Book a discovery call Menu
01

Product · Order-to-cash

Value Lens Private beta

The Margin Leak Finder for Order-to-Cash.

Value Lens reads SAP-like sales, pricing, cost and agreement facts through read-only tools, computes margin in exact integer cents under a versioned policy, and raises evidence-backed cases your finance team can accept or dismiss with a reason. Agents explain. Humans decide. SAP stays the system of record.

Private beta on synthetic SAP-like data. Production SAP connector in development. Detection rules are ours, not customer-approved accounting logic.

Request beta access Request a demo

A 45-minute call with an architect, not a salesperson. Replies within one business day. How we run engagements

Value Lens Findings queue Private beta Synthetic data Policy v1.3Week 38
Case Severity Rule Leakage Coverage Status
VL-0231 HIGH net_loss 12,406.18 0.94 Open

Sources

  • sales_order.get 4500012877
  • pricing.get PR-88213
  • cost.get MAT-10442
  • agreements.get AG-2031

Calculation · Policy v1.3

  • net price 84,120.00 − cost 96,526.18 = −12,406.18
  • rule net_loss → HIGH
  • integer cents, deterministic

Timeline

  • 09:12 facts read (4 tools, read-only)
  • 09:12 margin computed
  • 09:12 rule fired
  • 09:14 agent hypothesis (labelled hypothesis, 3 of 5 authorised reads used): "cost updated after pricing"
  • awaiting reviewer
VL-0228 HIGH excessive_discount 7,930.00 0.88 Open
VL-0225 MEDIUM cost_above_priced_cost 2,114.55 0.91 Reviewed
VL-0219 MEDIUM excessive_discount 1,480.00 0.97 Dismissed · reason: approved promotion

Numbers from deterministic code. The model explains; it never calculates.

02

The question

Where Did We Lose Money This Week That We Should Not Have Lost?

Standard SAP reporting tells you margin by product, customer and period. It does not tell you which orders shipped below cost, which discounts exceeded the agreement, or which cost updates landed after pricing, until the quarter closes and someone reconciles by hand. Value Lens asks the question every week, answers it with source records, and leaves the decision to a person.

Net loss

Orders priced below the cost that applied at the time.

Excessive discount

Discounts beyond the agreement or policy threshold.

Cost drift

Cost changes after pricing that nobody re-priced for.

Rule set in the beta. Thresholds are policy-versioned; severity is MEDIUM or HIGH.

03

How it works

Six Steps, One Audit Trail.

  1. 01

    Read, read-only.

    Facts come in through read-only MCP business tools: sales_order.search, sales_order.get, pricing.get, cost.get, agreements.get. No tool in the product can write.

  2. 02

    Calculate deterministically.

    A margin engine in exact integer-cent arithmetic under a versioned policy (POLICY v1.3 in the beta). The same inputs always produce the same number. The LLM never touches it.

  3. 03

    Detect by rule.

    Net-loss, excessive-discount and cost-drift rules with MEDIUM and HIGH thresholds raise candidate cases.

  4. 04

    Build the evidence.

    Every case carries its source references, a calculation snapshot and a coverage score that says how much of the picture the tools could see.

  5. 05

    Investigate within bounds.

    A host-mediated agent may run a bounded number of authorised reads to propose a hypothesis. It has no financial authority and its hypotheses are labelled as hypotheses.

  6. 06

    Review by a person.

    The findings queue is sorted by severity and leakage. A reviewer marks a case reviewed or dismisses it with a reason. Every action lands on the case timeline.

Six points on a line: 01 read, 02 calculate, 03 detect, 04 evidence (the lens glyph), 05 investigate, 06 human review (bracketed). 01 02 03 04 05 06 READ CALCULATE DETECT EVIDENCE INVESTIGATE HUMAN REVIEW Six points on a line: 01 read, 02 calculate, 03 detect, 04 evidence (the lens glyph), 05 investigate, 06 human review (bracketed). 01 02 03 04 05 06 READ CALCULATE DETECT EVIDENCE INVESTIGATE HUMAN REVIEW
04

Inside the beta

One Case, End to End.

This is the first golden case the beta reproduces exactly, on synthetic SAP-like data under the demo policy. Every number below is produced by deterministic code and can be recomputed from the stored inputs.

Synthetic order line · Demo policy

FactValue
Quantity10 EA
List unit price$100
Allowed discount (demo policy)10%
Recorded discount20%, no override, no eligible agreement
Charged unit price$80
Net goods total$800
Reference unit cost$60

Result

  • $300

    expected margin

  • $200

    actual margin

  • $100

    discount variance, the leakage

  • MEDIUM

    severity

Rules fired: net leakage (R-01) and excessive discount (R-02). Evidence: 4 of 4 checklist items resolved, coverage 100%. Annualised exposure, illustrative 52-week extrapolation: $5,200. The reviewer can mark the case reviewed or dismiss it with a reason; neither action changes a single cent.

Demo policy, not customer-approved accounting logic. Synthetic data. USD and EA only in the beta.

05

Architecture

Four Parts, One Direction of Trust.

Each part can only do what the part before it allows. The browser renders; the backend calculates; the tools read; the adapter is the only code that touches a source.

Browser, backend API, MCP tool server and source adapter in a row, each joined to the next by a line ending in a point. The adapter reads synthetic SAP-like records now; a dashed box marks the SAP connector in development. A hairline beneath the four parts is labelled shared contracts: runtime-validated schemas, integer-cent money. BROWSER BACKEND API MCP TOOL SERVER SOURCE ADAPTER SYNTHETIC SAP-LIKE RECORDS SAP CONNECTOR IN DEVELOPMENT SHARED CONTRACTS: RUNTIME-VALIDATED SCHEMAS, INTEGER-CENT MONEY Browser, backend API, MCP tool server and source adapter in a row, each joined to the next by a line ending in a point. The adapter reads synthetic SAP-like records now; a dashed box marks the SAP connector in development. A hairline beneath the four parts is labelled shared contracts: runtime-validated schemas, integer-cent money. BROWSER BACKEND API MCP TOOL SERVER SOURCE ADAPTER SYNTHETIC SAP-LIKE RECORDS SAP CONNECTOR IN DEVELOPMENT SHARED CONTRACTS: RUNTIME-VALIDATED SCHEMAS, INTEGER-CENT MONEY

Browser

Displays backend results. Holds no policy, no fixtures, no calculation and no model credentials.

Backend

Normalises the retrieved facts, runs the calculation and detection engines, creates cases and persists them with their audit events. All financial fields are owned by code.

MCP tool server

Exposes a small catalogue of read-only business tools over a real MCP transport. Every request and response is schema-validated; unknown keys, oversized pages and out-of-scope dates are rejected.

Source adapter

The only code that reads a source. Synthetic SAP-like records in the beta; a production SAP connector replaces it without moving SAP concepts into the calculation or the screen.

06

Tools

Five Tools. All of Them Read.

  • sales_order.search

    bounded business filters, 5 rows by default, 25 at most

  • sales_order.get

  • pricing.get

    price stages and conditions for one line

  • cost.get

    reference and current unit cost

  • agreements.get

    eligible agreements, or an explicit 'none' the case can rely on

There is no generic query tool, no raw table access and no write tool. Scope, tenant and actor are set by the backend, never by a user or a model.

07

Detection

Nine Rules. Deterministic Severity.

Detection turns a valid calculation into a finding. The beta ships the first two rules end to end and the remaining seven across eleven synthetic scenarios.

Rules

  • R-01 net leakage on the line
  • R-02 excessive discount
  • R-03 to R-09: further price, cost, charge and agreement causes, plus two repeated-pattern rules across lines (customer and material)

Severity (demo policy)

  • Below $50: no finding
  • $50 to $99.99: LOW
  • $100 to $199.99: MEDIUM
  • $200 and above: HIGH

Boundaries are exact to the cent and configured, never model-selected.

Queue order is fixed: severity, then leakage, then transaction date, then case ID. Re-running the same source, policy and rule versions produces the same cases and no duplicate exposure.

08

Evidence

Coverage Is Counted, Never Estimated.

Each case carries an evidence checklist. Coverage is the share of applicable checks whose source records resolved: 4 of 4 is 100%, 3 of 4 is 75%, and a case with nothing applicable shows NOT ASSESSED rather than a number. A missing cost never earns a score, and no narrative can make a gap disappear.

  • 4 of 4 resolved100%
  • 3 of 4 resolved75%
  • nothing applicableNot assessed
09

Design principles

Trust Cannot Be Retrofitted.

PrincipleIn the product
Numbers from code, never from the modelDeterministic engine, integer cents, versioned policy
Agents explain, humans decideBounded investigation, labelled hypotheses, reviewer actions only
Every case traceableSource references and calculation snapshot on every case
Read-only by designNo write tool exists in the catalog
Complete audit trailCase timeline with tools called, policy version, decisions, timestamps
SAP-agnostic domain contractsReplaceable adapter; the production SAP connector is in development
10

Agent

The Agent Explains. It Cannot Decide.

After a case exists, a model may investigate it. What it may read, say and never touch is fixed in code.

What it may read

Up to three additional authorised reads per case, inside the case's customer, material and week, through the same tool catalogue, with provenance on every read.

What it may say

A hypothesis, approved observation and action categories, and references that must resolve to real evidence. Numbers in its text are bound by code to the stored calculation.

What it can never touch

Money, severity, approval, review status or any source record. Invented records and contradictions are rejected. If the model is unavailable, the case stays complete and useful.

Retrieved documents are treated as untrusted text. Model credentials stay on the server and never appear in prompts, errors or audit records.

11

Audit

Every Case Can Be Rebuilt from Its Trail.

Each run records an attempt and an outcome for every tool call, the normalised inputs, the calculation snapshot with its policy and rule versions, each rule evaluation, the coverage assessment, each model request and result, and each reviewer action. Run, case, tool and investigation IDs are correlated. The case timeline shows what was read, calculated, detected, explained and reviewed, and a reconstruction check recomputes the stored cents from the stored inputs.

Review actions

Mark reviewed, or dismiss with a reason. The action and its audit event commit together; if either fails, nothing changes.

What a review cannot do

Alter a source, a calculation, a rule result or an exposure. Reviews survive restarts and reruns of the same version.

12

Also from SVLS LABS

Building the integrations around it?

See SAP Intelligence Suite
13

Scope

In the private beta (phase A)

  • Synthetic SAP-like order-to-cash data set
  • Read-only tool catalog (search, get, pricing, cost, agreements)
  • Deterministic margin engine, policy v1.x
  • Net-loss, excessive-discount and cost-drift rules with MEDIUM/HIGH thresholds
  • Evidence-backed cases with coverage scores
  • Bounded agent investigation with labelled hypotheses
  • Findings queue, review actions, case timeline

We update this list when the product changes, and we would rather show you a shorter list that is true.

Phase A · 15 milestones · 47 tasks

  1. M0Bootstrap
  2. M1Domain contracts
  3. M2Synthetic SAP-like records
  4. M3Source adapter
  5. M4Read-only MCP tools
  6. M5Deterministic calculation engine
  7. M6Detection and rules
  8. M7Case generation
  9. M8First working screen
  10. M9Evidence and calculation detail
  11. M10Agent investigation
  12. M11Audit trail completion
  13. M12Remaining scenarios
  14. M13Testing and hardening
  15. M14Demo readiness

The first usable screen arrives at M8, the investigator at M10, and the complete eleven-scenario demonstration at M12. After Phase A: the production SAP connector, customer accounting policy, production identity and hosting, and write-side controls.

Private beta on synthetic SAP-like data. Production SAP connector in development.

14

Guardrails

What We Refuse to Build into the Beta.

  • No real SAP integration or customer data in Phase A
  • No SAP writes, postings or execution of recommendations
  • No financial arithmetic, thresholds, severity or coverage inside the model
  • No generic query tools or raw table access
  • No hidden calculations: inputs, signed results and versions stay inspectable
  • No invented evidence, approvals or numeric claims
  • No multi-agent choreography or plugin platform
  • No production authentication or hosting decision yet
15

Stack

  • TypeScript monorepo
  • Next.js and React
  • Node.js API
  • MCP server on the TypeScript MCP SDK
  • Runtime-validated contracts
  • Integer-cent money
  • SQLite in the beta
  • Vitest and Playwright
  • Provider-neutral model gateway (local or hosted model)
16

Private beta

Request Beta Access.

The beta is a guided walkthrough on synthetic order-to-cash data, followed by an early-access seat. You will receive the Value Lens beta brief first: what the beta covers, what data it needs, and what it will not do yet.

Private beta on synthetic SAP-like data. Production SAP connector in development.

A 45-minute call with an architect, not a salesperson. Replies within one business day. How we run engagements

Your role

We use your details only to reply to you. See Privacy.

17

Questions

Is this connected to my SAP system?

Not yet. The beta runs on synthetic SAP-like data; the production SAP connector is in development.

Can it change anything in SAP?

No. There is no write tool in the product, and none is planned.

Where do the numbers come from?

Deterministic code, integer cents, versioned policy. Never from the model.

Are the detection rules our accounting policy?

No. They are our rules with configurable thresholds. Customer-specific policy configuration is on the roadmap.

Who reviews a case?

Your finance reviewer. Value Lens raises and explains; it does not decide.

What does the beta cost?

Beta participation terms are agreed per organisation. Participation is priced per organisation; no prices are published on this site.

Why integer cents?

Because $0.01 matters in a margin bridge. Money crosses every boundary as an integer string in minor units, and every expression uses exact integer operations. No floating point, no rounding surprises.

What happens when data is missing?

The case is blocked and says so. A missing cost, a total that does not reconcile or a failed read never becomes a zero-loss result or an empty success.

Can we bring our own accounting policy?

That is the point of the versioned policy. The beta ships a demo policy; a customer policy is configured, versioned and recorded on every calculation it touches.

Which SAP editions will the connector support?

The adapter boundary is edition-neutral. The production connector is in development; we will confirm editions with beta participants.

18

Next step

Agents Explain. Humans Decide. SAP Stays the System of Record.

Request beta access Book a discovery call

Private beta on synthetic SAP-like data. Production SAP connector in development.

A 45-minute call with an architect, not a salesperson. Replies within one business day.

How we run engagements