Recursive learning loops · the Weekly Learning Cycle

Every call makes the next one better.

Graded, curated, fine-tuned, tested, approved, measured. Every week, on your own calls.

See a learning cycle live Get a Call Readiness report

7 stages · one loop100% of calls graded1,240 golden scenarios illustrative1-click rollback

RBI · IRDAI · DPDP mapped

Most voice bots are as good on day 300 as on day 1.
"It learns" is not a mechanism. This is.

Seven stages, one loop

A named pipeline, not a promise.

01 · Listen

Every call, every outcome, every correction

Recordings, transcripts, customer outcomes (PTP kept, payment received, policy renewed), human-agent takeovers and compliance flags stream in as signals — 100% of calls, not a sample.

02 · Grade

Every call scored on your rubric

Grounding, flow, compliance, tone. Failures are clustered: "repeated the EMI amount", "missed already-paid", "asked branch-or-link twice".

03 · Curate

Redact first, then choose the lessons

Aadhaar, PAN, card and phone numbers masked in text and audio. Your best agents' calls and the corrected failures become the training set — approved by your reviewer.

04 · Fine-tune

Your model, your data only

Speech and language models fine-tuned per tenant. No shared brain. The weights are yours, stored in your environment for VPC and on-prem.

05 · Simulate

Proven before it speaks

The candidate replays 1,240 golden scenarios — barge-in, angry customer, already-paid, wrong number — with Hindi/English parity and frozen-line checks. Any drop blocks release.

06 · Release

A human signs. Canary. One-click rollback

Named approval from your QA or compliance owner, staged rollout 5% → 25% → 100%, automatic rollback on a guardrail breach.

07 · Measure

Outcomes, not minutes

Containment, promise-to-pay, right-party contact, complaints and cost per resolved call — tracked per release, reported every morning. Then the next week begins.

Weekly cycle · collections · hi-IN

swipe → 7 stages

Listen

calls · outcomes · corrections

48,210 signals · 7d

Grade

every call scored on rubric

100% graded · 14 clusters

Curate

redact · select · approve set

PII 100% · 3,120 examples

Fine-tune

your model, your data only

candidate v42

Simulate

golden set · barge-in · parity

1,240 scenarios · 0 red

Release

human approval · canary · rollback

5% → 25% → 100% · 1-click back

Measure

PTP · containment · complaints

PTP +8.5 pts over 6 wks

↻ next week's calls feed the next cycle · client memory: scripts · glossary · frozen lines · escalation matrix

Compliance lines never learnHumans approve every releaseRegression before releaseOne-click rollback

Illustrative week for one collections cohort. replace with a live agent's counters

Recursive loops · turn → call → week → portfolio

TurnGrounding · policy · persona checks before a word is spoken

CallEvery call graded by the evaluation model; failures clustered

WeekGraded calls → fine-tune → simulate → human sign-off → release

PortfolioLessons shared across journeys and languages under one policy

Loops inside loops

Recursive, not just repeated.

The per-turn check produces the call grade. Call grades produce the week's training set. The week's release produces the portfolio's lessons. And the evaluation model that grades all of it is retrained on your reviewers' corrections — the loop that grades the loop.

Evaluation models The per-turn check

What actually learns

Your model. Your data. Nothing shared.

We fine-tune speech and language models on your graded, redacted calls. The improvement lives in weights you own.

tenant model · learning policy · v42

$ formant policy --tenant collections-hi --show

01 signals_inrecordings · QA grades on 100% of calls · outcomes · human corrections

02 redactAadhaar · PAN · card · phone masked in text and audio · complaints excluded

03 fine_tuneper tenant · your environment · no cross-client training, not even anonymised

04 provegolden set · barge-in · Hindi/English parity · frozen-line check

✓ frozen lines verified · 5 strings · hash match

enforced · learning policy v42

Frozen vs. learns

🔒 Frozen · never learnsversion-hashed, verified on every release

"Yeh call record ho rahi hai." — recording disclosure
"Main AI assistant hoon." — AI disclosure
Grievance redressal line and helpline number
RBI Fair Practices Code calling-hour and tone rules
IRDAI product-claim restrictions

Learns · under approval

Objection handling and phrasing per language
Intent recognition (new customer phrasings)
Number, date and amount pronunciation
Handle-time and turn-taking behaviour

Same call, one week later

v41 missed it. v42 learned it.

"Already paid" — flagged 212 times in a week. The next release looks up the UTR and closes the call.

replay · hi-IN · collections · v41 → v42

v41beforeQA 62

Customer

Maine kal UPI se pay kar diya tha.

Agent

Aapki 5 tareekh ki EMI ₹4,850 abhi tak pending hai. Kya aap aaj payment kar payenge?repeats the amount · misses "already paid"

v42after · learned from 212 corrected callsQA 94

Customer

Maine kal UPI se pay kar diya tha.

Agent

Theek hai, dhanyavaad! Main abhi check kar rahi hoon… UTR mil gaya, payment 15 September ko receive hua hai. Aapko aage koi call nahi aayegi.stage: already_paid → utr_lookup → close ✓

Compounding, measured

Rising week over week, with the dip left in.

When v41 breached a live guardrail, canary caught it. We show that, because unseen failure cannot be trusted.

cohort · early bucket · releases v36 → v42

Containment (calls resolved without a human)Promise-to-pay rate (right-party contacts)Rollback event

Illustrative shape of six weeks of weekly learning cycles on one early-bucket cohort; every release marked, the dip kept in. Pilot data replaces this. replace with cohort data + methodology note

Rollback

41 min

canary breach → previous version back illustrative

Regression

1,240

golden scenarios before every release illustrative

Governed release

Every version has a name, an approver and a way back.

versions · collections-hi · v38 → v42

v38

17 Aug

baseline · approved

v39

24 Aug

objection handling

v40

31 Aug

Marathi legal-threat transfer

v41

7 Sep

rolled back 41 min · fixed · re-released

v42

14 Sep

already-paid · live

every node: approver · regression score · one-click rollback

release log · agent collections-hi

$ formant releases --agent collections-hi --last 3

v42 14 Sepalready-paid via UPI (212 corrected calls) · no EMI repeat · "19 tareekh" dates
frozen RBI disclosure · consent · grievance · approved client QA lead + FDE
regression 1,240 / 1,240 · ● live · canary → 100%

v41 7 Sep"branch ya link?" asked once · 9 new salary-delay phrasings
regression 1,198 / 1,198 · ▲ canary breached tone guardrail → auto-rollback 41 min → fixed → re-released

v40 31 AugMarathi legal-threat warm transfer · shorter greeting (−4 s AHT)
regression 1,150 / 1,150 · ✓ approved client compliance + FDE

rollback one click · previous version warm

Example release notes for illustration. replace with a live agent's notes

No learning without control

How a bank can accept a self-learning agent.

Compliance lines never learn.: Disclosures, consent and grievance lines are frozen, version-hashed strings.
Humans approve every release.: Named sign-off from your QA or compliance owner and our engineer, logged.
Regression before release.: Golden scenarios and last week's failures; any compliance drop blocks it.
Canary, then scale.: 5% → 25% → 100%, auto-rollback on breach, previous version always warm.
Learn only from what you allow.: Per-client boundaries, PII redacted before storage, pause per journey.
Every change is explainable.: Which calls taught it, who approved it, scores before and after — for RBI, IRDAI and audit.

Autonomy ladder

Climb at your pace. Own every level.

Most BFSI deployments start at L3. The weekly cycle takes you to L4, under the policy your compliance office signed.

Get your Call Readiness report

autonomy · L1 → L5

L1
Scripted
IVR-style prompts; every path pre-written.
—
L2
Assisted
Agent drafts, human agent speaks or approves.
✓
L3
Autonomous with review
Agent handles calls; 100% graded after; humans review flags.
✓
L4
Self-learning, governed
Weekly cycle: learn from graded calls, regression-tested, human-approved releases, frozen compliance lines.
● FormantAI today
L5
Portfolio
Learning shared across journeys and languages under one governance policy.
○ next

FAQ

The questions your risk team will ask.

Does the model train on our customers' data?

Yes — on your consented, redacted call data, into a model that belongs to you and serves only you. Never used for a shared or foundation model, even anonymised. Training runs inside your environment for VPC and on-prem.

Can the agent learn something it should not say?

Regulator-mandated lines are frozen strings the model cannot alter. Everything else is regression-tested against your golden scenarios and a compliance rubric, then approved by a named human. A canary breach restores the previous version automatically.

Can we pause learning, and how often does it run?

Weekly by default; cadence, approvers and exclusions are set in the pilot charter and you can change them. Learning can be paused or frozen per agent or journey — the agent keeps running on its current approved version.

Who owns the fine-tuned model?

You do. Weights are stored in your environment in VPC and on-prem deployments; in the India-region cloud they are tenant-isolated and exportable.

How do we audit what changed?

Every version has a changelog (source calls, changes, approver, before/after scores), a regression report and a rollback record — exportable for RBI, IRDAI and internal audit.

Under NDA, before any contract

Watch one week learn on your calls.

Book a learning-cycle demo

Readiness

Know what your calls can teach.

Get a Call Readiness report