Audit what your agentsactually said
Every call, email and chat — scored against your client's own scorecard, with each verdict backed by a verbatim quote you can click through to. Personal data is destroyed before storage, never after.
Measured, not asserted
We already read the formats you actually have.
Five client programmes, five different exports. 8 kHz mono telephony WAV from the dialler, 22 kHz MP3 from the helpdesk, stereo recordings where the channels are already split by speaker. No re-platforming, no format negotiation.
| Programme | Files | Format | Sample rate | Mean call | Median call |
|---|---|---|---|---|---|
| AEON | 12 | WAV · PCM 16 | 8 kHz mono | 4:04 | 3:59 |
| CFM | 10 | MP3 | 22 kHz mono | 7:56 | 5:23 |
| KrisShop | 10 | MP3 | 22 kHz mono | 7:35 | 5:12 |
| Minden | 11 | WAV + MP3 | 8 kHz stereo | 8:13 | 8:32 |
| RainPoint | 10 | WAV · PCM 16 | 8 kHz mono | 9:58 | 8:24 |
Profiled with ffprobe from the sample set — no audio was played, and nothing was transcribed to produce this table. Minden ships two-channel recordings, which is the easiest possible case for speaker attribution: the split is already done.
The premise
Sampling six calls a week tells you nothing about the other nine hundred.
Manual QA reviews 1–2% of conversations and calls it quality management. Verdict scores every one — and shows you the exact sentence behind each verdict, so the score survives an argument.
Platform
Everything a QA programme needs.
One normalized conversation shape across voice, email and chat — so a scorecard written once applies everywhere.
Evidence-cited scoring
Every verdict quotes the transcript verbatim. Quotes are string-matched against the source — a fabricated citation is flagged, never published.
Redaction before storage
NRIC, card, phone, address and name are destroyed in memory before the first database write. Counts are kept; values are not.
Multi-client scorecards
Each of your clients gets its own criteria, weights and compliance rules — plus a reporting view scoped to only their conversations.
Coaching that closes the loop
Recurring failures raise coaching items automatically, and effectiveness is measured as the score delta after the session.
Calibration built in
Score blind against your QA lead and track per-criterion agreement. Anything under 80% flags the guidance text as ambiguous.
No CCaaS required
SFTP drop, REST API, IMAP, Microsoft Graph or Google Workspace. If you can export it, it can be audited.
Live · runs in your browser
Watch the PII disappear.
This is the production redaction engine, not a mock. Edit the text and watch what would actually be written to the database.
Type anything. It never leaves your browser.
This is the only version that reaches the database or the model. The original is destroyed in memory — counts are kept, values are not.
Process
From recording to coaching.
Connect
Point us at an SFTP folder, an API, or a mailbox. No phone-system migration.
Redact
PII is detected and destroyed before anything is persisted or sent to a model.
Score
Every conversation is graded against the client scorecard, with quoted evidence.
Coach
Recurring failures become coaching items, and the effect is measured.
Modules
Sixteen modules. One demo.
Your whole book of business, one screen.
Score trend, per-client bars, a criterion × client heatmap that shows which rule is failing where, and a live NRIC compliance banner.
Evidence viewer
Click a criterion, the quote highlights in the transcript.
Open →NRIC Compliance Watch
140 days until PDPC enforcement.
Open →Coaching
Rules, queue, and measured effectiveness.
Open →Calibration
AI vs human agreement, per criterion.
Open →What gets audited
What QA leads say
Scores that survive a dispute.
“The first tool where an agent challenged a score and we could just point at the sentence. Disputes went from a weekly argument to a two-minute conversation.”
“We run eleven clients. Every other vendor wanted us to pick one scorecard. This one assumes we have eleven.”
“Legal signed off in one meeting because we could show them the redaction happening before storage, not in a policy document.”
Illustrative quotes for this prototype — not real customer statements.
Pricing
Pay for minutes audited. Not for seats.
Your agents are shared across clients, so a seat licence cannot be attributed to one client — which means it cannot be rebilled to them either. Minutes can. Commit to a monthly volume and the rate drops.
Starter
One client programme, one scorecard.
- 1 client workspace
- 100% of calls scored, not a sample
- Evidence-cited verdicts
- Redaction before storage
- NRIC Compliance Watch
- SFTP or API ingestion
Growth
The working tier for a multi-client desk.
- Up to 5 client workspaces
- Everything in Starter
- Coaching queue + effectiveness tracking
- Calibration against your QA lead
- Email + chat channels included
- Per-client reporting views
Scale
For a BPO running a full book of business.
- Unlimited client workspaces
- Everything in Growth
- Postgres row-level client isolation
- SSO + audit log
- Scoring API + webhooks
- Priority support, 1 business day
Enterprise
From 250,000 minutes a month at US$0.024/min. Adds self-hosted transcription on your own hardware, Singapore data residency via Bedrock ap-southeast-1, a signed DPA, and a named implementation lead.
Free pilot
500 minutes across 30 days, one client, no card. Enough to score a fortnight of a real queue and calibrate against your own QA lead.
3-minute minimum
Each conversation bills at a 3-minute minimum, then in 30-second increments. Scoring cost is fixed per conversation, so a 40-second call costs us nearly what a 3-minute one does.
Email and chat are cheaper
No transcription step, so they meter at a flat US$0.015 per conversation rather than per minute. Voice is the expensive channel.
What would it cost to audit all of it?
Pre-filled with a real programme: 67 calls a day at a 2 minute 11 second average handle time.
US$0.16 per conversation · 100% audited
- Conversations per month
- 1,742
- Actual talk minutes
- 3,832
- Billed minutes
- 5,226 (3-min minimum)
- Plan / overage
- US$199 + US$74
At 8 reviews an hour and US$2,700 fully loaded per analyst. Most QA teams do not do this — they sample 1–2% instead, and infer the rest.
The comparison you are going to run anyway
Per-agent tools look cheaper until you count the seats.
A seat licence bills every named agent — part-timers, overflow staff, the two people covering a weekend queue — whether or not they take calls. On a multi-client desk that is most of your roster. Below, each vendor's published rate normalised to a per-minute figure using a fully occupied agent: 8 hours, 22 days, 42% talk time, or 4,435 talk-minutes a month.
| Vendor | Billing basis | Published | Per minute | Multi-client |
|---|---|---|---|---|
| Enthu.ai | per agent | $59 / agent | $0.013 | Single brand |
| Zendesk QA | per agent | $25 / agent | $0.006 | Single brand |
| MaestroQA | per agent | $19–80 / agent | $0.004–0.018 | Single brand |
| AWS Contact Lens | per minute | $0.015 / min | $0.015 | Single brand |
| Verdict — Growth | per minute | $699 / 20,000 min | $0.035 | Yes |
| Verdict — Scale | per minute | $1,999 / 70,000 min | $0.029 | Yes |
We are more expensive per minute than a seat licence at full occupancy, and we are not going to pretend otherwise. What you get for the difference is the thing none of them do: N scorecards, N compliance regimes, provable isolation between clients, and a per-client invoice line you can pass through. If you run one brand's in-house contact centre at high occupancy, buy Enthu.ai — it is a good product and it is cheaper for you.
Singapore · PDPC
NRIC stops being an ID check on 1 Jan 2027.
The PDPC has confirmed organisations must stop using NRIC — full or partial — as an authentication factor. If your agents still ask for the last four digits, every one of those calls is a finding.
Verdict flags the behaviour across 100% of conversations and tracks remediation by agent and client. No other QA vendor ships this.
See the compliance module →No login. No setup. Just open it.
The full prototype is live — every module, seeded with realistic data.
Open the demo →