Trust & Data Governance
Client documents are the most sensitive data in an engagement. Here is exactly how we handle them.
This page answers the four questions every firm asks before giving a vendor access to a prepared-by-client request list: where data sits, who can see it, whether we train on it, and what happens when you leave. Answers are concrete, specific, and updated when anything changes.
Four principles
The architecture decisions behind the governance posture.
These are design choices, not policies written for a marketing page. Each one shows up in the codebase and is the reason a SOC 2 readiness review would be short.
Tenant Isolation
Every engagement is scoped to one firm, enforced at the database row.
Supabase Postgres with row-level security is the storage layer. Tenant identifiers are required on every read and write, enforced by policy, not by application code. A bug in application code cannot cross tenants because the database rejects the query.
No Training
Your engagement data is never used to train any model, including ours.
Anthropic is configured with zero data retention. SynthGL does not fine-tune on customer documents. The Engagement Memory accumulates structured facts per tenant only. Cross-tenant learning from anonymized corrections exists as an explicit opt-in, disabled by default, controlled per engagement.
Retention
Engagement removal is archive-first today; purge windows are contractual.
The product delete path archives engagements and preserves audit history today. For pilots, purge timing and backup rotation are handled in the agreement and operating runbook; automated hard-delete evidence is a readiness item before broader rollout.
Provenance First
Every AI-surfaced number traces back to the specific cell in the specific file.
Findings are rendered with inline citations to the source document, sheet, and cell. Associates and reviewers can click through to the origin. The accounting rules engine is deterministic, so the same inputs always produce the same output.
Subprocessors
Five vendors touch your data path. Here is each one, what they do, and what they see.
We notify customers 30 days before adding or replacing a subprocessor that touches customer data. The table is the canonical list; anything not on it is not in the data path.
Supabase
Primary database and object storage
Data in path
Engagement documents, extracted financial entities, user accounts, audit events
Vercel
UI hosting and edge network
Data in path
Application code and static assets only. No customer documents pass through Vercel.
Fly.io
API hosting
Data in path
Transit-only. Customer data is written to Supabase, not stored on Fly.
Anthropic
LLM inference for Engagement Memory and inline agent features
Data in path
Structured facts and excerpts from customer documents, scoped to the active query
WorkOS
Authentication (OIDC, SSO)
Data in path
User identity metadata (email, name, organization)
AI learning policy
What the AI layer does with your data, in four modes.
Accounting AI vendor-diligence coverage asks what the partner's policies are around AI learning. Here is ours, broken out by mode.
Per-engagement memory
Default: On
Structured facts accumulate within one engagement (entities, cross-references, reconciliations). Scoped to that engagement's tenant. Does not cross into other engagements or firms.
Cross-engagement memory (same firm)
Default: On, firm-configurable
Patterns from one engagement can inform another within the same firm - e.g., recognizing recurring client account names. Scoped to the firm's tenant.
Anonymized corrections (cross-tenant)
Default: Off
When an associate overrides an AI-surfaced answer, the correction can be anonymized and folded into the shared rule library that improves detection for all customers. Per-engagement opt-in.
Model fine-tuning on customer data
Default: Never
Not offered. SynthGL does not fine-tune Anthropic models on customer content, and Anthropic does not train on SynthGL API traffic.
Engagement Memory is a provenance index rather than a machine-learning system. It accumulates structured facts - entities, cross-references, reconciliations - that a rules engine and a read-only tool surface reason over. No gradient descent, no fine-tuning, no model weights updated on your data.
Procurement questions
The seven questions that show up on every firm's intake form.
Pulled from Accounting Today AI vendor-diligence coverage, the AICPA AI vendor due-diligence guidance, and the procurement forms we have already seen from boutique advisory firms. Direct answers, no marketing fog.
How is multi-client data stored and separated?
Every row in the database carries a tenant identifier. Supabase row-level security policies enforce that queries can only touch rows matching the authenticated tenant. This is enforced at the database, so an application bug cannot leak across firms. Object storage is scoped the same way - each engagement has a tenant-prefixed bucket path.
What are your retention and purging policies?
Default retention matches the engagement lifecycle. The product delete path archives engagements and preserves audit history today, so reviewers do not lose the evidence trail accidentally. For pilots, purge timing, backup rotation, and any regulatory hold requirements are handled in the agreement and operating runbook. Automated hard-delete evidence is a readiness item before broader rollout.
What is your policy around AI learning from our data?
Your engagement data is never used to train any model. Anthropic is configured with zero data retention. SynthGL does not fine-tune on customer content. The Engagement Memory is a structured store, not a model - it holds typed facts (entities, cross-references, reconciliations) scoped to your tenant. Cross-tenant learning, where it exists, is limited to anonymized corrections that a firm can opt into per engagement; it is disabled by default.
How do you handle depreciation and debt-like items?
This is the accounting AI vendor test we keep seeing in trade coverage and advisory-firm diligence: ask how the system handles depreciation and debt-like items. Both worked examples live on a dedicated page: How we model accounting. It walks through the depreciation presentation styles we classify today, the EBITDA D&A add-back rule, the default financial-statement line-item (FSLI) classification for the nine debt-like categories raised in that vendor-diligence prompt, and the live /nwc/ev-bridge API. Scope lines mark what is V1 today, what is V1.5, and what is V2.
Who are your subprocessors?
The full subprocessor table is on this page above. The short list: Supabase (primary database and storage), Vercel (UI hosting), Fly.io (API hosting), Anthropic (LLM inference, zero retention configured), WorkOS (authentication). We notify customers 30 days before adding a subprocessor that touches customer data.
Do you have SOC 2?
Not yet. SynthGL is pre-seed and working with boutique firms where SOC 2 is not a procurement requirement. Readiness assessment is planned; full Type II is targeted after the first cohort of design partners and before mid-market expansion. In the interim, the current control documentation for access control, change management, backup, monitoring, and incident response is available for review under NDA.
Can we sign a DPA?
Yes. A standard SaaS DPA is available on request and can be attached to the Design Partner or Founding Customer agreement. If your counsel prefers to work from your own template, we will negotiate from that baseline.
What happens if SynthGL is acquired or goes out of business?
Customer data is exportable at any time in structured form (normalized entities, original documents, audit log). Design Partner and Founding Customer agreements include a 90-day continuity clause covering export and assisted migration on termination, and a cap on price changes during the locked term. Engagement data in recurring workflows (409A, quarterly monitoring) stays exportable across periods.
What to do next
If your IT or legal team has a question not covered here, ask it directly.
Send procurement questions to trust@synthgl.com. A DPA and architecture diagram are available under NDA before any pilot starts. Most procurement cycles at this stage close in a single call once both documents are in the review queue.