Data from every call
AI listens for you, scores the call and writes straight into your CRM.
Call to Text: connects to your phone system, 100% of calls become text.
Keep your phone system · No software on staff devices · Results generated right after the call ends
You measure every ad dollar. Yet at the exact moment customers decide, you hear nothing
Calls are where deals are actually won or lost — and the only stage of the funnel nobody measures.
Managers listen to a handful of calls a week and trust self-reported summaries for the rest
Transcribe 100% of calls into searchable text — read a 30-second summary instead of listening for 8 minutes, and search any keyword: "discount", "competitor", "let me think about it"
QA listens to samples and scores by hand in Excel — every reviewer uses a different yardstick
Score 100% of calls against your own scorecard — one consistent yardstick, every criterion backed by a verbatim quote
Reviewing recordings is mathematically impossible: 20 agents × 30 calls/day × 5 minutes ≈ 1,100 hours per month (illustrative estimate)
AI processes everything in parallel. People only open the flagged ones: low scores, upset customers, risky promises, big deals
At the end of the day, staff log a single line — "customer interested, will call back". Needs, budget and objections never make it into the system
AI extracts needs, budget, location, objections and appointments, then writes them straight into CRM fields. Staff just confirm
Marketing picks ad messaging by guesswork, while the answers sit inside thousands of calls nobody reads
Cluster topics and objections by frequency — know exactly what customers ask, complain about and compare you against
Staff over-promise to close deals; the company only finds out when a customer complains or demands a refund
You define the prohibited-phrase list. Violations are flagged instantly with a Zalo/Email alert linking to the exact moment
One 8-minute 14-second call, after AI processing
The same call: what you have today on the left, what Speech Insight AI returns on the right.
BEFORE
What you have: an 8:14 recording
AFTER AI PROCESSING
What you get: a 30-second read
AI analysis
Customer wants a 2-bedroom unit, budget around 2 billion VND, needs a loan. Main objection: priced higher than a competitor. Agreed to a weekend site visit.
- Greeting & needs confirmation
- Budget discovery
- Objection handling
- Next-step close
Compliance alert — 05:33
Risky promise detected: “the title deed will definitely be issued this quarter”
Replay from 05:33CRM record auto-filled
crm.clickstar.vn
Deal stage: Consulting → Meeting scheduled
✓ Reminder task created 1 day ahead
AFTER AI PROCESSING
What you get: a 30-second read
AI analysis
Customer wants a 2-bedroom unit, budget around 2 billion VND, needs a loan. Main objection: priced higher than a competitor. Agreed to a weekend site visit.
- Greeting & needs confirmation
- Budget discovery
- Objection handling
- Next-step close
Compliance alert — 05:33
Risky promise detected: “the title deed will definitely be issued this quarter”
Replay from 05:33CRM record auto-filled
crm.clickstar.vn
Deal stage: Consulting → Meeting scheduled
✓ Reminder task created 1 day ahead
The agent typed nothing. Every result above was produced by AI.
Every call passes through 9 layers of analysis — automatically, with zero clicks
Vietnamese transcription built for sales calls
Handles regional accents, fast speech, overlapping talk and office noise. Loads your own dictionary: product names, projects, courses.
- Per-sentence timestamps
- Industry-specific dictionary
- Phone number & amount detection
- Full-text search
Speaker separation
Clearly separates agent from customer, unlocking metrics the human ear cannot measure.
- Talk-time ratio
- Interruption count
- Unusual silence gaps
- Speaking pace
Call summary and action items
A 30-second read: who the customer is, what they need, what they objected to, what was promised and what happens next.
- 5-line summary
- Promises made by the agent
- Next step with deadline
- Link to original transcript
Scoring against your own playbook
No imported one-size-fits-all rubric — we use the exact checklist your team already trains on.
- Weighted criteria
- Verbatim evidence quotes
- Scores 100% of calls
- Compare by agent and team
Sentiment and escalation detection
Measures sentiment segment by segment instead of stamping one label on the whole call.
- Sentiment over time
- Churn-risk alerts
- Complaint signals
- Estimated satisfaction score
Data extraction and CRM auto-fill
The heart of the service: turns a conversation into structured fields sitting in the right place in your CRM.
- Maps to your existing CRM fields
- Updates deal stage
- Creates tasks and reminders
- One-tap confirmation by staff
Topic and objection detection
Answers the most expensive question: which objections actually kill deals and which are just noise.
- Topic clustering
- Objection ranking
- Cross-referenced with close rate
- Weekly trends
Prohibited-phrase and compliance alerts
You define the prohibited phrases and risky-promise patterns that matter in your industry.
- Custom prohibited phrases
- Over-promise detection
- Zalo, Email, Telegram alerts
- Compliance log
Trend reports and team ranking
Feeds directly into Clickstar Data Dashboard so call quality sits next to ad spend and revenue.
- Ranking by scorecard
- Period-over-period comparison
- Filter by campaign and source
- Automated report delivery
The brain reading your calls is Claude by Anthropic
We are upfront about what we run on. Below are the 4 technical reasons for that choice — and a clear line between what the AI does and what people decide.
Analysis model
Claude — Anthropic
We match the model tier to each company’s call volume and complexity, rather than forcing one configuration on everyone.
Understands Vietnamese as people actually speak it
Not textbook Vietnamese: customers mix in English, abbreviate, trail off, and speak in 3 regional accents. The same objection — "too expensive" — comes in 10 phrasings; the system still groups them under one label.
Reads the whole call in one pass
A 45-minute call is not chopped into fragments and stitched back together. That is how the system knows a sentence at minute 38 answers a question asked at minute 4 — exactly what chunking always loses.
Returns structured data, not prose
Output is constrained to a predefined data shape — not a paragraph your system then has to guess its way through. What arrives goes straight into the CRM.
Scores against your own playbook
Your scorecard, product dictionary and internal rules are loaded in — not a boxed checklist sold to everyone. Change the playbook and the scoring changes; the system does not need rebuilding.
What the AI does, what people decide
This boundary is set explicitly up front so your team knows exactly what they are working with.
- Listens to and transcribes 100% of calls, missing none
- Scores against the criteria you provide, applied identically to every agent
- Extracts needs, budget, objections and the agreed next step
- Flags the calls a human needs to review
- What goes into the scorecard and how it is weighted
- What score counts as passing and which calls get re-reviewed
- Performance reviews, rewards, penalties and coaching
- What is stored, for how long, and who may see it
The AI does not replace managers in judging people. Every score can be opened up: which part of the call it was based on, and you can replay that exact moment.
Where your data goes
- Your call data is not used to train AI models.
- Sensitive details in transcripts are masked automatically before analysis.
- Full audit log: which call was processed, when, and who opened the result.
The same call — here is what the AI sees
Pick the industry closest to yours, then follow 3 steps: read the conversation, see the signals the AI picks up, and get the data back.
Hello, you were asking about the vocational study programme in Germany, is that right?
Yes, I saw it on Facebook. How much does the whole thing cost?
It depends on the pathway — let me send you the detailed sheet over Zalo.
I just want a number. Roughly how much?
I would have to check with the advisory team on that…
Never mind, I will look into it myself. [hangs up]
Price question dodged — customer asked directly, agent redirected to another channel
Customer had to ask twice: an impatience signal
“I have to check with the team” — agent has no price range to answer on the spot
Customer ended the call; no next step was agreed
Only 49 seconds; the agent spoke 71% of the time yet delivered none of the information asked for
{
"call_id": "cl_3b81e0",
"duration_sec": 49,
"quality": {
"score": 41,
"criteria": {
"chao_hoi": true,
"khai_thac_nhu_cau": false,
"tra_loi_cau_hoi_gia": false,
"chot_buoc_tiep_theo": false
}
},
"sentiment": { "overall": "negative", "drop_at_sec": 34 },
"talk_ratio": { "agent": 0.71, "customer": 0.29 },
"extracted": {
"nhu_cau": "Vocational study in Germany",
"objection": "khong_duoc_bao_gia",
"next_step": null
},
"flags": [
"khach_chu_dong_cup_may",
"cau_hoi_gia_khong_duoc_tra_loi"
],
"crm": { "stage": "can_goi_lai", "task_created": true }
}Quality score
41/100 — failed 3 of 4 criteria
Task created automatically
Call back within 24 hours with a concrete fee range, assigned to that same agent
Logged into the weekly report
Error group “price question left unanswered” — if it keeps growing, the problem is the script, not one person
Not a black box — structured data your engineers can use right away
Every analysed call returns a complete JSON record, delivered by webhook to your system or pulled in batches via API.
{
"call_id": "cl_9f2a17",
"duration_sec": 494,
"agent": { "id": "sale_08", "name": "Nguyễn Thu Hà" },
"transcript_url": "https://.../cl_9f2a17.txt",
"summary": "Asked about a 2BR unit, budget ~2B VND,
needs financing. Felt the price was above
a competitor's. Site visit booked for
Saturday.",
"quality": {
"score": 82,
"criteria": {
"chao_hoi": true,
"khai_thac_nhu_cau": true,
"xu_ly_tu_choi": true,
"chot_buoc_tiep_theo": true
}
},
"sentiment": { "overall": "positive", "risk": false },
"talk_ratio": { "agent": 0.62, "customer": 0.38 },
"extracted": {
"nhu_cau": "2-bedroom apartment",
"ngan_sach": 2000000000,
"khu_vuc": "Thủ Đức",
"can_vay": true,
"objection": "gia_cao_hon_doi_thu",
"next_step": { "type": "site_visit", "date": "2026-07-04" }
},
"compliance": [
{
"type": "over_promise",
"at_sec": 333,
"quote": "the title deed will definitely be issued this quarter",
"severity": "high"
}
],
"crm": { "record_id": "kh_41022", "stage": "hen_gap", "task_created": true }
}Field names and CRM mapping are configured to match your own system.
From single calls to the whole team picture
Once thousands of calls exist as data, AI answers questions that used to be pure guesswork.
Why customers don’t close — ranked by frequency
Aggregated automatically across every call in the period
Every objection is clickable — see the underlying calls and replay the exact moment.
Ranked reasons for lost deals
AI normalises thousands of phrasings of the same objection, then cross-references close rates to reveal which ones actually kill deals.
Topics customers keep raising
Clusters recurring questions week by week — feeding straight into ad copy, sales scripts and FAQ pages.
Team ranking and coaching
Compares each agent on the same scorecard and pinpoints which criterion drags the score down, with the exact audio excerpt.
Early risk alerts
Detects abnormal spikes in negative sentiment or repeated over-promising — before they become complaints.
From hang-up to CRM record — nobody lifts a finger
Connect the recording source
One-time, 1–3 business daysPull recordings plus call metadata via API, webhook, SFTP or cloud folder sync. Your live calling system is untouched.
Tune the scorecard and dictionary
During the pilotLoad your call script, QA checklist, prohibited-phrase list and company-specific vocabulary.
Normalise and transcribe
Right after the call endsChannel separation, noise reduction, Vietnamese recognition, speaker separation and per-sentence timestamps.
AI analyses the content
Immediately after step 3Summarise, score, measure sentiment, tag topics and objections, extract fields, scan for prohibited phrases.
Push to CRM and alert on exceptions
Near-instantUpdates records, deals, timeline and tasks. Flagged calls trigger an alert linking to the exact problematic moment.
Report and optimise
Daily / weeklyTrend dashboards, team ranking and topic insights delivered to leadership, marketing and training.
All 6 steps run automatically. The only human job is reviewing what AI flagged.
Exact processing times are confirmed with real measurements on your infrastructure during the pilot.
Your call data is not used to train AI models
Encrypted in transit and at rest
Role-based access: staff see only their own calls, managers see their team
Automatic masking of sensitive data in transcripts (card numbers, ID numbers, OTP codes)
Configurable retention with a documented deletion process
Complete access log: who opened which call, and when
Recordings are your customers’ personal data. We treat that as the default, not an option
Call recording and analysis in Vietnam falls under Decree 13/2023/ND-CP on personal data protection. We hand over recording notice and consent templates so you deploy compliantly from day one.
The AI model we run on is stated openly above. The remaining deployment parameters — where data is stored, private/on-premise deployment options and retention periods — are confirmed in writing during the technical session, before you send us a single recording.
A transcription tool hands you a text file. We deliver the result where you actually make decisions.
Recording sources in
We read the recordings you already have.
- IP / Cloud contact centers with API or webhook
- Self-hosted open-source PBX: Asterisk, FreePBX, Issabel, 3CX
- Online meeting recordings: Google Meet, Zoom, Microsoft Teams
- Call-recording apps on staff mobile devices
- Folder sync, SFTP, Google Drive, OneDrive, S3
If your phone system can export recordings with call metadata, we can connect to it.
Where results land
The advantage of the Clickstar ecosystem.
Third-party CRMs connect via API or webhook. Export to Google Sheets, CSV or an open API for your data team.
You do NOT need to replace your phone system.
Keep your hotline numbers and the way your team already calls. Speech Insight AI sits in the layer after the call, reading only the recordings you already have. Your staff install nothing.
3 plans by minute quota
You pay for the recorded minutes actually processed — calling less does not cost the same as calling a lot.
Basic
Contact for pricing · roughly 2,000 calls a month
- Transcribe and summarise 100% of calls within the quota
- Quality scoring against one scorecard of yours
- Extract needs, budget and objections
- Connect one recording source, keep your existing phone system
- Weekly report with ranked reasons for lost deals
- Recording notice templates per Decree 13/2023
Advanced
Contact for pricing · roughly 4,000 calls a month
- Everything in Basic, with double the quota
- Multiple departments and locations on one system
- Auto-fill results into the CRM after every call
- Compliance and negative-sentiment alerts sent to managers immediately
- Separate scorecards per product line
- Agent ranking with the audio excerpts for coaching
- Role-based access and a complete audit log
Premium
Minute quota designed around your actual needs
- Everything in Advanced, with a custom quota
- Dedicated API and webhooks into your existing internal systems
- Industry dictionary and scorecards tuned to your business
- Bespoke agreement on data location and retention period
- SLA commitment and a dedicated engineer
- Regular reviews together with your operations team
The 11 questions sales leaders ask most — answered directly
Connecting the recording source takes 1 to 3 business days, depending on your existing call system and the work of writing the connector. Next comes a pilot period to tune your scorecard and custom dictionary, then full operation: 100% of calls are turned into text, scored against your own criteria and flagged when they break the rules.
Yes. The foundation of the service is Call to Text: every call is automatically converted into Vietnamese text (also known as voice-to-text, speech-to-text or call transcription), with agent and customer clearly separated. Unlike plain transcription tools, AI then summarises the call, scores it against your own playbook, extracts needs, budget and objections, and writes the results straight into your CRM.
Pricing is based on the volume of recorded minutes processed each month, plus a one-time setup fee (call system integration, scorecard and custom dictionary tuning). You only pay for what your team actually calls. After the trial analysis, we quote based on your real volume.
Data is encrypted in transit and at rest, access is role-based (staff see only their own calls, managers see their team), retention is configurable and every access is logged. Sensitive details in transcripts are automatically masked. Call recording and analysis in Vietnam falls under Decree 13/2023/ND-CP on personal data protection — we hand over recording notice and consent templates so you deploy compliantly from day one.
The analysis runs on Claude by Anthropic. We chose it for 3 technical reasons: it understands everyday spoken Vietnamese (mixed with English, abbreviated, across regional accents) rather than only textbook Vietnamese; it reads a long call in a single pass, so context between the start and end of the conversation is not lost; and it returns results in a predefined data shape that writes straight into your CRM instead of a block of prose. We match the model tier to each company’s volume and complexity rather than applying one configuration to everyone.
No. Your call data is not used to train AI models. Sensitive details in transcripts are masked automatically before analysis, and the whole process is fully logged: which call was processed, when, and who opened the result. The specific terms on storage location and retention period are written into the contract before you hand over any recording.
Accuracy depends most on your call system’s recording quality (separate vs mixed channels, codec, background noise), not just the AI model. The hardest part is always proper nouns and industry jargon — we handle that by loading your own dictionary (product names, projects, courses). The honest answer: we run a trial on your real calls and give you measured results before you sign anything.
Any system that lets us pull recordings plus call metadata (extension, agent, duration, customer ID) via API, webhook, SFTP or cloud folder sync. That includes IP/Cloud contact centers, self-hosted open-source PBX (Asterisk, FreePBX, Issabel, 3CX), online meeting recordings (Google Meet, Zoom, Teams) and even mobile call-recording apps.
No. You keep your existing phone system, hotline numbers and the way your sales team already calls. Speech Insight AI sits in the layer after the call — it only reads recordings you already have. Your staff install nothing.
A pilot for one department typically takes 1–2 weeks. You only need 3 things: a sample set of recordings, the call script or QA checklist you already use, and the list of CRM fields you want filled automatically. Clickstar handles the technical integration.
Yes — if you roll it out secretly. We recommend the opposite: announce it to the team upfront, publish the scorecard before you start scoring, and position it as coaching rather than fault-finding. In practice new hires benefit most, because they get concrete per-call feedback instead of vague advice.
We only publish numbers we have actually measured on real data: transcription accuracy, pilot results, calls processed. Instead of handing you a number we invented, we run a trial on your own calls and give you measured results — before you sign anything.
Send us 50 real calls. Get back an analysis of your own team.
No phone system change, no upfront commitment. We transcribe, score and extract the data, then review the results together with you. After that session, you decide whether to proceed.
We sign an NDA before you send any recording. Trial data is deleted after the review if you decide not to proceed.