RGI Source-to-Pay
Earned Autonomy for Source-to-Pay
RGI Source-to-Pay is a decision copilot for procurement: it scores every requisition, PO and invoice against the judgment your firm has already proven — not a rulebook — then decides it or routes it, reasoning attached.
Not another drafting agent. Suite copilots, agentic-AP tools and process miners draft for your approval and stop; RGI earns the right to decide, class by class, and shows the CFO the ledger.
Built by Arindam Banerji, PhD — architect of the Accenture Insights Platform; advised Reckitt, BP and Sumitomo on supply chain.
Enrich→
Decide / abstain→
Act→
Verify & write-back→
Evolve
one loop · the loop is the product
47
good invoices your rules rejected this quarter — each one doing its job
the problem
5% → 41%
of eligible decisions auto-decided, each step earned
modeled trajectory
30 days
read-only — a signed ledger before you trust it
how it starts
from $60K
band-priced on decisions under management — not per-seat, not a cut of savings
pricing
Roadmap-forward. No production customers yet — figures are modeled and labeled throughout.
The problem
Your AP team rejected 47 good invoices this quarter. Each one was a rule doing its job.
Procurement is a judgment problem wearing an optimization costume. Six pressures break the rule-and-queue status quo:
P1 · judgment
Rules are brittle
- Thresholds fire on a number, not on why it moved
- So they reject good invoices and wave maverick spend through
- The judgment that'd get it right lives in a few experts' heads — and leaves when they do
P2 · non-stationary
The ground keeps moving
- Commodity, tariffs, FX, supplier health, lead times shift weekly
- Rules tuned in January are wrong by April
- A model trained once decays in silence
P3 · non-linear
The hard calls interact
- One PO carries price, demand, supplier, geo, alternative-availability and spend-pattern risk at once
- No linear rule draws that boundary
- Thresholds over- and under-fire on the calls that matter
P4 · scattered
Systems don't talk about cause
- Invoice in the ERP, clause in the contract repo, price move in a commodity feed, process in the mining tool
- Connectors move bytes; nothing links them into why
P5 · every call is a dollar
Wrong is asymmetric
- Often the largest controllable spend line
- Wrong auto-approval = a write-off; wrong rejection = a late supplier + maverick spend
- At volume, getting it wrong breaks the firm
P6 · the agent gap
Today's agents draft, never graduate
- Agentic-ERP fleets monitor, summarize, route for approval — and stop
- Nothing measures earned authority, shows the CFO a ledger, or pauses when accuracy slips
How it decides · one invoice, this morning
How it decides one invoice — judgment where a rule only sees a number
Here is a single invoice as it hits the queue — and the two ways it gets handled. The copilot doesn't apply a bigger rulebook; it scores the decision against the shape of every similar call your firm has verified.
Invoice #4471 · $214,900 · 5.2% over the PO · Tier-1 supplier, 11-yr history
A threshold rule
5.2% > 5.0% policy line
→ route to manual review
REJECT
The 47th good supplier paid late this quarter.
Judgment — the six factors, weighed
copper +4.8% on the 30-day index
contract §7.3 allows pass-through to 110%
5.2% ≤ 1.10 × 4.8% = 5.28%
ACCEPT · 0.91
Supplier paid on time; the clause is cited and the call is reversible.
the decision — geometry, not a reward
P(a | f, c) = softmax( −‖ f − μc ‖² / τ )
The action is the nearest learned prototype (μ) to this decision's factor vector (f) — inspectable, reversible, and improved by every verified outcome. Not a rule that fired, and not a black-box reward.
Earned autonomy
How it earns the authority to decide on its own
This is the difference in one picture. An agent that drafts for approval never graduates. RGI grows the share of spend it decides on its own, one risk-class at a time, each step earned against verified outcomes and held by a conservation gate.
Coverage grows — each step earned
Auto-decided share of eligible decisions, accuracy held above the floor by the conservation gate. Modeled trajectory.
What it will never do
Autonomy is earned per class and bounded. The dial is always visible, and the hard limits never move:
Routine reorder, on-contract
auto once the class is proven
Price anomaly
propose-first until earned
Sole-source / emergency
always escalates to a human
Restricted / sanctioned party
always holds — never auto
Illustrative: "41% auto-approved, here's the proof — 450 decisions, 91% blended, above the floor for eight weeks, and it pauses itself if accuracy drops." No vendor we've seen publishes that ledger.
The autonomy contract — per class, you set it
Accuracy floor
Auto-decides only while that class's audit-verified accuracy holds at or above the floor you set (default 90%).
Dollar ceiling
Auto only under a per-decision cap (default $50K); above it, policy escalates — human approval, then controller, then dual sign-off.
Pause & rollback
On a slip it reverts the class to propose-only and alerts the owner; every auto-decision is a reversible, signed ERP write in a 7-year ledger.
the full autonomy contract — floor, ceiling, verified-outcome lag, pause, rollback, record
Accuracy floor
The headline 91% is blended across auto + human-routed decisions; the auto-only slice is held to the stricter floor you set (default 90%).
Dollar ceiling
Per-decision cap (default $50K) and a per-day aggregate cap. Above it — >$50K → human approval, >$250K → controller, >$1M → dual sign-off.
Verified outcome — and its lag
Learns only from outcomes confirmed in SAP / the contract / the commodity index — price & 3-way-match in minutes–hours, goods-receipt & payment in days — plus a continuous 30% audit sample. Never from its own prediction.
What a pause does
Reverts the class to propose-only, alerts its named owner, flags recent auto-decisions for review. It re-earns autonomy only after verified accuracy climbs back above the floor on a fresh window.
Rollback
Every auto-decision is a reversible, signed write to the ERP with the prior state kept — one click to reverse, and the reversal is itself logged.
On the record
Every decision, pause and reversal lands in a 7-year signed ledger (SOX / EU-AI-Act-shaped), under the oversight chain you approved — analyst → category manager → controller.
End to end · one incident
One commodity shock — the rule rejects on the number; RGI reads the contract and clears it
The scenario: mid-stream, copper jumps on the 30-day index. A 5.2%-over-PO invoice lands from a tier-1 supplier with 11 years of history. Here is the same incident through RGI's loop — and the same morning without it.
The same morning — without
- ~800 flagged overnight; triaged by hand across SAP → Celonis → PDFs → commodity feed.
- The 5% rule auto-rejects the invoice — the 47th this quarter; §7.3 unseen, a good supplier paid late.
- A container cleared into a full warehouse; a single-source credit slip × a sanction nobody connects.
- 5pm: queue still ~800, maverick spend leaked, no ledger for the CFO.
With RGI Source-to-Pay
- Overnight work pre-scored; ~40% auto-decided, the rest a ranked exception queue with reasoning.
- Effort follows stakes — one read on the $500 bill, three on the $50K capex.
- Cross-graph links the credit slip × the regional event × the >70% dependency, flagging it early.
- 5pm: queue shrank, one class retired to zero, the CFO gets the signed ledger.
Illustrative incident on modeled values — the ledger figures are the modeled earned-autonomy trajectory, not a customer result. Cross-graph discovery and effort-follows-stakes are roadmap pillars.
What it costs today
What getting these calls wrong costs — the quiet leaks dwarf the headline blowups
The everyday leaks cost more than the headline disasters — and both trace to the same missing judgment. RGI was at none of these firms; the pairing shows which pillars are built for each failure mode.
| Failure mode | Cost | What goes wrong | Pillars built for it |
|---|---|---|---|
| Maverick spend | ~$18M per $1B purchases/yr | Off-contract buying at volume, unenforced (APQC). | Situation Analysis · Process Action · Earned Autonomy |
| Duplicate & erroneous payments | 0.1–0.5% of AP spend, every year | The same invoice paid twice, or an off-contract price waved through — ≈$5–25M a year on $5B of spend. | Situation Analysis · Verified Reasoning · Earned Autonomy |
| Invoice & vendor-payment fraud (BEC) | $2.77B reported 2024 (FBI IC3) | Payment redirected to a spoofed supplier or a bogus bank-detail change — authorized because the change looked routine. | Situation Analysis · Verified Reasoning · Process Action |
| Sanctions at the transaction (ZTE, 2017) | $892M penalty | Procurement to restricted entities — compliance failed at the transaction. | Situation Analysis · Process Action · Verified Reasoning |
And the headline blowups trace to the same gaps: the 2021 chip shortage (~$210B lost auto sales — single-source exposure invisible until production halted), Target Canada (~$7B, rushed master data), Nike/i2 (~$100M, bad demand signals). Industry figures: FBI IC3 2024 (BEC $2.77B) · ACFE 2024 Report to the Nations · AP duplicate-payment audits (0.1–0.5% of spend).
Where it plugs in
Where it fits — under your existing P2P stack, requisition to payment, no rip-and-replace
Named roles across the arc — one judgment engine, not seven products. You start at the price-variance / invoice seam and extend as it earns coverage.
Sourcing / RFx
Sourcing Optimizer
long cycles, compliance gaps
Requisition
Requisition Advisor
approval delays, maverick spend
Purchase order
PO Compliance
unauthorized changes
PO → invoice seam
Price-Variance Guardian
◆ flagship — hidden price drift
Goods received / QC
GR & Quality
manual GR, siloed QA
Invoice / 3-way match
Invoice Concierge
◆ flagship — working capital frozen
Payment
Payment & Supplier
duplicate/late pay, DSO
RGI Source-to-Pay — the judgment layer
reads your systems · scores every decision against proven judgment · earns authority, logs an auditable ledger
↑ reads the data↓ writes the decision back
your stack — unchanged
SAPOracleAribaCoupaCelonisEDW · Snowflake/Redshift/Databricks/Fabriccontract repocommodity feeds
software-only · no infra licensing
sits alongside · no rip-and-replace
doesn't build your EDW — routes to it
Quick wins — the Shadow's first artifacts
What you get in the first weeks — before you grant any authority
Each is a read-only output of the 30-Day Shadow, on your own data: the first slice of the economics above, shown as a tangible artifact and mapped to the sub-process and dataset it comes from. This is what the signed ledger is made of. The two live wins are what the 30-Day Shadow produces now; the four near are the roadmap, as coverage extends.
| Sub-process | Dataset | What the Shadow produces | |
|---|---|---|---|
| Invoice / 3-way match | AP invoice lines × PO-match | A ledger of invoices your rules rejected but were right to pay (the §7.3 line) — and the count of suppliers you'd have paid late. | live |
| PO → invoice seam | PO lines × contract × commodity | A price-drift tally — line items above the benchmark that slipped through; the recoverable-leakage number. | near |
| Payment / AP | Payment run × vendor remittance | A duplicate-and-overpayment list with invoice IDs — direct cash to claw back. | near |
| Requisition | PR data × budget × catalog | A day-one triaged queue beside what your team did — the agreement rate and the calls it would've caught. | live |
| Sourcing / supplier risk | Supplier master × credit × geo | An at-risk list where a credit slip × a regional event × a >70% single-source dependency line up. | near |
| Data quality (DataOps on procurement data) | Supplier · SKU · contract · inventory | A data-estate scan: duplicate vendors, broken PO↔contract links, phantom inventory — the faults that corrupt decisions, surfaced first. | near |
The data-quality scan is the same engine as RGI DataOps, run on procurement data — the answer to "our ERP data is too messy for this."
The product
The whole product — twelve pillars on one closed loop
The loop is the product, and it is what the agent fleets don't have: every verified decision reshapes the next one. Read the stations in the order the loop runs them; the twelve pillars hang off them.
live running in the demo, not at a customer · near building now · roadmap planned
Enrich · sense the situation
1 · Procurement Context Graphlive
ERP, contracts, commodity indices, supplier history and process-mining signals in one causal graph the copilot reasons on from day one — a reasoning substrate, not RAG.
2 · Situation Analysislive
Each PO read into a factor vector, classified into one of six situations, scored across five actions — and it abstains when it hasn't earned the right.
Act · plan and execute
3 · Planning & Optimizationnear
Plans the multi-step action and calls proven solvers where the problem is combinatorial; a feasibility check runs first. Judgment chooses the plan; the solver optimizes within it.
4 · Process Actionlive
Posts or holds the PO, drafts the debit memo, notifies the supplier — under an approval gate, verified in source, logged as immutable evidence.
Verify & write-back · the compounding step
5 · Judgment Memorylive
Every verified outcome reshapes one prototype — provenance back to the PO that taught it, a quality score, one-click rollback. Your best buyer's judgment, kept and sharpening after they leave.
6 · Self-Evaluationlive
Scores its own calls against verified outcomes — never an LLM judge or curated ground truth, which fail here. The eval that lets autonomy grow.
Evolve & wrap · tune, discover, govern
7 · Adaptive Investigationlive
Learns how much to investigate per decision — one read on the $500, three on the $50K — and recovers when conditions shift or the data is gamed. The geometry still makes every call.
8 · Tech-Process Fusionnear · the wedge
A procurement process becomes an editable object: propose an edit, verify the KPI moved, learn which edits work. This is the causal connectivity.
9 · Cross-Graph Discoverynear
Attending the firm's graphs to each other surfaces risks nobody queried — a supplier credit slip × a regional event × a single-source dependency, before it bites.
10 · Earned Autonomylive
Grows the auto-decided share one class at a time, gated and logged in a CFO-ready ledger — and pauses itself when accuracy slips.
11 · Verified Reasoninglive
Every decision inspectable and reversible — the prototype, the factors, the verified outcome it learned from. The decision is the explanation an auditor asks for.
12 · The 30-Day Shadowlive · start here
Read-only on one contained family for 30 days → a signed ledger of what it would have decided, credited if you proceed.
See it — the demo spine
Nine moments at a procurement desk — the call a rule makes vs. the call RGI makes
live real engine runs, demo data · plant seeded on your data · roadmap labeled, future. These are illustrative scenes, not customer quotes — behind every one, a single update rule does the learning.
how a verified outcome moves the geometry
μc,a ← μc,a + η ( f − μc,a ) · ηconfirm = 0.05, ηoverride = 0.01
A confirmed decision nudges the prototype toward the new evidence; an override nudges 5× less — so one bad or adversarial call can't swing the firm's judgment. This is the step repeated across all nine scenes.
The rule that kept saying no
"A rule fires on a number; judgment reasons about why — and you stop paying good suppliers late."
Situation · Verifiedlive
Before the first invoice
"Most vendors need six months of your data; this was useful before the first invoice arrived."
Judgment Memorylive
When paying more is right
"Every optimizer would clear the container; sometimes the right answer is to pay more."
Situation · Planningplant
Earned autonomy, week over week
"41% auto-approved, here's the proof — no vendor we've seen publishes that ledger."
Earned Autonomylive
The queue that shrank
"Decision ten-thousand isn't handled like decision one — the queue shrinks, classes retire."
Earned · Adaptiveplant
The rule it kept manual
"It retires the rules it earned and keeps the ones it didn't — restraint you can audit."
Earned · Verifiedplant
$50K needs more than $500
"One confidence signal, three owners — finance, security, audit — and nobody had to lose."
Situation · Earnedplant
One read on easy, three on hard
"Effort follows stakes — and the safety gate never moved."
Adaptive Investigationroadmap
The version we froze
"The distance between the live line and the frozen twin is the thing you're buying."
Self-Evaluationplant
The economics
Start with ~$22M in year one (modeled) — then it compounds
Modeled on a $7.5B manufacturer (range $5–10B), expected band. Value lands by process, on the sub-dataset it runs on — realized process-by-process as coverage extends. The table is the floor, not the ceiling.
How the ~$22M is built — a number you can challenge, not a headline
• ≈ $4B addressable procurement spend — not the $7.5B revenue (direct + indirect the copilot touches).
• ≈ 0.55% captured in year one → ~$22M — deliberately sub-1% of addressable spend.
• Maturing toward ~$109M (~2.7% of spend) as coverage compounds — the prize, not year one.
• Year one is built from the entry seams (price-variance + invoice); the mature pool adds the later pillars as coverage extends.
Where the value is
Mature recurring pool by process, $7.5B manufacturer (expected). Year-one realized ≈ $22M on the entry process.
the compounding law — measured in simulation
I(n, t) ∝ n2.15 · t1.387
Value grows super-linearly in the number of verified decisions (n) and in time (t) — because each decision sharpens the next. A one-time project is linear; this curve bends up. That is why year-one is the floor.
Distributor ($5B)
Mature pool ~$63M/yr; ~$13M in year one; one-time working-capital release ~$96M (modeled). Mix weights transaction volume, duplicate-pay and inventory-carry higher.
Modeled from public procurement benchmarks · conservative · labeled · separate from pricing.
What it costs
Priced on decisions under management — three tiers
Not per-seat (which punishes automating the work away), not a share of savings. Software-only, no infrastructure licensing. The 30-day shadow is credited to year one.
S · Guardian
$60K / yr
- ≤ 150K decisions/yr · ≤ $1B spend
- 1 class · earned autonomy on 1, $25K ceiling
- Lands under a single VP signature — one seam
M · Control Tower
$220K / yr
- ≤ 750K decisions/yr · ≤ $5B spend
- 4–5 classes · earned autonomy on up to 3, $50K ceiling
- The $5–10B manufacturer / distributor core
L · Enterprise
Custom
- Uncapped · $5B+ · multi-entity
- All + custom classes · autonomy across all eligible
- Multi-ERP · option to add a second copilot
Band-step — exceed your decisions-under-management band and you move to the next tier, no per-decision metering. Shadow ~$20K, credited. Figures indicative; set per engagement.
The moat
Why competitors can't close the gap
The layer that learns your firm's judgment and earns the authority to use it. Hold that against the field: everyone now ships "autonomous," but the autonomy is scoped and static — it doesn't compound. And the ones that say they "learn over time" retrain in batches or tune retrieval — the decision surface stays frozen between releases; RGI reshapes it on every verified outcome.
The gap is what you're buying
Live system vs a frozen day-one twin, on the firm's own decisions.
the moat math — measured in simulation
cumulative advantage c = 1.387 · discovery pairs D(n) = n(n−1)/2 ∝ n2.15
A single-loop rival returns the same value each cycle (d²/dt² ≈ 0). RGI accelerates (c > 1), and each copilot added multiplies the discovery surface rather than adding to it.
Why a competitor can't catch up
- Firm-specific — learned from your verified outcomes; can't be forked or bought.
- Temporally irreversible — every decision is embedded; a rival starting today starts far behind.
- Model-independent — survives any LLM swap; the advantage is in the geometry, not the weights.
- Single-loop rivals can't match the trajectory — a drafting agent never compounds.
Who else is in the market
Everyone clusters bottom-left — automating the task, drafting for approval. The top-right — compounding judgment, earning audited authority — is where RGI sits; we haven't seen another there.
Where to start
How to start — a 30-day read-only shadow on your own data
Not a six-month implementation and not a pilot that asks for trust up front. Read-only for 30 days, then it earns authority the way a new hire does — by being right, visibly, on the calls you already make.
Shadow — it watches, decides nothing
Read-only on one contained PR or invoice family. It scores every decision it sees and records what it would have done — no write authority, no risk, nothing to rip out.
The ledger — proof on your numbers
A signed ledger of what it would have decided: the caught invoices, the drift, the duplicates, the at-risk suppliers — your agreement rate and the dollars behind it. Credited to the engagement if you proceed.
Earned authority — one class at a time
Go live proposing; it earns write authority per situation-class, each step gated and logged — and pauses itself if accuracy slips. You watch the share grow; you never hand it over blind.
The 30-day shadow runs ~$20K, credited to the engagement if you proceed. What it asks of IT: read-only extracts of your PO / invoice / contract tables · runs in your VPC or ours · a few hours of one analyst — read-only, no ERP write credentials.
See whether your judgment compounds — before anyone asks you to believe it does.
Built by Arindam Banerji, PhD — 20+ years in enterprise AI/data, specializing in supply chain. Advised Reckitt, BP and Sumitomo on AI & supply-chain strategy; led supply-chain risk for a networking major and the DoD. Built and commercialized the Accenture Insights Platform; co-created SOA / e-Speak at HP Labs (Smithsonian). 9 US patents.
On the enterprise supply-chain visibility program Banerji led to market: "Infosys and Microsoft poised to capture the supply-visibility and collaboration opportunity" — AMR Research
dakshineshwari.net
Arindam Banerji, PhD · banerji.arindam@gmail.com