SOC audits, with the fieldwork done by AI.
A SOC 2 audit is mostly evidence: screenshots of access lists, exports of change tickets, samples pulled from logs, and weeks of back-and-forth to get them. Your team spends a quarter assembling it, and the auditor spends another reviewing it.
Cerulin replaces that fieldwork with AI agents. They connect to the systems your controls live in, collect the evidence directly, test every control against its population instead of a sample, and hand a licensed CPA a finished workpaper to review and sign.
Talk to usContinuous testing, not a once-a-year scramble.
Agents test your controls throughout the review period, so issues show up when they happen instead of in the auditor’s findings letter.
Read-only connections to your cloud provider, identity provider, ticketing, HR, and code repositories. No agents installed on your servers, no passwords shared.
Each control is mapped to the Trust Services Criteria and tested against the full population: every terminated employee, every production change, every access grant in the period.
Exceptions are flagged with the evidence attached. At the end of the period, a CPA reviews the agents’ workpapers, forms the opinion, and issues the report.
The controls that eat your audit season.
Who has access to what, whether it matches their role, and whether it was removed on time when they left.
Every production deploy traced back to a ticket, a reviewer who was not the author, and a passing build.
Confirms alerts are configured, routed to someone, and actually acknowledged when they fire.
Tracks your subprocessors, pulls their own SOC reports, and flags any that have lapsed.
Background checks, security training, and policy acknowledgements matched against your HR system.
Checks storage and databases for encryption at rest, and that backups exist and have been restored at least once.
AI does the fieldwork. A CPA signs the report.
A SOC report is an attestation, and only a licensed CPA firm can issue one. Our agents never form the opinion; they prepare the evidence a human auditor needs to form it, and every conclusion can be traced back to the source.
Each test records what the agent queried, what it found, and which model version made the call, so a reviewer can re-perform any of it.
Evidence is encrypted, isolated per customer, kept in the United States, and never used to train a model.
Anything the agents can’t resolve with certainty goes to an auditor, not into the report.