FibeAI Suite

Growth workspace

Agent team & evals

The DeployOne voice and reasoning agents running Fibe's growth pilot, with their evaluation suite and guardrails.

Active agents

6

4 voice-enabled

Calls handled (30d)

34.5K

↑ 12.8% · pilot estimate

Avg eval pass rate

89.2%

↑ 3.1 pts vs last month

Guardrail flags (30d)

17

↓ 22% after prompt patch

Qualifier

Pre-screens inbound loan interest via voice

ModelDeployOne-Voice v2 + GPT-4o
ToolsCRM lookup, bureau ping, KFS reader
Runs (30d)18,420
Success rate88.4%
Cost / unit₹2.10/call

Nudge

Reminds warm leads to complete KYC/e-sign

ModelDeployOne-Voice v2
ToolsWhatsApp send, calendar hold
Runs (30d)9,860
Success rate71.2%
Cost / unit₹1.35/call

Winback

Re-engages dropped-off applicants

ModelDeployOne-Voice v2 + GPT-4o-mini
ToolsOffer engine, CRM update
Runs (30d)6,240
Success rate34.6%
Cost / unit₹1.80/call

Insight Analyst

Summarises call themes for growth team

ModelGPT-4o
ToolsTranscript store, BI export
Runs (30d)34,510 calls analysed
Success raten/a
Cost / unit₹0.42/call

Creative Studio

Drafts ad and script variants from signals

ModelGPT-4o + image gen
ToolsBrand kit, asset library
Runs (30d)1,120 assets
Success raten/a
Cost / unit₹6.50/asset

Reputation Watch

Clusters app-store and review themes

ModelGPT-4o-mini
ToolsPlay Store API, App Store API
Runs (30d)Daily sweep
Success raten/a
Cost / unit₹0.08/review

Eval suite

Automated evals run against golden transcripts

Pilot estimate
EvalDatasetPass rateLast runTrend
Qualifier — eligibility accuracy500 cases91.2%2 hrs ago↑ improving
Qualifier — disclosure compliance500 cases96.8%2 hrs ago→ stable
Nudge — tone & consent320 cases88.0%6 hrs ago↑ improving
Winback — objection handling280 cases76.4%1 day ago↓ regressing
Insight Analyst — summary faithfulness400 cases93.5%1 day ago↑ improving

Eval score trend

Composite weighted score across all suites

Failure-case review

Sample transcripts flagged by evals or guardrails, pending human verdict

F-2291WinbackHigh risk

"So basically there's no real cap on how much you can borrow..." — agent overstated loan ceiling beyond ₹10L policy.

F-2287QualifierMedium risk

Agent skipped reading the cooling-off period clause before confirming interest.

F-2270NudgeMedium risk

Called outside the approved 8am–7pm window due to timezone mismatch.

Guardrail policy

Live rules applied to every voice turn

Never state a rate below Fibe's published ~18% p.a. floor without a human loan officer confirming eligibility.
Loan ceiling capped at ₹10 lakh in all agent responses; no speculative amounts.
Calling window restricted to 8:00am–7:00pm IST per RBI fair-practices guidance.
Any consent revocation or hardship mention triggers immediate handoff to a human agent.
All calls opened with recorded-line disclosure; guardrail blocks call start if disclosure step is skipped.