Case Study · Telebyte Reference Deployment
From 2% manual sampling to 100% call coverage — Consumer Duty evidence for an 8-agent UK insurance brokerage
- Client
- FCA-regulated UK insurance brokerage (name withheld by request)
- Sector
- Direct-to-consumer protection and health insurance intermediary
- Scale
- 8 dialler agents · 540,000+ CRM contacts · ~9M historical call recordings
- Stack
- ViciDial · Asterisk · GoHighLevel · Next.js · Whisper · RoBERTa · MariaDB
- Engagement
- Full stack — dialler, CRM, customer-facing site, transcription pipeline
- Telebyte role
- Build, host, and run, every working day
- Call coverage, post-deployment
- 2% → 100%
Call coverage, post-deployment
- Historical recordings indexed
- 9M+
Historical recordings indexed
- Complaint-call retrieval (was: days)
- < 5 min
Complaint-call retrieval (was: days)
- Encrypted UK retention, FCA-compliant
- 7 yrs
Encrypted UK retention, FCA-compliant
The problem
By late 2023, the brokerage’s compliance function had reached the same conclusion every UK intermediary was reaching: Consumer Duty had quietly made manual call sampling indefensible. A team reviewing 1–2% of calls each week could no longer evidence — to the board, the FCA, or themselves — that the other 98% were producing good customer outcomes.
The brokerage had three further problems sitting underneath that one:
- Twelve years of historical call recordings sat on a NAS — searchable only by date and agent, not by anything a complaint investigator actually needed.
- Vulnerability identification depended entirely on whether the agent had remembered to flag it — there was no second-pass check, and no audit trail on the 99% of calls that weren't flagged.
- Board-level outcome reporting was constructed from CRM notes and dispositions — narrative, not evidence.
“We knew our agents were doing the right thing on the calls we listened to. We just couldn’t tell the FCA — or honestly, ourselves — what was happening on the rest of them. That gap got loud after the first Consumer Duty supervisory letter landed.”
The approach
Telebyte already had the engineering grounding — twenty-plus years across UK dialler platforms and a decade running IT for the brokerage’s parent group. The build decision was to assemble a transcription and analytics stack from open-source components Telebyte could host, audit, and modify in the UK — rather than wire the brokerage’s audio into a US enterprise vendor’s API.
What got built
- OpenAI Whisper transcription pipeline, on a UK VPS with GPU-accelerated backfill workers for the historical archive.
- RoBERTa-based sentiment scoring on every call, with a custom vulnerability indicator pass tuned on the brokerage's own hand-labelled snippets.
- Full-text search over the entire transcript corpus in MariaDB — sub-second queries over 9M+ recordings.
- Daily QA digest emails to compliance and line management — flagged calls, low-sentiment outliers, missed disclosures.
- Encrypted UK object storage with a 7-year lifecycle policy matching FCA recordkeeping.
- Direct integration into the existing ViciDial dialler and GoHighLevel CRM, so a contact record links to its transcribed call history without manual cross-referencing.
Why this stack, and not a vendor API?
Two reasons. First, the brokerage’s call data is regulated personal information — keeping it in-country, on infrastructure Telebyte controls, removed an entire class of GDPR and FCA outsourcing concerns. Second, the per-call economics of vendor APIs at 9M+ historical recordings would have made the backfill alone cost six figures. Self-hosting made it tractable.
The outcome
Within four months of the search and sentiment layer going live, the brokerage’s Consumer Duty position changed from defensible-with-effort to materially evidenced. Every call — outbound sales, inbound service, retention, complaints — is transcribed, scored, and searchable. Every flagged vulnerability indicator has an audit trail. Every board report draws on the underlying transcript corpus, not constructed CRM narrative.
What changed, measurably
- Call coverage moved from ~2% manual sample to 100% AI-assisted review, with humans focused on the flagged 3–5% rather than picking calls at random.
- 9 million+ historical recordings backfilled and indexed — twelve years of complaint, retention, and outcome evidence now searchable in seconds.
- Complaint-call retrieval dropped from a typical 1–3 day exercise to under five minutes, including the surrounding context calls.
- Vulnerability identification became continuous, not agent-dependent — the second-pass model surfaces indicators on every call regardless of agent disposition.
- Board outcome reporting now cites specific call evidence per product line, with trend analysis the board can challenge.
“The first time we used the search to investigate a complaint — pulled every related call across eighteen months in under a minute — I realised we’d been doing this job with one hand tied behind our back for years.”
What this means for other UK firms
The reference deployment is now the basis for Telebyte’s productised offering. The same transcription, sentiment, and search stack is available to other FCA-regulated firms — brokers, claims handlers, IFA networks, debt management — either as a managed service (from £299/seat/month) or as a self-hosted install (from £4,500).
What other firms get is the second deployment of a system that’s already been running, every working day, on a UK contact centre handling regulated phone sales at production scale. Not a pilot. Not a vendor demo. The actual working stack.
The full case study, formatted for sales hand-off — three A4 pages.
Would the same stack work for you?
Telebyte runs a 30-minute readiness call — no slides, no sales pitch. Bring your current setup, agent count, and any specific FCA concerns. You’ll get a straight answer on whether Telebyte Call Analytics fits, where the install would land, and what it would cost.