Skip to content

AI adoption for high-consequence work

AI proposes.
Code computes.
You authorise.

Zarene AI lets large enterprises put AI inside the processes they already run — without replacing core systems, and without surrendering any decision that carries financial, legal or regulatory consequence. Models read documents and draft language. They never produce a figure. Every number is computed by deterministic code, checked against the rest of the document, and traced to its source. Nothing is released until a named person approves it.

Request a walkthroughRead the architecture

Statement analysis · FY2024

Awaiting approval

Balance sheet does not foot

Cash and equivalents$6,400,000,000p.41
Trade receivables$11,900,000,000p.41
Investment securities$22,100,000,000p.41
Other assets$7,900,000,000p.42

17/18 checks passed

Approve

What the reviewer actually sees

The product, shown with synthetic data

Statement analysis · FY2024 · Draft 3

Awaiting approval
Balance sheet does not foot — 200,000,000 difference. Assets 48,300,000,000 against liabilities and equity 48,100,000,000. Page 42.
17 of 18 checks passed.
Synthetic balance sheet figures. Select a figure to see its provenance.
Line itemValue
Cash and equivalents
Trade receivables
Investment securities
Other assets
Total assets
Total liabilities
Total equity
Total liabilities and equity

Provenance

Cash and equivalents

Source
Page 41
Extraction confidence
99%
Checked by
Footing check

1 validation failure must be resolved before release.

An analysis awaiting approval. All figures are synthetic. Select any figure to see where it came from and which check consumed it.

Zarene AI by the numbers

0
credentials that can move money or alter records
1 click
from any output to its source
2 of 6
workstreams in production
100%
of released decisions reconstructable from the ledger

Four stages. Two of them are not AI.

The separation is the product.

Wherever a mistake would cost money, breach a rule or require an explanation, the work is done by code that behaves identically every time.

  1. 01AI

    Extract

    A vision-language model reads the document and returns labelled fields, each with a page reference. It transcribes. It does not calculate.

  2. 02Deterministic code

    Validate

    Does the arithmetic hold. Do the derived figures agree with the underlying records they claim to summarise. Are the labels consistent across every table. Failures are raised here, before a human reads a draft.

  3. 03Deterministic code

    Compute

    Every total, ratio and variance is computed in typed code from validated inputs. Same inputs, same outputs, every time, with or without a model provider.

  4. 04AI

    Narrate

    The model writes commentary from figures that already exist. Its output is checked against the computed set; any figure that does not match a computed value is rejected before the draft is shown.

  5. 05Human

    Approve

    A named user approves. The approval records who, when, which draft and which engine version. Until then, nothing leaves the platform.

The error class that source-grounding misses

In a live customer’s working papers, a table carried two correct figures in the wrong columns.

Both numbers were right. Both appeared, verbatim, in the source document. Every assurance method that asks does this figure appear in the source passes that page cleanly. Only a method that recomputes the figure from the data it belongs to catches it.

Synthetic ratio table comparing the figure as printed against the figure as recomputed by the engine, by period.
 Prior periodCurrent period
As printed1.421.18
As computed1.181.42
Figures transposed between periods

This is the case that should concern anyone approving an AI programme. The failure is not hallucination. Nothing was invented, and no confidence score would have flagged it. The document was internally inconsistent, and the check most vendors rely on cannot see that class of error at all.

Two further errors surfaced in the same set: a statement that did not foot, and a line reported as zero against non-zero adjustments elsewhere in the same document. All three were raised by the engine before the draft reached a reviewer.

Deliberately narrow

Six workstreams. Four of them exist today.

The architecture is domain-agnostic. The deployments are not. We build where being wrong is most expensive, prove the workstream against a real operation, and extend from there. Today that means financial operations. The same four stages apply to any process where a document is read, a figure is derived, and someone is accountable for the result.

In production

Trade reconciliation

Counterparty confirmations matched against internal records to a tolerance you set. Breaks are itemised with the fields that disagree.

In production

Financial statement analysis

Filings and management accounts read into a validated figure set, with metrics computed and commentary drafted for review.

In production

Credit assessment

Counterparty financials, rating reports and a transaction's own terms spread into a deterministic scorecard, with an investment recommendation and a repayment-capacity forecast drafted for review.

In development

Payment operations

Instruction review and exception handling. The platform will draft and check; it will not hold credentials that can move money.

In development

Transaction documents

Extraction of terms, dates and obligations from executed documents, each traced to its clause.

In production

Market & sector research

Briefs on a named company or sector, drawn from a verified source universe and constrained search, with every claim checked against a stored, snapshotted passage before it ships.

Four workstreams are in production. Two are in development and are described here as such. We do not show what we have not built.

A short call, not a sales cycle.

Bring one process you would like to see automated and one you would not. Thirty minutes is usually enough to tell whether this is a fit.

Request a walkthrough