SOC 2 and Compliance Automation: from Spreadsheets to AI agents

Compliance before automation was a calendar event. Sarbanes-Oxley in 2002 made internal controls a board-level problem for public companies, PCI DSS arrived in 2004 for anyone touching card data, and in 2011 the AICPA's SOC reports replaced SAS 70 as the way a service provider proved its controls to customers. Each framework meant a fresh binder of policies and a fresh round of evidence.

The evidence was screenshots. Someone logged into the firewall, the HR system and the cloud console, captured a screen, pasted it into a Word document, and filed it in a shared drive named for the audit year. Control status lived in Excel. Consultants wrote the policies, and the auditor sampled whatever the team could find in the weeks before fieldwork.

I lived this. As Product and Compliance Manager at Sageworks, I built the SOC 2, PCI and ISO programs, and the hard part was never the controls. It was proving, months later, that each control had actually run. If you are starting today, my simple guide to SOC 2 for startups covers what the report asks for.

Large enterprises bought on-prem GRC suites like Archer, which EMC acquired in 2010 and folded into RSA. Those tools organized risk registers and policy libraries well, but they did not collect evidence. A human still fed them, so they mostly gave the spreadsheet a better database.

The shift started with a new category of vendor. Vanta launched in 2018 to automate SOC 2 evidence collection. Drata followed in 2020, and Secureframe and Sprinto joined the same wave. The pitch was simple: connect your cloud, identity provider, code repository and HR system, and the platform pulls the evidence for you.

That changed the unit of work. Instead of screenshots, an API call checked that MFA was on, that disks were encrypted, that offboarded employees lost access. Continuous monitoring replaced the pre-audit scramble, and one set of evidence could map to SOC 2, ISO 27001 and HIPAA at once. I compare how the main platforms differ today in Vanta vs Drata vs Sprinto vs Secureframe.

The second shift was on the sales side. Trust centers from vendors like SafeBase put the SOC 2 report, policies and subprocessors behind a self-serve page, so buyers stopped emailing for the same PDF. Pricing settled into annual contracts sized by framework count and headcount.

What did not change: policies still needed a human writer, auditors still sampled evidence by hand, and the integrations only covered what had an API. Anything outside the cloud stack, like board minutes or vendor reviews, was still an upload.

Generative AI arrived in compliance through the most hated chore: writing. Secureframe shipped Comply AI in June 2023, which generated infrastructure-as-code fixes for failing cloud tests. Within a year most platforms offered AI policy drafting, control explanations and suggested questionnaire answers drawn from past responses and the trust center.

Questionnaire automation was the clearest win. A 300-question spreadsheet from an enterprise buyer could be pre-filled in minutes from a knowledge base, then reviewed by the security lead. Conveyor, Vanta and Drata all sold this. The human still read every answer, because a wrong answer in a security review becomes a contractual representation.

Policy drafting got faster too, but it exposed a trap I write about in what founders actually need in SOC 2 policies: a generated policy that describes controls you do not run is worse than no policy, because the auditor will test it.

Meanwhile the scope grew. ISO/IEC 42001, the first AI management system standard, was published in December 2023, and the EU AI Act entered into force in August 2024. Compliance teams now had AI to govern as well as AI to help them.

In 2025 the vendors stopped selling suggestions and started selling workers. Drata agreed to buy SafeBase in February 2025 to own the trust center. Conveyor shipped Sue in April 2025, an agent that triages inbound security reviews, drafts answers, chases subject matter experts in Slack and updates Salesforce. Vanta launched its AI Agent in June 2025, starting with policy onboarding and evidence evaluation against audit requirements.

By March 2026 the whole category had rebranded around agents. Vanta introduced compliance, third-party risk and customer trust agents that it says keep humans in the loop for final decisions. Drata said agentic vendor risk assessments were available to all customers and put agentic questionnaire response into beta. Sprinto launched what it calls an Autonomous Trust Platform that refreshes evidence, prepares audit artifacts and runs vendor due diligence. Secureframe and Sprinto both shipped MCP servers so assistants like Claude can read and act on live control data. Secureframe's hosted version can also update test statuses and create remediation items, which makes write scopes a buying question, not a footnote.

Vendors say their agents now pull and check evidence, flag control drift, pre-fill questionnaires from approved sources, and assess vendor documents, though none of those claims is independently benchmarked. What they do not do: sign the audit opinion, accept a risk, or vouch for a control they cannot observe. The auditor is still a licensed human firm.

The March 2026 allegations against one AI-first compliance startup, backed by Insight Partners, showed the failure mode. An anonymous whistleblower claimed it fabricated evidence of board meetings and tests that never happened. The startup says it only offers templates and does not issue compliance reports, and it denies providing fake evidence. TechCrunch reported in April that Y Combinator had parted ways with it, and that two of its customers at the time, LiteLLM and Context AI, had since disclosed security incidents. Nothing reported links those incidents to their compliance reports. Automation that produces the artifact without the underlying control is not compliance. And as agents spread, Drata opened limited early access to AI agent governance in June 2026, because nobody owns agent behavior yet.

My bet: by 2031 the annual audit stops being an event. Agents will keep evidence current every day, and auditors will consume machine-generated, signed evidence streams instead of sampling screenshots. The SOC 2 report becomes a live attestation with a freshness date, not a PDF that is eleven months stale by the time a buyer reads it.

The early signs exist. Drata says its vendor-risk agent already collects documents from vendors' trust centers and runs assessments. The next step is the buyer's agent and the seller's agent settling a security review between themselves, with humans reviewing only the exceptions.

The scope will keep growing. ISO 42001 is becoming the expected answer to "how do you govern your AI", and the EU AI Act's stand-alone high-risk obligations now apply from 2 December 2027 under the adopted digital omnibus, Regulation (EU) 2026/1744. Even companies that only call a model API carry duties, which I cover in the EU AI Act for the downstream provider.

What has to be true: auditors and the AICPA must accept agent-collected evidence with a verifiable chain of custody, and evidence must be cryptographically signed at the source so nobody can template it. That is the same shift cryptographic consent predicts for signatures: proof that is checked by math, not by a person squinting at an image.