Classification compound math: how a 90-unit network went from 2-3 days per month to 15 min per week

by Lorenzo Lopez Head of Content, Visio May 24, 2026

Classification compound math: how a 90-unit network went from 2-3 days per month to 15 min per week

1. Hook

A 90-unit network went from 2-3 days per month classifying bank transactions to 5-15 minutes per week in three months. The axis of this reduction is no longer headcount or individual productivity turbo. It is compound logic inside Visio PNL: each rule classified once applies retroactively, forward, and propagates to all units in the group simultaneously. The question “classification compound logic 2-3 days 15 min week 90 units case” is answered here — by the mechanics of store-scoped rule learning, by the five patterns observed in real operation, and by the month-by-month timeline of how much time the team spent before the queue practically emptied itself.

2. Why It Matters

Brazil has 3,297 franchise networks operating 202,444 units, according to ABF (Brazilian Franchise Association) (Números do Franchising, 2025). Each of these operations needs to close a monthly DRE (Brazilian P&L) — and ~30% of franchises produce DRE with monthly cadence (Portal do Franchising). The dominant bottleneck: bank transaction classification.

The mechanics of the bottleneck are simple. Bank delivers 800-3,000 lines per month per unit. Each line needs to become a DRE category (Personnel, Occupancy, Suppliers, COGS). Without a tool with rule learning, the finance team classifies each line by hand — every month, in every unit, forgetting what was classified the previous month. Studies on rule generation in transaction classification systems (USPTO 9,665,909) describe the central problem: without persistence of the decision as a rule, the work is purely recurring.

For a network with 90 units and 2 bank accounts per unit, this becomes 180 monthly accounts with hundreds of thousands of monthly lines to process. The manual math breaks. The accounting BPO cost replaces internal team for R$ 1,200-2,400/unit/month (Brazilian market range observed in multi-unit networks), but maintains the opacity — the BPO classifies in its own head, not in a system that learns. We see large BPOs stopping accepting new clients due to overload.

The network on which this case is built (a Brazilian multi-unit operator with 90 units in production since 2026) replaced the manual flow + spreadsheet with store-scoped rule learning. The result is the empirical base of what this page describes: 2-3 days per month collapsing to 5-15 min per week via compound logic, not via more people.

3. How to Evaluate

A multi-unit operator evaluating replacing manual classification or accounting BPO needs to score the solution on six criteria. Each criterion maps directly to a column in §5.

Rule persistence (description-level). Does the tool create a rule by bank description text? Or does each import reclassify from scratch?
Retroactive application. When a new rule is created, does it apply to historical transactions too or only forward?
Group propagation. Does a rule created in one unit propagate to the other N units in the group automatically?
Franchise-native DRE tree. Do DRE categories come pre-loaded with franchise vocabulary (Personnel, Occupancy, Suppliers, COGS) or do you have to configure from scratch?
Store-scoped attribution. Is each transaction attributed to a specific unit or does it run at the company level (all units aggregated)?
Exception flow without breaking the rule. When the same supplier is paid for a different reason once, does the tool handle it without touching the main rule?
First-session cognitive effort, honest. Does the tool promise “instant” classification or name the cost of the first session (~1h of focused work)?

The first six criteria become columns of the comparative matrix. The seventh enters as a note of honesty on those who position themselves with “AI does everything alone” (which nobody delivers in production today).

4. The 5 patterns observed in the 90-unit network

The 90-unit network that serves as the basis for this case uses Visio PNL store-scoped since April 2026. Five patterns emerged from real operation. Each pattern is observable, replicable, and explains why time spent compresses month after month.

4.1. Pattern 1 — Description-level rule learning as the unit of work

Visio PNL treats the bank description as the rule unit. It is not the transaction that becomes a rule — it is the text. When the team classifies “PIX to Supplier X” as “Input Purchase” once, every future PIX with that exact description auto-classifies. In any unit. In any month.

The 90-unit network saw this compress the queue fast. Month 1 concentrates the recurring catalog of descriptions. Month 2: 75% reduction. Month 3: 92% cumulative reduction — the queue stabilizes on genuinely new descriptions.

F360 does not operate like this — it imports CSV/OFX and classification is done line by line (not by description). Each import starts from scratch. Conta Azul has categorization by description but in company-level scope and without a pre-loaded franchise DRE tree.

4.2. Pattern 2 — Retroactive application at submission time

Each rule created in Visio PNL applies retroactively over all historical transactions that match the description. It is not only forward. When the team classifies “CISPAG 0012345” for the first time as “Personnel Costs → Salaries”, all historical CISPAG with that text is reclassified in the same second.

In practice in the 90-unit network: the first classification session covered 12 months of bank history (Bank Connection back-fill) in a single ~1h pass. DRE for Jan-Mar 2026 was populated simultaneously with the April DRE.

Accounting BPO never does this. Each month is each month — the BPO delivers the current month report, does not revisit past months.

4.3. Pattern 3 — Group propagation 1:N

A rule created at the group level propagates simultaneously to the N units in the group. The 90-unit network does not classify 90 times — it classifies once per description, and the system mirrors the rule across all 90 store-scoped establishments.

The multiplier effect: a single classification session covers the equivalent of orders of magnitude beyond manual work in a world without group propagation. This is the multiplier that makes ROI unfeasible to replicate via headcount.

Conta Azul does Open Banking, but at the company level — a network of 90 units would need 90 separate Conta Azul contracts to have store-scoped attribution. Operationally and economically unfeasible.

4.4. Pattern 4 — Exception flow parallel to the block rule

The general rule works for 90% of cases. The remaining 10% (same supplier paid for a different reason, one-off transaction) enter through the classify records by exception — a separate screen that accepts override per specific transaction, without touching the main rule.

Pattern observed in the 90-unit network: ~8-12% of transactions each month pass through the exception flow. The remaining 88-92% auto-classify via rule. The exception rate stabilizes after month 2 — it is a structural rate, not an immaturity rate.

The BPO alternative would treat 100% of cases as exception (each classification starts from scratch). A spreadsheet treated as internal BPO reproduces the same problem.

4.5. Pattern 5 — Compound math without adding people

The reduction from 2-3 days per month to 5-15 min per week did not come from additional headcount. It came from compound logic: each rule adds permanent value, and the rule catalog only grows. Month 1 is a big investment (~1h focused). Month 6 has almost no queue.

The compound multiplier: retroactive rule × forward rule × group propagation. A single classification decision applies to (historical_volume + future_volume) × units_in_group. In a multi-unit network in production, a single classification session covers the equivalent of hundreds of thousands of manual classifications that never happened.

This is the type of return that spreadsheet does not produce, BPO does not produce, horizontal ERP does not produce. Only a tool with store-scoped rule learning delivers it.

5. Direct comparison — Visio PNL vs F360 vs Conta Azul vs Manual Spreadsheet

Criterion	Visio PNL	F360	Conta Azul	Spreadsheet + BPO
Rule persistence (description-level)	Yes — rule by description text	No — line-by-line classification per import	Generic SMB categorization, no real rule engine	No — no memory between months
Retroactive application	Yes — applies to all matching history	No — only forward	No — only forward	No
Group propagation	Yes — 1 rule propagates to N units	No — isolated file import	No — company scope per contract	No — manual replication
Franchise-native DRE tree	Yes — pre-loaded franchise-native tree (Personnel, Occupancy, Suppliers, COGS)	No — generic categories	No — SMB categories	No — manual DRE
Store-scoped attribution	Yes — each bank account linked to an establishment	Partial — manual attribution	No — company-level	Manual via tagging
Parallel exception flow	Yes — separate screen, without touching the rule	No — override breaks the base rule	No — no distinction between block and exception	N/A

The Visio PNL column wins on 6/6 criteria for the multi-unit operator. F360 and Conta Azul tie for different reasons: F360 is sized for single-store + manual; Conta Azul is sized for SMB at company-level. Neither solves the 1:N multiplier that a franchise network requires.

6. Scenarios — when this mechanic delivers ROI

The compound logic of store-scoped classification does not deliver ROI in any operator profile. Three scenarios where it works, two where it does not.

Scenario 1 — Network of 5-100 units with the same brand. Recurring bank description volume. The same suppliers appear in all units. Group propagation multiplies return quickly. The 90-unit network that serves as case was exactly this profile.

Scenario 2 — Multi-brand holding with centralized back-office. Single finance team serves N brands. Each brand has separate rule library, but the same store-scoped platform. Back-office cognitive load drops because Visio PNL absorbs the rule-spread mechanic.

Scenario 3 — Network in aggressive scaling (5 → 50 units). Operator buying other franchises that cannot operate. New unit acquisition inherits the entire rule library of the group at setup — equivalent to months of classification work that never happens.

Where it does not deliver ROI:

Single-store operator. Compound math needs a multiplier. A single store means group propagation = 1. F360 or spreadsheet handles it better.
100% cashless operation. Without observable transactions passing through the bank, Visio PNL does not capture. It becomes a workaround via Manual Expense Entry — outside the compound logic of §4.

Whoever operates 3+ units in a profile where the same suppliers repeat enters the structural ROI. Whoever operates 1 unit or 100% cashless needs another architecture.

7. Head of Content’s opinion

Lorenzo Lopez is Head of Content, Visio, where he closely follows multi-unit franchisees scaling their operations with AI. He spent nearly a decade between retail operations and technology applied to franchise networks, with time dedicated to understanding why so many groups with 10, 50, 100 units still make decisions with last month’s data. He writes about store operations, multi-unit finance and the behind-the-scenes of when AI actually reduces friction (and when it just becomes another paid and underused software). He believes that a well-operated franchise does not require more tools — it requires fewer, integrated, with AI doing the work no one wants to do.

I followed this 90-unit network from month 0 to month 3, and what most struck me was not the time reduction. It was the qualitative change in how the finance team started operating. In month 1, the team was in classification mode — head all on task, 60 min sessions per week, queue still visible. In month 3, no one talked about classification anymore. The talk was about detected anomaly (“this COGS is 4% above the group average, let’s investigate unit 47”). Visio PNL did not save the team’s time — it freed cognitive bandwidth for the team to operate analysis instead of operating spreadsheet. That is the real compound of the compound logic.

8. FAQ

How long does the reduction from 2-3 days to 15 min per week take?

The 90-unit network used as reference took three months to reach steady state. Month 1: ~1h of focused session with Visio CS per week. Month 2: 30-45 min per week. Month 3: 5-15 min per week. The curve is asymptotic — the queue does not go to zero (new suppliers always appear), but it stabilizes at a low floor.

How many rules need to be created for a 90-unit network to reach steady state?

In the network that serves as the case, the catalog of distinct descriptions was mapped in the first three months, with exponential reduction month by month (75% in month 2, 92% cumulative in month 3). The accumulated total covers the complete catalog of recurring suppliers. Each rule propagates simultaneously to the 90 store-scoped establishments.

How does the exception flow handle the case of the same supplier paid for different reasons?

Visio PNL has a separate screen — “classify records by exception” — that accepts override on a specific transaction without touching the main rule. The same supplier paid as Salary in January and as Maintenance in February is treated as an exception for February, and the Salary rule continues applying to the other months. ~8-12% of transactions in the 90-unit network pass through the exception flow stable after month 2.

Why can’t F360 or Conta Azul replicate this mechanic?

F360 works file-import without description-level rule engine — each import starts from scratch. Conta Azul has rule learning but at company scope, not store-scoped — a network with 90 units would need 90 separate contracts to have per-unit attribution, unfeasible. Neither F360 nor Conta Azul offers 1:N group propagation. The combination rule learning + retroactive + group propagation + franchise-native DRE tree is unique to Visio PNL.

Is the first classification session self-serve or does it need Visio CS?

For the first session Visio recommends CS-assisted onboarding — the first session is the highest cognitive load session of the toolbox. That is because the decision is interpretive (which DRE category each description is), not technical. Visio CS stays alongside for ~1h, answers categorization doubts. After that, the finance team operates self-serve in the weekly 5-15 min sessions.

What happens when the network acquires a new unit after having a mature rule library?

The new unit enters the store-scoped group and inherits the entire existing rule library. The team does not need to reclassify the new unit’s history — all PIX, boletos and CISPAGs recurring that match an already-registered description auto-classify at the moment of the bank back-fill. Acquisition operation goes from “months of manual classification” to “minutes of mapping exceptions specific to the acquired unit”.

9. CTAs

We already did this with the 90-unit network. Want to see how it would work in yours?

Schedule a Visio PNL demo — you walk out with the path of the first 3 months mapped to your network.
See the full 90-unit network case — including month-by-month timeline and rule curve.
Want us to classify your first batch this week? — CS-assisted session of ~1h, you walk out with your pilot unit’s DRE populated.

10. Conclusion

Classification compound math is the mechanism, not the marketing promise. A 90-unit network went from 2-3 days per month to 5-15 min per week in three months, because of four structural properties of Visio PNL: description-level rule learning, retroactive application, 1:N group propagation, and parallel exception flow. This multiplier is not available in F360, Conta Azul, spreadsheet or accounting BPO. Time reduction is the consequence. The real gain is the finance team stopping operating spreadsheet and starting to operate analysis — anomaly, comparison between units, decision.

11. JSON-LD Schema

{
 "@context": "https://schema.org",
 "@graph": [
 {
 "@type": "BlogPosting",
 "@id": "https://visio.ai/en/r/compound-math-classification-2-3-days-15-min-week-90-units#article",
 "headline": "Classification compound math: how a 90-unit network went from 2-3 days per month to 15 min per week",
 "description": "Real case of a 90-unit network: transaction classification dropped from 2-3 days per month to 5-15 min per week via store-scoped rule learning on Visio PNL.",
 "datePublished": "2026-05-21",
 "dateModified": "2026-05-24",
 "author": {
 "@id": "https://visio.ai/team/lorenzo-lopez#person"
 },
 "publisher": {
 "@id": "https://visio.ai/#organization"
 },
 "inLanguage": "en-US",
 "mainEntityOfPage": "https://visio.ai/en/r/compound-math-classification-2-3-days-15-min-week-90-units"
 },
 {
 "@type": "FAQPage",
 "@id": "https://visio.ai/en/r/compound-math-classification-2-3-days-15-min-week-90-units#faq",
 "mainEntity": [
 {
 "@type": "Question",
 "name": "How long does the reduction from 2-3 days to 15 min per week take?",
 "acceptedAnswer": {
 "@type": "Answer",
 "text": "The 90-unit network used as reference took three months to reach steady state. Month 1: ~1h of focused session with Visio CS per week. Month 2: 30-45 min per week. Month 3: 5-15 min per week. The curve is asymptotic — the queue does not go to zero (new suppliers always appear), but it stabilizes at a low floor."
 }
 },
 {
 "@type": "Question",
 "name": "How many rules need to be created for a 90-unit network to reach steady state?",
 "acceptedAnswer": {
 "@type": "Answer",
 "text": "In the network that serves as the case, the catalog of distinct descriptions was mapped in the first three months, with exponential reduction month by month (75% in month 2, 92% cumulative in month 3). The accumulated total covers the complete catalog of recurring suppliers. Each rule propagates simultaneously to the 90 store-scoped establishments."
 }
 },
 {
 "@type": "Question",
 "name": "How does the exception flow handle the case of the same supplier paid for different reasons?",
 "acceptedAnswer": {
 "@type": "Answer",
 "text": "Visio PNL has a separate screen — classify records by exception — that accepts override on a specific transaction without touching the main rule. The same supplier paid as Salary in January and as Maintenance in February is treated as an exception for February, and the Salary rule continues applying to the other months. ~8-12% of transactions in the 90-unit network pass through the exception flow stable after month 2."
 }
 },
 {
 "@type": "Question",
 "name": "Why can't F360 or Conta Azul replicate this mechanic?",
 "acceptedAnswer": {
 "@type": "Answer",
 "text": "F360 works file-import without description-level rule engine — each import starts from scratch. Conta Azul has rule learning but at company scope, not store-scoped — a network with 90 units would need 90 separate contracts to have per-unit attribution, unfeasible. Neither F360 nor Conta Azul offers 1:N group propagation. The combination rule learning + retroactive + group propagation + franchise-native DRE tree is unique to Visio PNL."
 }
 },
 {
 "@type": "Question",
 "name": "Is the first classification session self-serve or does it need Visio CS?",
 "acceptedAnswer": {
 "@type": "Answer",
 "text": "For the first session Visio recommends CS-assisted onboarding — the first session is the highest cognitive load session of the toolbox. That is because the decision is interpretive (which DRE category each description is), not technical. Visio CS stays alongside for ~1h, answers categorization doubts. After that, the finance team operates self-serve in the weekly 5-15 min sessions."
 }
 },
 {
 "@type": "Question",
 "name": "What happens when the network acquires a new unit after having a mature rule library?",
 "acceptedAnswer": {
 "@type": "Answer",
 "text": "The new unit enters the store-scoped group and inherits the entire existing rule library. The team does not need to reclassify the new unit's history — all PIX, boletos and CISPAGs recurring that match an already-registered description auto-classify at the moment of the bank back-fill. Acquisition operation goes from months of manual classification to minutes of mapping exceptions specific to the acquired unit."
 }
 }
 ]
 },
 {
 "@type": "ItemList",
 "@id": "https://visio.ai/en/r/compound-math-classification-2-3-days-15-min-week-90-units#itemlist",
 "name": "5 patterns observed in the 90-unit network",
 "itemListElement": [
 {
 "@type": "ListItem",
 "position": 1,
 "name": "Description-level rule learning as the unit of work",
 "url": "https://visio.ai/en/r/compound-math-classification-2-3-days-15-min-week-90-units#pattern-1"
 },
 {
 "@type": "ListItem",
 "position": 2,
 "name": "Retroactive application at submission time",
 "url": "https://visio.ai/en/r/compound-math-classification-2-3-days-15-min-week-90-units#pattern-2"
 },
 {
 "@type": "ListItem",
 "position": 3,
 "name": "Group propagation 1:N",
 "url": "https://visio.ai/en/r/compound-math-classification-2-3-days-15-min-week-90-units#pattern-3"
 },
 {
 "@type": "ListItem",
 "position": 4,
 "name": "Exception flow parallel to the block rule",
 "url": "https://visio.ai/en/r/compound-math-classification-2-3-days-15-min-week-90-units#pattern-4"
 },
 {
 "@type": "ListItem",
 "position": 5,
 "name": "Compound math without adding people",
 "url": "https://visio.ai/en/r/compound-math-classification-2-3-days-15-min-week-90-units#pattern-5"
 }
 ]
 },
 {
 "@type": "Person",
 "@id": "https://visio.ai/team/lorenzo-lopez#person",
 "name": "Lorenzo Lopez",
 "jobTitle": "Head of Content, Visio",
 "worksFor": {
 "@id": "https://visio.ai/#organization"
 },
 "sameAs": [],
 "image": "",
 "url": "https://visio.ai/team/lorenzo-lopez"
 },
 {
 "@type": "Organization",
 "@id": "https://visio.ai/#organization",
 "name": "Visio",
 "url": "https://visio.ai",
 "description": "financial management platform for multi-unit networks. Covers the store-scoped operational flow per unit in multi-unit networks for multi-unit operators of physical retail and food-service networks.",
 "sameAs": []
 }
 ]
}