Pattern learning

The problem

Expense qualification follows precise tax rules, but also choices specific to each company:

• pro/personal split of a phone;
• categorization of a recurring supplier;
• tax treatment of a type of expense.

These are decisions that repeat themselves.

Without learning, the agent asks the same questions for every invoice from the same supplier.

---

The solution: two levels of rules

Factory rules

Predefined tax rules, drawn from the BOFiP and accompanied by their legal references: a manager’s meal taken alone at the standard rate, fuel deductible according to the engine type, public transport with no recoverable VAT. These rules are immutable, versioned and updated with the product.

Learned patterns

Choices that Simon remembers, by supplier and category:

TOTAL + fuel → 80% deductible, recoverable VAT

landlord + rent → 20% deductible, non-recoverable VAT

ORANGE + telecom → 50% deductible, recoverable VAT

Key insight

Each manual correction enriches the system. A supplier qualified once is qualified automatically the following times.

---

The qualification flow

flowchart TD
    A[Document ingested] --> B{Existing pattern?}
    B -->|Yes| C[Apply silently]
    B -->|No| D{Factory rule?}
    D -->|Without validation| E[Auto-qualify]
    D -->|With validation| F[Propose to the user]
    D -->|No| G[Manual qualification]
    F --> H{Confirms?}
    H -->|Yes| I[Qualify + remember]

1. Pattern found → silent qualification (no question)
2. Rule without validation → automatic qualification (fuel, transport…)
3. Rule with validation → proposal then learning
4. No rule → manual qualification

---

Obsolescence detection

When a factory rule changes (an update of the legal rates), Simon detects that the existing patterns use a rate different from the new default rate and proposes the update.

---

Why this design

Repetition is the norm

The same suppliers come back every month. Learning eliminates the repetitive work.

Corrections improve the system

Each manual correction is remembered. The system becomes more accurate with use.

Traceability is preserved

Each qualification carries the trace of its source: tax rule, learned pattern, agent decision, or manual choice. You can always understand why a document was qualified in a certain way.