Minimise first
Do not send transaction data when the task can be completed without it.
What leaves the machine, and what never does.
The safest ledger upload is the one that never happens. Audit Scoot runs every analysis against the local engagement and limits the optional assistant to the auditor's question and a profile of the ledger's columns.
The difference is the data exposed to the model context.
Data or controlDirect ledger upload to general AIAudit Scoot
Journal lines, balances and resultsCan enter provider contextRemain local
Account, counterparty and preparer names, narrativesCan enter provider contextRemain local
Column profile: names, types, counts, numeric rangesIncluded with the fileSent with the question, only for an AI analysis
Population testModel-dependent interpretationVersioned query executes locally
Client documents (registers, listings, statements)Can enter provider contextRemain local; only their names and columns go with a question
Review evidenceChat transcript or manual reconstructionWorksheets that state what they tested, with their audit logic, in a workbook indexed by the SHA-256 of each source
This comparison describes architecture and data scope. It is not a numerical claim that any product is “1,000× safer”.
What stays local, what can transit, and why.
Audit Scoot does not rely on a vague “secure AI” promise. The boundary is defined by data class.
Journal rows, balances, names, narratives, documents and results
They stay in the local engagement file and are analysed on the auditor's machine, with every client document loaded beside the ledger. By default the application makes no network connection at all.
Your question and a profile of the columns
Column names and types, distinct counts, share of blanks, numeric minimum and maximum, the year end and materiality thresholds, and the names and columns of the documents loaded. Sent to Google's Gemini model under Audit Scoot's account, only when a user asks the assistant; the query it returns is validated and runs locally.
Update check and diagnostic archive
With a release feed configured, the app fetches a signed manifest at most weekly. A diagnostic archive is written only when a user exports one, and they send it themselves.
Execute read-only
Generated procedures inspect the local engagement population rather than modifying it.
Keep review human
Custom analyses are drafts for professional review, not autonomous audit conclusions.
Export the trace
Every worksheet states what it tested and goes into the workbook beside the fingerprints of its sources, instead of disappearing into a chat response.
For your IT review.
The answers your IT and security team will ask for, as the application ships. The full security and privacy document is available on request.
| Does it need admin rights? | No. A per-user install that runs as the user; the installer supports /S for a silent deployment. |
|---|---|
| Which ports does it open? | One loopback port on 127.0.0.1, random at each launch and protected by a token drawn per launch. Nothing listens on the network. |
| Outbound connections? | None by default. When a user asks the assistant: generativelanguage.googleapis.com (Google's Gemini API). If a release feed is configured: that feed's URL. |
| Which files can the application read or write? | Only those the user chose in a Windows Open or Save dialog, picked or dropped, or files in the engagement library folder. The engine refuses any other path, so a script in the window could not read or overwrite a file nobody chose. |
| Telemetry or analytics? | None. |
| Where do engagement files live? | Documents\Audit Scoot by default, or wherever the firm keeps working papers. They are not encrypted by the application: keep them on encrypted volumes or in the firm's document system. |
| What stays on the machine after a session? | The engagement files, the settings database and a rotating log that by rule holds no client value. Temporary copies are purged at every start; crash dumps stay local. |
| How are releases supplied? | Every release ships a CycloneDX software bill of materials and SHA-256 hashes of the installer; update manifests are signed (Ed25519). |
Questions an audit firm should ask.
Does the AI receive journal entries?
No. Nothing leaves the machine by default: the application makes no network connection. Only when a user asks the assistant a question does the application send Google's Gemini model, under Audit Scoot's account, the question and a profile of the ledger's columns — names, types, counts and numeric ranges, plus the year end and materiality thresholds — and the names and columns of any client documents loaded beside it; never a ledger or document row, an account or counterparty name, a narrative or a file name. Gemini returns a query, which the application validates and runs locally.
Can our IT turn the AI off?
Yes. Nothing is sent until a user asks the assistant a question, and a firm policy setting refuses it outright: the engine then answers every AI request with a refusal and sends nothing. Analyses the assistant built earlier can still be re-run: that executes their saved query locally and asks no model.
Do the client documents we load leave the machine?
No. Registers, listings and statements loaded on the Documents page are read, stored and analysed locally like the ledger. When someone asks the assistant a question, the names, row counts and columns of the loaded documents go with it so the query can read them — never a row.
Where does the generated query run?
Locally, read-only, against the engagement. The application validates it first — SELECT statements on the ledger and the loaded documents only, nothing that writes — and keeps it with the engagement with the question, its parameters and who asked, so it can be re-performed later without the model. The auditor remains responsible for reviewing its suitability and results.
Is Audit Scoot ISO 27001 or SOC 2 certified?
No current certification is claimed. The core risk reduction comes from data minimisation: the service does not need the ledger population to generate or execute tests.
Does this make every AI use risk-free?
No system is risk-free. Audit Scoot materially reduces exposure by removing transaction data from the AI request, limiting what is sent to a profile of the columns, using read-only local execution and keeping professional review explicit.
How does this support audit documentation?
Each worksheet states the population it tested, and the exported workbook indexes every source file by its SHA-256 fingerprint with the ledger's control totals; the engagement file keeps an audit trail of changes under each user's Windows account. These can contribute to documentation prepared under the applicable auditing standards. Audit Scoot does not certify compliance or replace professional judgement.
Run your next audit file test without the Excel crash.
One free engagement to evaluate, valid 60 days. The ledger never leaves the machine.
- 20 minutes
- Runs on a fictional engagement
- No client data required