How to run your compliance program with AI: connectors, browser agents and a human in the loop

This is a practical, level-deeper guide. If you are new to the idea, start with can AI run my compliance program? and do I still need a compliance platform?, then come back here for the mechanics.

Why compliance centralised in the first place, and what just changed

Compliance platforms boomed for a good reason. Evidence for a SOC 2 or ISO 27001 audit lives scattered across a dozen systems: cloud consoles, code repositories, HR records and policy documents. Collecting it by hand meant knowing where to look in each one, what the auditor actually needed, and pulling it together in one place for the auditor.

Platforms solved that with hundreds of built-in integrations that pull evidence in. For most teams that was transformative compared with folders of screenshots.

What changed is that the integration layer stopped being proprietary. AI assistants like Claude and ChatGPT now connect to the same tools directly, through open connector standards such as the Model Context Protocol (MCP). They can read, collect, organise and explain compliance evidence from the systems where your controls actually operate.

The user need the platforms served (stop doing manual evidence collection) can now be met without introducing a new system to maintain. There are three ways to do it, and most programs will use a combination of all three.

The three ways AI connects to your compliance program

Connectors are the default, the browser fills the gaps, and a human in the loop covers judgment calls and sign-ins. The right question for each piece of evidence is: which method produces the strongest artefact with the least ongoing effort?

Connectors (APIs)

• How it works: AI reads systems directly through authorised API connections.
• Coverage of a typical startup stack: about two-thirds of tools, and roughly 90% of evidence by volume.
• Evidence produced: structured, timestamped system records (such as JSON exports) and documents.
• Best for: recurring collection, configurations and populations.

Browser agent

• How it works: AI operates a signed-in browser session, page by page.
• Coverage: nearly everything with a login page.
• Evidence produced: screenshots and records captured from the user interface.
• Best for: tools without connectors, admin consoles and one-off checks.

Human in the loop

• How it works: AI tells you exactly what is needed, where it lives and how to capture it. You act.
• Coverage: everything.
• Evidence produced: whatever you capture, correctly specified the first time.
• Best for: sign-ins and MFA, judgment calls and sensitive systems.

Method 1: Connectors, evidence straight from the source

Connectors are the most reliable method. They work the same way as the built-in integrations that made compliance platforms valuable: authorised, API-level access to the systems where controls operate. The difference is that they are now a property of your AI assistant rather than a platform. You grant access once, and the AI can pull the records on demand or on a schedule, with no one at a keyboard.

What connectors produce is the strongest form of audit evidence: structured, timestamped system records exported as data directly from the system of record. Think of the actual configuration of your production databases, the real user list from your identity provider, or the branch-protection rules on your repository. Auditors can test these at the record level, and their provenance is self-evident in a way no screenshot can match.

Coverage is already further along than most teams expect. The connector directories in AI tools like Claude cover more than two-thirds of the tool categories a typical SaaS company needs for compliance: cloud infrastructure, code repositories and ticketing, knowledge bases like Notion and Confluence, and HRIS systems like BambooHR and Deel.

Weighted by how much evidence actually comes from each tool, in our experience that is closer to 90% of a startup's audit evidence, because the highest-volume sources (cloud and code) are the best covered. Anything a connector can't reach yet usually still has an API, which an AI agent can call with a read-only credential. It is the same mechanism, just without the pre-built wrapper.

Method 2: Browser agents, closing the gap

Browser agents take coverage close to 100%, because they can reach anything you can log into. Instead of calling an API, the AI operates a real, signed-in browser session. It navigates an admin console, opens the right settings page and captures what it finds. Where no connector or API exists, or where the evidence is genuinely a screen (an admin toggle, a vendor portal), this is the method that gets it.

It comes with honest trade-offs in both directions. The downside is fragility. Browser agents depend on the user interface, so redesigns, popups and loading quirks can break a run in ways an API call never does. They also generally need you signed in and nearby for authentication.

The upsides are underrated, though. Screenshots are the evidence format auditors have tested for decades, so a well-captured, timestamped screenshot often moves through an audit with less friction than a novel data export. A targeted screenshot of the one relevant setting also avoids handing over sprawling system records that invite questions beyond the control being tested. Scoping evidence to the control is good audit hygiene, not concealment, and the browser is naturally good at it.

The practical pattern: use the browser for the long tail. That means the handful of tools without connectors, periodic spot-checks of admin settings, and anywhere the auditor's request is literally "show me this screen."

Method 3: Human in the loop, AI as navigator

The third method is the most overlooked. The AI doesn't act on your systems at all. It tells you exactly what to do, and you do it.

That sounds modest until you remember what made evidence collection hard in the first place. It was never the clicking. It was the figuring out. Which of the forty screens in the AWS console does the auditor mean? What does "evidence of access reviews" actually need to show? Is this screenshot sufficient, or will it bounce back in three weeks with a follow-up request?

AI collapses that overhead even when it touches nothing. Tell it your stack and your framework, and it can specify each piece of evidence precisely: the system, the path, what must be visible, and why the auditor needs it. You capture it right the first time.

The task is also smaller than the industry's framing suggests. A startup with a right-sized control set of 20 to 30 configured controls typically needs on the order of 30 to 50 screenshots and exports in total, most of them sitting in one cloud console, such as AWS or Vercel running on top of it. Collected with a precise shopping list, that is hours of work, not weeks. See what are Compliance Shapes? for how a right-sized control set is designed.

This mode also stays the right one for a few things permanently: sign-ins, MFA and passkeys; systems you'd rather no agent touched; and the judgment calls, such as what to scope in or how to word a management response, where the AI should inform the decision, not make it. That boundary is part of how you set an agent's authority and escalation.

Where compliance evidence actually lives

Four categories of tools hold the vast majority of a startup's audit evidence. Map your own stack against this and you know which method does what before you start.

Cloud infrastructure

• Typical tools: AWS, or Vercel on top of it; databases like MongoDB Atlas and Snowflake.
• What the auditor needs: security settings and configurations, including encryption, logging, network rules and backups.
• Best method today: connector or API. This is the highest-volume and best-covered category.

Identity provider

• Typical tools: Google Workspace, Microsoft 365 or Entra ID, or dedicated SSO like Okta.
• What the auditor needs: user populations for sampling, MFA status and group memberships. The reviews and approvals sit outside the tool.
• Best method today: browser or a read-only admin API. Connector coverage is thinnest here, but evidence volume is low.

Code repository and CI/CD

• Typical tools: GitHub, GitLab, Bitbucket and pipeline tooling.
• What the auditor needs: branch protection, peer review, automated testing, approval stages and release controls.
• Best method today: connector. Well covered and configuration-heavy.

Organisational systems

• Typical tools: HRIS (BambooHR, Deel, Rippling) and knowledge bases (Notion, Confluence, Google Drive).
• What the auditor needs: policies, onboarding and offboarding records, board minutes, management reviews, and risk and vendor assessments.
• Best method today: connector. APIs index and search content far better than paging through a user interface.

One nuance on the last category: for document-heavy organisational evidence, connectors beat the browser even though both can reach it. An API-connected AI can search and cross-reference an entire workspace to find the approval, the minutes or the policy version it needs, rather than navigating page by page.

The interface layer: your systems are already the system of record

Centralising compliance into one dedicated platform can quietly create the duplication it set out to remove. Your compliance already lives in live systems across the company: policies in Notion, Confluence or Google Drive; onboarding and offboarding in your HRIS; changes in your repo; board minutes wherever the company keeps them. Copying all of that into a separate system means maintaining two versions of the truth, and the copy is the one that drifts.

The agentic alternative inverts it. Leave every record in the system where the control actually operates, and let AI connect the dots across them: assembling the evidence set, checking completeness against your control list and flagging what's stale. The Compliance Brain gives the agent the company context to know what complete looks like.

The remaining question is the interface layer: where do humans and AI meet to work on compliance together? That can be Google Drive, Notion, OneDrive or Confluence, whatever your team already lives in. It doesn't need to be a software category purpose-built for compliance. It needs to be the place your team will actually keep current, with the AI doing the connecting underneath.

Frequently asked questions

Do you need a compliance platform to pass SOC 2 or ISO 27001?

No. Auditors test controls and evidence, not tooling. A program run through AI connectors, a browser agent and a well-specified evidence list produces the same (often stronger) artefacts an auditor needs. Platforms can still add value; see do I still need a compliance platform?

What's the difference between an AI connector and a browser agent for compliance?

A connector reads systems through authorised APIs and returns structured, timestamped records without anyone watching. A browser agent operates a signed-in session through the user interface. It reaches more tools, but produces screenshot-grade evidence and needs a human nearby for logins.

Will auditors accept AI-collected evidence?

Yes, when it is a source-system artefact, such as an API export or a screenshot of the live system, with clear provenance and timestamps. What auditors won't accept is an AI-written summary standing in for the underlying record. The record itself is the evidence. More on this in who can audit agentic compliance?

How much evidence does a startup audit actually involve?

For a right-sized program of 20 to 30 configured controls, typically 30 to 50 discrete artefacts, most from the cloud console, plus the recurring records (reviews, tickets, HR events) that accumulate in the systems where they happen.