutes)

You choose a set of records; the AI studies all of them and writes a report. The finished report is kept, so opening it again later is instant and free.

sequenceDiagram
    participant U as You
    participant P as User Portal
    participant E as AI engine
    U->>P: filter records, press "Ask AI to analyse"
    P->>E: run the job (usually a few minutes)
    E-->>P: report ready, tokens deducted
    P-->>U: result appears - cached for next time
    Note over E: fails? auto-retry, then refund tokens

What you can do here

1Check the health of your customer servicea few minutes, weekly
  1. Open Business Overview. The bars show how many conversations happened on each topic, compared with last month.
  2. A bar that jumped (say, Delivery) means something changed. The donuts show whether people were happy or unhappy about it.
  3. Scroll down: every matching conversation is listed in one line each, with its feeling and topic.

You see trouble the day it starts trending, not in next month's report.

2Read what actually happened in one conversation2 minutes
  1. Click any row to open it in Records.
  2. Left side: the full conversation, word by word, with who said what and when. Words your customers use that machines usually get wrong (product names, branch names) are auto-corrected using your company's glossary — corrections are shown.
  3. Right side: the AI's short summary — what the customer wanted, what happened, and the labels attached.
  4. Need more depth on an important case? Press Generate case review for a full report on that single case (uses tokens).

No more listening to recordings to find out why a customer was angry.

3Ask the AI to study hundreds of conversations at once5 minutes to ask, hours to cook
  1. Filter the records you care about (e.g. all negative Delivery conversations this month).
  2. Press Ask AI to analyse. Confirm — this uses tokens. A progress state shows while it runs; usually a few minutes (very large record sets, or busy AI servers, take longer).
  3. The report appears on AI Analysis: what's going wrong, where, and what to do about it. It is cached — re-opening it later is instant and free.
  4. Every report stays in the list with its token cost, so spending is always visible. Failed runs retry automatically and refund their tokens if they still fail.

One report replaces a week of manually reading transcripts.

4Watch the raw numberslive
  1. Open Raw Data for plain counters: calls received, transferred, failed, busiest hours, and the raw call log.
  2. This is the un-processed truth from the phone system — useful for staffing decisions and for checking the AI numbers against reality.

Trust, but verify: the AI views and the raw counts sit one click apart.

What we need from HKT

  1. Confirm the analysis pricing assumption. We assume a flat token cost per AI analysis job (for example, 10 tokens), charged to the company pool. If pricing should depend on how many records are analysed, please define that rule.
  2. Confirm speaker separation. The transcript view labels who said what; this depends on diarization support in the transcription pipeline, which is not yet confirmed on your side.
  3. Confirm the merged Raw Data page. We propose folding the quantitative raw-data view into this portal instead of a separate portal. Also: how long must raw call logs be kept, and how should caller numbers be masked, per your privacy rules?
  4. Provide each company's topic list. Topics are per-company; we need the initial list per client (we supply a default 7-topic starter set).