Under the Hood: Ambill’s Dynamic AR Categorization & Follow-Up Engine

How Ambill turned fragmented, manually-owned collections across a 40+ branch dealer network into a self-calibrating, fully auditable follow-up machine.

For manufacturers and distributors selling through multi-tier dealer networks, accounts receivable follow-up is rarely a technology problem on paper — and almost always one in practice. Thousands of dealers. Dozens of branches. Collection owned informally by whichever field associate "knows the account." No system connecting collection activity, escalation state, or billing decisions.

The result is predictable: reliable dealers get over-called, risky dealers slip quietly into 60+ day buckets, and the AR Head discovers a problem account only after it has already become a provisioning line item.

We recently built and deployed an AR categorization and follow-up orchestration engine for one of India's leading electrical products manufacturers — a business running its dealer receivables across 40+ branches on an in-house ERP. This post walks through how the engine works: how every dealer is scored, categorized, and routed through an automated IVR, WhatsApp, and email follow-up sequence, with humans pulled in exactly where they add value.

⏱️ TL;DR for CFOs

The Core Problem: In dealer-network businesses, collections follow-up is fragmented and person-dependent. There is no systematic answer to "which dealers deserve a gentle nudge, which need a call before the due date, and which need the Branch Head involved today" — so effort is misallocated and risk surfaces late.
The Solution: Ambill's engine computes a composite behavioural score from 15 signals across 7 families (payment behaviour, credit exposure, trend, relationship value, financial impact, contactability, and field intelligence), grades every dealer against the live portfolio, and maps each category to an automated IVR + WhatsApp + email follow-up cadence with defined human escalation paths.
The Bottom Line: Every dealer is re-scored every refresh cycle, every category assignment is traceable to its underlying signals (no black box), and follow-up intensity is matched to actual risk — so finance bandwidth concentrates on the accounts that genuinely need it, and deteriorating accounts are flagged before their ageing looks alarming.

## The Core Process Flow

The engine runs alongside the client's ERP as an independent service — no ERP replacement, no ledger migration. Dealer, invoice, and payment records sync bidirectionally via API and database-level connectors, and five logical stages compose the pipeline:

Feature store — a nightly ERP delta sync populates a dealer-level feature table, versioned per refresh cycle so category history is fully auditable.
Categorization engine — computes the composite behavioural score and assigns (or re-assigns) each dealer's category.
Rules engine — maps category $\times$ invoice ageing to an active follow-up strategy and cadence.
Channel orchestrator — sequences IVR, WhatsApp, and email per the active strategy, with call and message outcomes feeding back into the feature store.
Dashboard layer — role-scoped views for the field associate, the Branch Head, and the AR Head.

Step 1: Fifteen Signals, Seven Families

Each refresh cycle, the engine computes fifteen features per dealer, organized into seven signal families:

Family	What it captures	Used for
P — Payment Behaviour	Recency-weighted average days-to-pay, promise-to-pay reliability, dispute frequency, chronic part-payment patterns, and decayed write-off history	Risk score
C — Credit Risk & Exposure	Credit-limit utilization, how many consecutive cycles a dealer has stayed maxed-out, and the Ind AS 109 ECL provisioning stage the account sits in	Risk score
T — Trend & Context	The slope of payment behaviour over recent cycles, and deviation from the dealer's own seasonal baseline — so festive-season stocking reads as normal, not risk	Risk score
V — Relationship & Value	Tenure $\times$ reorder consistency, plus a strategic-importance flag that caps follow-up aggressiveness regardless of score	Risk score
I — Financial Impact	DSO contribution, revenue share, and cost-to-collect ratio	Prioritization (separate axis)
E — Contactability	Per-channel read/response rates from IVR, WhatsApp, and email logs	Channel routing only
F — Field Intent Signal	A subjective 1–5 willingness-to-pay score logged by the field associate after each dealer interaction	Risk score (calibrated)

Two families are deliberately kept out of the risk score. Financial impact forms a separate prioritization axis — a large, reliable dealer will always carry large outstandings, and size alone should never push a good customer into a risky category. And contactability never touches risk at all: how a dealer prefers to be reached decides which channel sequence the orchestrator uses, not what category he sits in.

The trend family deserves a special mention for CFOs: the payment-trend slope flags a worsening account before its days-past-due level looks alarming. That is the difference between a follow-up call this month and a provisioning conversation next quarter.

Step 2: Grading on a Curve, Not Against Fixed Benchmarks

Every dealer is scored relative to the current dealer population — is this dealer better or worse than the portfolio norm on each measure? — with extreme one-off outliers capped so a single freak event cannot distort the scale.

This design choice matters more than it sounds. Because the reference point is the live portfolio itself, the model automatically re-centers as the dealer base grows or the payment culture shifts. There is no annual "re-benchmarking exercise," no stale thresholds hardcoded three years ago that nobody remembers the rationale for.

Scoring is also layered, not flat. Signals first roll up within their own family — the five payment-behaviour signals combine into one payment-behaviour read, the credit-exposure signals into one credit read — and those family-level reads then combine into the single composite score. The relative weights are fitted to the client's own ledger history during deployment, not launched with generic defaults. Directionally:

Demonstrated payment behaviour carries the most influence.
Direction-of-travel and credit discipline carry meaningful weight.
Relationship history earns benefit-of-doubt at the margins.
The human field signal starts small until proven.

The layered structure has a practical payoff: every family read is a standalone, auditable intermediate the AR Head can inspect on the dashboard — "this dealer is Category B mainly because his trend is worsening, not because of disputes."

Step 3: Keeping the Human Signal Honest

The field-review score is the only input not derived from the ERP — and the noisiest. It reflects one field associate's read of one dealer interaction. Three safeguards keep it useful without letting it distort categorization:

Every rater is graded on their own curve: One associate rates everyone harshly, another generously. Each rater's reviews are adjusted against their own historical rating pattern before being pooled, so a strict rater's "3" and a lenient rater's "3" are read correctly.
Disagreements are surfaced, not averaged away: Where multiple associates have reviewed the same dealer over time, the system tracks rater agreement. A dealer whose reviews persistently conflict is flagged for manual category review instead of being silently auto-scored.
Trust is earned, not assumed: The field signal starts with a deliberately small influence on the composite score, increased only once live data confirms it actually predicts payment behaviour. A human rating is the easiest input to game and the hardest to audit — so it starts on probation.

Step 4: Four Categories, Assigned by Distribution

Categories are assigned by where a dealer sits within the portfolio's score distribution — relative cutoffs, re-fit every refresh cycle, not fixed thresholds:

Category	Position in portfolio	Typical share
A — Reliable	Top band of the score distribution	~45%
B — Moderate/erratic	Middle band	~35%
C — Chronic/high-risk	Bottom band	~20%
D — Unscored	Insufficient history	Cohort-dependent

Crucially, a dealer's category is not overwritten on a single-cycle score change. A dealer only migrates when the new score persists across two consecutive refresh cycles, or moves by a large, unambiguous margin in one. This hysteresis band prevents dealers from flapping between B and C on ordinary month-to-month noise — which would otherwise generate constant escalation churn for the field team.

Every transition is logged with a timestamp, prior category, new category, and triggering score delta. That transition log is itself a signal: a dealer who flips categories frequently is, independent of his current letter, a higher-attention account.

Step 5: Category Drives Cadence; Exposure Drives Urgency

Category alone sets follow-up intensity. Exposure is layered separately as an Impact axis for prioritization within a category — a large B-category dealer and a small B-category dealer receive the same cadence but different escalation urgency.

Category	Primary channels & timing	Escalation trigger
A	IVR + WhatsApp reminder near/at due date	No response past due date $\rightarrow$ single associate check-in call
B	IVR + WhatsApp + email from due date; earlier associate involvement	No response within a shortened SLA $\rightarrow$ associate call $\rightarrow$ Branch Head
C	Proactive associate call before due date + IVR/WhatsApp reinforcement	Any slippage $\rightarrow$ immediate Branch Head visibility
D	Conservative default cadence, treated as B/C	Standard escalation until history accumulates

The rules live in a configuration table, not in code — so credit control can retune thresholds without raising an engineering change request.

The IVR layer is a bounded, keypad-driven flow: the call states the outstanding amount and due date, and the dealer's keypad response maps to a fixed intent set — confirm a payment date, request a callback, or dispute the amount. Every outcome (answered, response captured, no-answer, invalid number) is logged and fed back to the rules engine for the next channel decision. No-response and dispute outcomes trigger automatic handoff to a human. This feedback loop is how the engine stays current between full refresh cycles.

Why Not Machine Learning on Day One?

A fair question — and a deliberate design choice. An ML classifier is intentionally deferred until 2–3 cycles of realized outcomes (paid-on-time vs. defaulted) exist as training labels. For a first deployment, a categorization the credit team can fully trace beats a black box it has to take on faith.

Every category assignment is traceable to its family sub-scores, underlying feature values, and the field-review input — by design, for AR Head audit. Once live outcome data accumulates, it becomes the training set for the upgraded model.

Conclusion: Strategic Financial Impact

For the CFO, systematic dealer categorization is not a collections tool — it is receivables risk management embedded into daily operations. Three advantages stand out:

Effort lands where risk lives: Reliable dealers get a light automated touch; erratic dealers get earlier human attention; chronic accounts get proactive calls before the due date. Finance bandwidth stops being spread evenly across accounts that don't need it.
Deterioration surfaces early: Trend-slope signals and category-transition logs flag a worsening account while it is still a follow-up conversation — not yet an ECL stage migration. Because the scoring links directly to Ind AS 109 provisioning buckets, the operational view and the reported-numbers view finally speak the same language.
Every decision is defensible: Versioned feature snapshots, logged category transitions, and traceable sub-scores mean the AR Head can answer "why is this dealer being escalated?" with data, not anecdote — and the audit trail writes itself.

Collections in a dealer-network business will always involve relationships, judgment, and the occasional difficult phone call. What it should never involve is guesswork about who to call, when, and how hard. That is what the engine now decides — every cycle, for every dealer, on the evidence.

Ambill builds AI-powered financial operations automation — accounts receivable, reconciliation, and payment collection — that runs alongside your existing ERP. [See Ambill in action].