How "AI that does the work under control" is scored for Essential Eight platforms
AI that does the work under control carries 10 percent of the index and measures AI or an agent that drafts answers and policies and does compliance tasks, with review and approval before anything is used. The highest score is 10 of 10, for Comp AI; Scytale scores 7 of 10, 2nd of 13 platforms on this criterion.
How is this criterion scored?
Add: (a) AI or an agent that collects evidence or carries out compliance tasks, documented 3; (b) AI that drafts or tailors policies 2; (c) AI that drafts answers to questionnaires (security questionnaires or the scheme's own questions) 2; (d) documented human review or approval before AI output is submitted or used, or an audit trail of AI actions 3. An AI feature with no documented review step scores 0 on (d). Cap 10. Vendor-only evidence caps the total at 7.
Weight in the default index: 10 percent. Situation weights range from 5 to 15 percent.
What does each platform score on it?
| Platform | Score | Reason |
|---|---|---|
| Comp AI | 10 | AI agents create tests and use a browser to verify a control and screenshot it (3), policies are generated from onboarding context (2), questionnaires are answered from approved policies with review before export (2), and policy changes need review "before they're saved" (3). A TechCrunch interview confirms an agent drafts a policy and "a person still reviews and approves it", so the cap does not apply. |
| 6clicks | 7 | 6clicks AI validates and maps evidence to controls and the homepage describes agents that collect evidence (3), it generates "control definitions and policy descriptions" (2), drafts assessment responses from uploaded documents (2), and states human oversight including "reviewing AI outputs" (3). All evidence is from 6clicks pages, so the total of 10 is capped at 7. |
| Delve | 7 | Autonomous AI agents "take screenshots, write reports, and perform validation of your evidence" (3), and AI autofills vendor questionnaires (2). A dedicated team member "reviews everything by hand" before an third-party firm examines it (3), though no approve step for AI questionnaire answers is described. AI policy drafting: none found on 18 Nov 2026; templates are starting points. All evidence is from Delve pages, so the total of 8 is capped at 7. |
| Drata | 7 | Drata documents agents for third-party risk, AI test failure insights and AI control mapping (3), AI answers to security questionnaires (2), and users who "approve, edit, or reject" those answers (3). AI policy drafting: none found on 18 Nov 2026; the AI only suggests controls for policies. All evidence is from Drata pages, so the total of 8 is capped at 7. |
| Scrut Automation | 7 | Named AI Teammates "draft policies, collect evidence, detect gaps, and prepare audit packages" (3), the Policy Architect generates policy drafts (2), questionnaires are auto-filled with an option to edit before submitting (2), and "Your team reviews and approves" (3). All evidence is from Scrut Automation pages, so the total of 10 is capped at 7. |
| Scytale | 7 | Scytale documents AI agents on its agentic compliance page, including a detected MFA gap (3), one-click policy generation "tailored to your unique needs" (2), AI-drafted questionnaire answers (2), and users who "approve, edit, or reject" AI answers (3). All evidence is from Scytale pages, so the total of 10 is capped at 7. |
| Secureframe | 7 | AI suggests control mappings and an Evidence Validation check runs on evidence (3), the policy editor has an AI writing assistant (2), AI drafts questionnaire answers (2), and those answers cite sources for the user to "verify and edit" (3). Evidence Validation states no manual review is needed for its own check. The third-party interview confirms only the MCP server for querying data, not a scored element, so the total of 10 is capped at 7. |
| Sprinto | 7 | Sprinto lists an Evidence Agent, Fix it agent, Vendor Agent and Trust Agent (3), "AI-assisted policies" tailored to the organisation (2), AI questionnaires with 20 a year in Foundation (2), and states "Your team reviews, approves, and steers" (3). All evidence is from Sprinto pages, so the total of 10 is capped at 7. |
| Thoropass | 7 | The MCP server announced 9 July 2026 lets AI agents work in the audit workflow (3), due diligence questionnaires are "powered by GenAI" (2), and the team can "verify the work before it moves forward" (3). AI policy generation: none found on 18 Nov 2026. All evidence is from Thoropass pages, so the total of 8 is capped at 7. |
| Vanta | 7 | The Vanta Agent is documented to "draft policies, complete questionnaires, and flag issues" (3), policies are drafted from templates (2), questionnaire answers are suggested from the knowledge base (2), and answers are "review and submit" with a stated human-in-the-loop principle (3). All evidence is from Vanta pages, so the total of 10 is capped at 7. |
| Avertro | 6 | AI framework translation suggests scores and evidence and maps an assessment across frameworks (3), and the user must "review and confirm before anything changes" (3). The Iris assistant is read-only. AI policy drafting and AI questionnaire answers: none found on 18 Nov 2026. Evidence is from Avertro pages only and the total of 6 is under the cap. |
| Huntsman Security | 0 | AI that does compliance tasks, drafts policies or answers questionnaires: none found on 18 Nov 2026 on the Auditor and Scorecard pages. With no AI feature, no review step scores. |
| Protecht | 0 | The Protecht AI assistant helps with form entry, incident search and executive report summaries; an agent that collects evidence or carries out compliance tasks: none found on 18 Nov 2026. AI policy drafting and AI questionnaire answers: none found on 18 Nov 2026. A general "human-in-the-loop design" statement does not say which outputs need approval, so (d) scores 0. |
Sources
- Platform profiles with sources for every score, read 18 Nov 2026
Last reviewed 18 Nov 2026