Off by default
The AI layer is switched off per organization until someone with authority turns it on.
Sofic is a modular ERP for nonprofits and the organizations that operate like them. It is built on real multi-basis fund accounting, with an AI layer designed into the schema, permissions and audit trail from day one. Two Genevity founders built it: a CPA who has spent her career inside nonprofit finance, and an infrastructure architect who has spent his keeping systems honest.
VISIT SOFIC.AISmall and mid-sized nonprofits have almost nothing in the middle. General-purpose tools bend classes and tags into standing in for funds and grants. That is a convention, not a control, so nothing stops a bad posting. The real fund balances end up in a side workbook that one person understands, and the audit finds out.
The incumbents that do understand fund accounting price accordingly: long implementations, bundles nobody asked for, and per-seat costs that punish growth. Meanwhile, most vendors are wiring a chatbot onto a ledger written a decade before language models existed.
The gap was never features. It was a correct ledger, at a price a nonprofit can pay, with AI that can be trusted near money.
Everything writes to one double-entry core, so there is no month-end reconciliation between your own systems. The core holds a segmented chart of accounts with combination validation, and a journal that keeps cash, modified-accrual and full-accrual books side by side. It produces statements under both FASB and GASB without an export step.
The controls live in the database, not in a training manual:
Fund accounting, payables and receivables, and fixed assets have shipped. Grants carry indirect-cost recovery, subawards and restriction release. The release amount is derived from what has actually been spent, never typed in. A vendor portal and the intelligent-workflow layer are in development.
“The assistant drafts, explains and warns. It does not approve, post or release money.”
THE RULE SOFIC IS BUILT UNDER
The AI layer is switched off per organization until someone with authority turns it on.
Its first job is review: it watches postings and raises flags in a queue. It has no path to change the books.
When it prepares an entry, it lands as a draft for a person to approve. Scheduled jobs that move money between a grant and the organization do the same.
Anything a machine keys is stamped in an append-only provenance record naming the agent and the tool, alongside the human responsible.
The assistant runs under the same permission model and segregation-of-duties controls as people. Nobody, human or model, approves their own work.
Its guidance comes from a rulebook where every rule carries its authority and verification status. Unverified rules are never asserted.
The knowledge behind the assistant is a separate product, Sofic Verified, now in development. A CPA authors each entry against the FASB and GASB Codification and a second person publishes it. The database refuses to let one person do both. Every published answer is cryptographically signed, so a finance team can check the citation itself instead of taking anyone’s word for it.
Some organizations must keep their books on their own premises because a grant condition or a board policy requires it. So Sofic ships with no front-end build step and no Node toolchain. It is server-rendered HTML that installs and runs on a server with no outbound internet at all. The same product runs self-hosted, in the customer’s own cloud, or hosted, so moving between them never means re-implementing.
Every rule in the rulebook has to name the code that implements it, and a test checks that the claim is true. That discipline has repeatedly found missing features that a schema review would have passed. Every screen is also photographed and reviewed. A green test suite is not the same as looking at the page.
The pattern behind Sofic carries over to any high-stakes system: governed AI, controls the database enforces, and software that runs on infrastructure you own.
START A PROJECT