RGI ยท Compounding Intelligence

RGI Source-to-Pay

Automation Assurance for Procure-to-Pay

Six things break in source-to-pay. Six mechanisms fix them. Your three-way match can't read a contract clause. Commodity shocks cascade from requisition to invoice. Maverick spend leaks through the gaps between SAP and your compliance tool. Your touchless rate hasn't moved in three years. No vendor publishes a verified-outcome ledger. And supplier concentration risk hides until it's a headline. Four views โ€” each one solves a different pair.

RGI-S2P Engine Loop
Six things that break โ€” how each one gets fixed
Problems
Copper spikes → 800 exceptions overnight
Credit downgrade + regional disruption — convergence unseen
Mechanisms
Multi-factor re-scoring against live commodity index
Cross-graph: supplier risk ↔ regional disruption
Commodity Shock
Response
$2.4M/yr
Problems
Touchless rate stuck at 32% — year after year
CFO asks for proof it's safe — no ledger, no rollback record
Mechanisms
Per-category accuracy → threshold → auto-approve unlock
Signed ledger + accuracy floor + cost-asymmetry constraint
Automation
Assurance
$3.1M/yr
Problems
Off-contract → OFAC near-match undetected
Sanctions gaps between systems — no cross-check
Mechanisms
Cross-system discovery chain (SAP + contract + sanctions)
Enrichment: verified outcomes add graph edges
Spend
Governance
$4.2M/yr
Problems
Invoice rejected wrongly — copper moved, contract allows it
Good supplier paid late — the 47th wrong rejection this quarter
Mechanisms
Six-factor risk profile reads contract + commodity + supplier
Scored disposition replaces threshold rules
Touchless Rate
Trajectory
$1.8M/yr

Modeled impact at $10B manufacturing revenue. Figures illustrative.

1

Automation Assurance

  • Your touchless rate is stuck at ~30% (Ardent 2025: 32.6%, barely changed from 2023) โ€” the boundary calls resist automation because the right answer depends on context no rule has
  • RGI breaks the ceiling: automation assurance โ€” starts at zero, grows class by class as verified accuracy proves each boundary call
  • The third line (dotted orange) is RGI without the accuracy floor โ€” it learns faster but chases noise. The gate is why the solid line holds
  • And the proof for the CFO: every auto-decision in a signed ledger โ€” accuracy floor, dollar ceiling, pause-and-rollback
decisions to simulate
outcome verification noise
Accuracy & touchless PO coverage โ€” over the decision stream
RGI accuracy (gated) RGI ungated โ€” no accuracy floor current workflow (rules-plus-agent) touchless PO coverage

SNAPSHOT From the running S2P copilot โ€” automation assurance โ€” accuracy floor & signed ledger โ–ถ
S2P Performance tab โ€” learning trajectory and accuracy floor
S2P Copilot ยท Performance tab โ€” accuracy floor GREEN, 80% accuracy, earned authority controls, signed ledger
accuracy floorGREEN
ฮฑ (learning rate)1.00
verified decisions292
correct decisions234
accuracy (q)80.14%
baseline75.0%
cost asymmetry5.0
headroom233.92
total scored429
auto-approve rate36.36%
savings estimate$522,548
annual target$680,000

Snapshot from the development copilot, early in learning โ€” conservation/status and performance/summary โ€” the same gate the simulation above models. Signal 234.00 exceeds ฮธ_min 0.0806. Status GREEN: the copilot is earning autonomy. 5 exception categories active.

The nine-station loop โ€” lit = what this demo exercises

Context pulled per PO โ€” requisition ยท contract ยท goods receipt ยท invoice matching ยท supplier master
pulled from your systems
โ†’
normalized to the 6 factors the scorer reads

This demo runs on synthetic PO data. In production, the same scorer reads live SAP (EKPO/EBAN/MIGO), Celonis P2P process-mining signals, and commodity feeds โ€” spanning requisition through three-way match and payment. No model retraining required.

PO disposition trace โ€” $ amount ยท spend category ยท static vs RGI routing ยท confidence ยท touchless or buyer
Agentic workflow
RGI
confidence (margin to the 2nd-best routing):highmediumlow

Click any PO row to see the SAP/Celonis context and the confidence margin behind that routing decision.

2

Commodity Shock Response

  • Copper spikes 25% โ€” and it doesn't just break three-way match. It triggers off-contract panic buying, pulls in alternate suppliers who may be on a restricted-party list, and surfaces invoice exceptions weeks later
  • The three-way match is blind to why the number moved โ€” rule sees 5.2% > 5.0%, rejects. But copper moved +4.8%, and contract ยง7.3 allows pass-through to 110%
  • Watch RGI's Situation Analyzer read six factors where a rule reads one โ€” and the accuracy floor pause affected categories before a bad PO clears
try a few โ€” recovery varies
Accuracy through the shock โ€” with commodity price overlay
RGI accuracy current workflow (rules-plus-agent) touchless coverage commodity price

SNAPSHOT From the running S2P copilot โ€” invoice #4471 six-factor fingerprint โ–ถ
S2P Insight tab โ€” 8-factor invoice fingerprint
S2P Copilot ยท Insight tab โ€” invoice #4471 factor fingerprint, ยง7.3 pass-through, six-factor scoring
S2P-INV-0001 contract_gap dominant: payment_terms_impact
match_status
0.698
amount_variance
0.142
duplicate_score
0.065
supplier_history
0.164
payment_terms
0.859
commodity_index
0.194
tax_regulatory
0.586
environmental
0.450

Snapshot from the development copilot โ€” invoice fingerprint S2P-INV-0001 โ€” the 8-factor model the copilot uses to classify exceptions. This invoice's dominant signal is payment_terms_impact at 0.859, categorized as contract_gap. The same factor weighting drives the simulation's PO routing above.

Accuracy gate โ€” spend categories paused & volatile updates damped
classes paused damped updates (high-displacement)

PO routing profiles โ€” how each shifted after the shock
Autonomy status โ€” which spend classes recovered
3

Touchless Rate Trajectory

  • The CPO metric: touchless order rate โ€” POs clearing requisition, three-way match, and payment without a buyer touch
  • This is the financial translation of the first two views โ€” accuracy becomes spend, the accuracy floor becomes cost avoidance
  • Industry benchmark: 32.6% average, 49.2% best-in-class (Ardent 2025). Teams above 60% show 3.5ร— higher AP productivity (Hackett 2025)
  • The escalated/held band = maverick-spend intercepts and compliance holds โ€” what a rules-plus-agent workflow leaves on the table
match view 2 to compare
optional โ€” rescales outcomes
for a spend number
Touchless share โ€” how the PO routing split evolves over the stream
touchless (auto-cleared) buyer-reviewed escalated / held (compliance + maverick intercepts)
SNAPSHOT From the running S2P copilot โ€” exception queue snapshot โ–ถ
S2P Exception Triage tab โ€” invoice queue with novelty detection
S2P Copilot ยท Exception Triage โ€” 93 queued invoices, novelty spike alert, scoring and situation context
engine versionv0.7.23
sourcegraph
total exceptions82
scored50
Exception categories in queue
InvoiceCategorySupplierAmount
S2P-INV-0001contract_gapSUP-017$12,450
S2P-INV-0002price_varianceSUP-003$8,920
S2P-INV-0003quantity_mismatchSUP-041$23,100
S2P-INV-0004duplicate_riskSUP-008$5,670
S2P-INV-0005format_complianceSUP-022$31,400

Snapshot from the development copilot โ€” exception queue preview โ€” first 5 of 82 exceptions. 5 exception categories active: price_variance, quantity_mismatch, duplicate_risk, contract_gap, format_compliance. The touchless rate trajectory above models how auto-clearing these grows over time.

4

Spend Governance & Compliance

  • Invoice in SAP, clause in contract repo, sanctions in compliance tool โ€” nobody connects shock โ†’ off-contract PO โ†’ restricted-party alternate
  • Maverick spend: ~$25M per $1B for bottom performers (APQC). BEC fraud: $2.8B in 2024 (FBI IC3). Supplier concentration: $210B in lost auto revenue from invisible single-source exposure (AlixPartners 2021)
  • RGI's cross-graph attention links facts across systems โ€” and enrichment compounds: the 1,000th investigation connects dots the 1st couldn't see
  • Interception funnel: how much risky spend each engine catches, how much leaks to invoice exceptions
try different seeds
Interception funnel โ€” risky spend caught vs leaked, by risk type
RGI intercepted agentic workflow caught leaked (neither caught)
SNAPSHOT From the running S2P copilot โ€” cross-system compliance evidence โ–ถ
S2P Evidence tab โ€” audit trail and compliance posture
S2P Copilot ยท Evidence tab โ€” cross-system chain, decision audit trail, compliance screening
total reviewed50
compliant0 (0.0%)
flagged50 (100%)
savings estimate$522,548
annual target$680,000
target attainment76.8%
Flagged invoices โ€” recommended actions
InvoiceCategoryAction
S2P-INV-0001contract_gapescalate to procurement
S2P-INV-0002price_variancerefer to commodity desk
S2P-INV-0003quantity_mismatchhold for GR reconciliation
S2P-INV-0004duplicate_riskauto-block โ€” duplicate score >0.9
S2P-INV-0005format_compliancereturn to supplier

Snapshot from the development copilot, early in learning โ€” compliance evidence and performance/summary โ€” 50/50 invoices flagged is expected: the copilot is in early learning (ฮฑ=1.0, accuracy floor GREEN) and errs toward flagging. As accuracy compounds, the auto-approve rate (currently 36.36%) climbs and the compliance funnel above narrows. The interception funnel in this sub-demo models the same mechanism at scale.

How does the judgment memory inside this engine compare to agentic (RAG/k-NN) memory architectures? โ†’ Side-by-side comparison  ยท  โ†’ Full dashboard

30-second summary โ€” plays after you've seen the views

Automation assurance is the flagship of RGI Source-to-Pay, part of the Compounding Intelligence platform. The copilot sits across requisition, PO disposition, goods receipt, invoice matching, and payment โ€” growing the share of PO spend it processes straight-through under a dollar ceiling and a monitored accuracy floor. When commodity prices shift, it pauses the affected spend categories back to your buyers until it relearns the new price landscape. Off-contract (maverick) spend is intercepted before it becomes an invoice exception; restricted-party and sanctions screening runs on every PO โ€” the two costliest P2P failure modes that agentic workflows treat as separate tool calls.

dakshineshwari.net
Arindam Banerji, PhD
banerji.arindam@gmail.com
Simulation on synthetic, seeded P2P decisions (requisition โ†’ PO โ†’ GR โ†’ invoice match โ†’ payment) โ€” a mechanism demo of how the earned-autonomy gate behaves across the procure-to-pay cycle. Not customer data and not a measured result. No production customers.  ยท  demo build v48 ยท October 2026