How it works
An engagement, end to end.
The system does the volume work. A person makes every judgment that carries professional responsibility — and the file records both halves.
Five gated checkpoints
Each gate is a recorded human judgment, snapshotted against the exact data state it was made on.
Inside each stage
Expand any stage for detail. Nothing here is hidden behind a form.
1 IntakeGL, trial balance, subledgers, documents, ERP connectors
Arbitrary column mapping means you do not reshape your client's export to suit us. Documents arrive through a client portal and are read where they land, rather than being re-keyed into a template first.
- General ledger, trial balance and subledger ingest
- PBC portal for client-provided documents
- Direct ERP connector import
2 Detection & riskEngines write findings routed for review — never conclusions
Detection runs across the population rather than a sample. Every engine writes a finding routed for human review; none writes a conclusion, and none advances the engagement on its own authority.
- Journal-entry testing · Benford · duplicate and misclassification detection
- Segregation of duties · graph and network fraud patterns
- ML anomaly detection · analytical review · compound risk reasoning
3 File buildYour methodology, versioned and pinned per engagement
Firms author their own workpaper and field universes. An engagement pins to a methodology version, so a change made at national-office level mid-season cannot mutate a file that has already been concluded.
4 Evidence & provingControls reperformed over 100% of the population
A control compiles into deterministic executable logic and is reperformed over the entire population, issuing a signed control certificate that can be re-run years later to reproduce the identical result — or to name exactly what changed.
5 CompletionArchive, inspection-response pack, firm quality dashboard
The archive is produced in a long-term preservation format with evidence embedded rather than linked. Inspection-response packs assemble from the stored chain, and a firm-level dashboard tracks quality-management components.
Standards as code
The corpus is structured data, not a PDF library. Clauses are addressable objects that findings, risks and procedures link to directly — which is what makes the one-button trace possible rather than aspirational.
Targeted invalidation
When new data arrives late — as it always does — the system neither silently re-runs everything nor pretends nothing changed. It flags exactly the artifacts the new data affects and leaves the rest of your concluded work intact.
This is the difference between software that respects a concluded file and software that quietly overwrites one.
Where it deliberately stops
Some judgments are legally reserved to a person. In those places the product is built to stop short.
| Area | What the software does | What stays with you |
|---|---|---|
| Findings | Detects, evidences, cites the clause | Accepting, editing or rejecting every one |
| Independence | Screens and surfaces what it found | The partner attests — the tool never asserts permissibility |
| The opinion | Assembles the supporting file | Entirely yours. It is a judgment, not an output |
Every AI-assisted surface is labelled, and the models and versions used on an engagement are recorded in a manifest that ships with the file.
Next
See whether the traceability claim holds up.
The fastest way to evaluate this is to point it at a real engagement and try to break the chain.