AI for GRC: Intelligence and Action in Practice with Michael Rasmussen and MJ Raber
Show Notes
Webinar Overview
Many GRC programs have their policies, risks, and controls well documented. Practitioners still end up connecting that information manually: linking the risk register to control evidence, following up over email, and tracking down the person who needs to make a decision.
In this conversation, Michael Rasmussen of GRC 2020 Research and MJ Raber, Senior Director of GRC AI at Trustero, apply Michael’s GRC 7.0 Orchestrate framework to work practitioners already do every day: control testing and third-party risk management. They discuss what AI agents can contribute, how practitioners direct and review that work, where human judgment remains essential, and how GRC leaders can tell whether the approach is helping. The session ends with a walkthrough of Trustero’s AI agents in a control test and a vendor assessment.
Key Discussion Highlights
From system of record to system of orchestration
GRC 1.0 through 6.0 centered on the system of record: a relational database with forms, workflows, and tasks. Michael compares relying on last quarter’s assessment to driving a car while looking in the rearview mirror.
GRC 7.0 keeps the system of record as the foundation for obligations, policies, risks, controls, and accountability. Around it, a system of orchestration connects three subsystems: intelligence, which interprets context and evidence; configuration, which lets people build and adapt capabilities; and action, which performs or coordinates authorized work. Human oversight applies across all three.
Control testing starts with context
A control statement on its own rarely gives an AI agent enough to evaluate. The agent needs the control objective, the policy or requirement behind it, the associated risk, the operational procedure, the evidence, and the test procedure.
MJ notes that the operational procedure is one of the hardest pieces to get right, because a vague procedure is easier to write. AI agents can help teams document who performs an activity, in which system, when, and how, so the process no longer depends on one person’s memory.
Making uncertainty visible
Michael separates evidence that supports a control, evidence that indicates a failure, and evidence that is insufficient to reach a conclusion. A missing record does not prove a control failed, and it does not justify a pass either.
The accountable person should be able to challenge the requirement applied, the evidence selected, the test criteria, and any assumptions that affect the result. In Michael’s words, a green approval button without that context does very little for oversight.
The practitioner’s role in configuration
As agents take on more of the work, practitioners spend more time defining what good work looks like and checking whether it was done well. For a control test, that means setting the objective, relevant account and system information, required conditions, expected output, review points, and what should happen when evidence is missing.
MJ describes Trustero’s TI Playbooks, specialized AI agents that practitioners create or adapt in natural language. Michael adds that new instructions should be tested on a known case first, then on contradictory records and unusual exceptions, and that limits of authority belong in workflows and permissions.
Third-party risk with context intact
A vendor assessment involves many tasks, systems, and people. Michael explains that each step should stay tied to the vendor’s service, the data or access involved, and the decision the organization must make. When an attestation is missing or a questionnaire conflicts with the evidence, the gap should stay visible and block only the conclusions that depend on it.
He describes three working modes: an agent completes a bounded task, a practitioner and an agent resolve an inconsistency together, or a step requires human input, review, or approval before proceeding. MJ connects this to Trustero’s long-running horizon agent, which follows the TPRM steps a practitioner sets and notifies them when their input is needed.
Measuring whether the approach is helping
Michael recommends setting a baseline and measuring value from four angles:
- Efficiency
- Effectiveness
- Resilience
- Agility
Efficiency should include the time spent reviewing and correcting results, along with the time saved gathering evidence. Effectiveness should account for missed issues and unsupported passes, since, as Michael puts it, you can make bad processes faster.
Five Major Takeaways
1. Start with one recurring piece of work
Pick a task with a clear objective, a verifiable result, an identifiable owner, and real practitioner time behind it, such as one control test or one evidence review.
Write down how the work happens today, what information it needs, and where it stalls. There is no need to perfect the GRC program first. Michael recommends small, incremental changes in the spirit of kaizen.
2. Make operational knowledge explicit
Document the control objective, requirements, risk, operational procedure, evidence, and test procedure so an agent, a reviewer, or a new colleague can understand how the control operates.
Clear procedures also separate a control that is well designed from one that operated effectively. The procedure describes what should happen, and activity records show whether it did.
3. Require conclusions that can be traced and challenged
An AI conclusion should show the supporting records, explain how they relate to the condition tested, and state where the information ends.
Reviewers should be able to question the requirement, evidence, test criteria, and assumptions, and to request more evidence. Sometimes a finding reveals a weak test procedure rather than a failed control.
4. Set limits of authority in the workflow
Decide which tasks agents complete on their own, which they work on with a practitioner, and which require human review or approval.
Reflect those decisions in workflows and permissions, and keep accountable people responsible for consequential calls, such as whether evidence is sufficient or a material risk can be accepted. Exceptions need an owner, a reason, and a review or expiration date.
5. Test the difficult cases before expanding
An expired document or a contradictory record can tell a team more about a capability than a polished demonstration.
Buyers should ask vendors to follow the connections using evidence from their own environment: trace a conclusion to its records, see what happens when information conflicts or approval is withheld, and confirm the record shows what happened afterward.
Chapter Markers
02:15 – Welcome and session agenda
02:45 – Michael Rasmussen’s background and the origin of the term GRC
04:15 – MJ Raber’s introduction and the GRC 7.0 Orchestrate framework
06:05 – GRC 7.0: from system of record to orchestration
07:20 – Intelligence, configuration, and action
08:05 – Why the control is central to a GRC program
09:50 – Why operational procedures are hard to get right
10:30 – Access control test example
11:20 – The context an agent needs to evaluate a control
13:05 – What AI intelligence is and is not
14:30 – Giving AI agents clear guidance
15:00 – When evidence is incomplete or contradictory
16:30 – Human review and oversight
16:40 – How Trustero customers refine controls and test procedures
18:00 – Tuning test procedures to each environment
19:30 – The practitioner’s role in configuration
20:10 – Defining and testing agent instructions
21:30 – GRC for AI
21:45 – TI Playbooks as specialized AI agents
23:15 – Deciding what agents do and where humans review
25:05 – Third-party risk management and orchestration
26:15 – Agent, collaborative, and human-approval modes
27:00 – Accountable judgment and exceptions
27:50 – Long-running horizon agents for TPRM
28:10 – When a vendor assessment is sufficiently supported
29:05 – Questions GRC teams should ask themselves
29:35 – Where GRC leaders should start
30:20 – Four angles of value: efficiency, effectiveness, resilience, and agility
31:20 – Testing difficult cases early
32:20 – What buyers should look for
33:40 – Where to read Michael’s GRC 7.0 research
34:30 – Product demo: segregation of duties control
36:00 – Natural-language test procedures and control test results
37:10 – Receptors and continuous control monitoring
38:00 – Refining procedures with TI Playbooks
39:00 – Building custom playbooks
39:55 – TPRM demo: vendor evaluation timeline and requests
41:45 – Closing
Concise Webinar Summary
Michael Rasmussen explains that GRC 7.0 keeps the system of record as the foundation and adds orchestration across intelligence, configuration, and action, with human oversight throughout.
MJ Raber applies the framework to control testing and third-party risk management. AI agents can gather evidence, run control tests, and coordinate vendor assessments, while practitioners define the work, set limits of authority, and make the consequential decisions.
Both speakers stress clear context and traceable conclusions. Controls need documented objectives, procedures, and test criteria, and AI conclusions need supporting records that a reviewer can examine and challenge.
The central message is that GRC teams can start small, measure efficiency, effectiveness, resilience, and agility, and expand AI-driven work based on what the results support.
Webinar Description
Join Trustero for a conversation with Michael Rasmussen, GRC analyst at GRC 2020 Research and the person credited with coining the term GRC, and MJ Raber, Senior Director of GRC AI at Trustero, as they discuss how Michael’s GRC 7.0 Orchestrate framework applies to day-to-day GRC work.
The discussion covers what changes when GRC technology adds orchestration to the system of record. Michael describes how intelligence, configuration, and action connect business objectives, obligations, controls, and evidence, so practitioners no longer have to link that information manually.
MJ applies the framework to control testing. Using an access control example, the speakers walk through the context an AI agent needs to evaluate a control: the objective, the requirement, the risk, the operational procedure, the evidence, and the test procedure. They also discuss how AI conclusions should handle missing or conflicting evidence, and what an accountable reviewer must be able to challenge.
Later in the conversation, they cover configuration and third-party risk management. Michael explains why practitioners should define and test agent instructions, set limits of authority in workflows and permissions, and keep uncertainty visible during vendor assessments. MJ describes how Trustero customers use TI Playbooks and long-running agents to take on manual work while practitioners review and approve key steps.
Michael closes with practical guidance for GRC leaders: start with one recurring task, set a baseline, measure four angles of value, test difficult cases early, and ask vendors to demonstrate their capabilities using evidence from your own environment.
The session ends with a product walkthrough in Trustero, covering a segregation of duties control test with natural-language test procedures and proposed corrective actions, TI Playbooks for refining and building agent instructions, and an evaluation timeline for a third-party vendor assessment.
Tune in for a practical discussion on how GRC teams can put AI agents to work on control testing and vendor assessments while keeping people accountable for consequential decisions.


