Finance & liquidity decisions

Context & background

Treasury and financial steering in the mid-market rarely work off a single truth. Account balances, open items, planned investments and grant commitments sit in separate systems and spreadsheets. From them a liquidity forecast is built that the board, the bank and shareholders are meant to trust.

The decisions that hang on it — draw a credit line, pull an investment forward, time a distribution — are material and partly irreversible. All the more it matters whether the forecast is reproducible: does the same data basis yield the same result tomorrow, and can every figure be backed before the supervisory board or auditor? The EU AI Act demands exactly this traceability for decision-preparing systems.

The problem

Spreadsheet forecasts are hard to trace and rarely reproducible. Nested formulas, manual overrides and grown links mean no one can say for sure why a number is what it is — or whether it stays the same on the next open.

Generative AI is the wrong answer here. A model that 'estimates' numbers produces false precision with no compute path and can invent values freely. In financial steering that is untenable: an invented liquidity figure can trigger a real wrong decision.

Concrete failure points

• Spreadsheet forecasts are hard to trace and rarely reproducible.
• Generative AI invents numbers — untenable in a financial context.
• Scenarios are hard to compare cleanly against each other.
• Uncertainty is masked as false precision instead of disclosed.

The causal chain

From data silo to risky decision runs a clear cause-and-effect chain:

1. 1
   Separate financial data sourcesare reconciled manually in a spreadsheet
2. 2
   Manual, grown spreadsheetsare not reproducible and hard to audit
3. 3
   Non-reproducible numbershide uncertainty as false precision
4. 4
   Falsely precise numberslead to risky allocation decisions

How cNode solves it

cNode ingests the financial sources deterministically and anchors them in the knowledge graph: balances, open items, plan values and commitments are unambiguously linked to periods and documents. Nothing is retyped, nothing freely interpreted.

The forecast itself is a deterministic computation. Via a fixed seed it is reproducible: the same data basis provably yields the same result — the basis of any audit and any comparison over time.

Uncertainty is modelled explicitly and disclosed as a p05/p50/p95 band. Scenarios can be compared cleanly against each other; the ranking stays stable even when assumptions are perturbed, so robustness becomes visible instead of vanishing behind a point score.

Only at the end does the language model cast the computed, evidenced results into a legible assessment. It does not compute and invents nothing — it explains the number the engine has already backed.

The deterministic process

1. Ingest financial sources deterministically and anchor them in the graph.
2. Deterministic forecast — reproducible via a fixed seed.
3. Confidence bands p05/p50/p95 instead of false precision.
4. Scenario simulation with a ranking stable under stress.
5. Source + compute path per figure; the LLM only verbalizes.

The outcome

Board and treasury get a dependable liquidity view in which every figure is backed by source and compute path and uncertainty is honestly disclosed. Allocation decisions are taken faster and documented at the same time — with a rationale that withstands a later audit.

Methodology & verifiability

The methodology cleanly separates computing from explaining. The number arises deterministically and reproducibly; the language model only verbalizes it. An invented number cannot enter the forecast at all — the model's only degree of freedom is the phrasing, not the value.

Reproducibility via a fixed seed, source and compute-path evidence per figure, and an audit trail per data point create alignment with EU AI Act Art. 12/13. Operation either in Frankfurt or on-prem up to air-gapped, data processing under GDPR Art. 28, no training on the data.