Executive Summary
When an executive asks “who checks what the agents did?”, the instinctive answer is a person, at every step. It feels responsible. In practice it is the answer most likely to fail quietly. A reviewer who approves dozens of plausible, routine actions a day stops reviewing and starts clicking, and the organization ends up holding an approval trail that proves authorization and says nothing about scrutiny [7]. The other extreme, unsupervised autonomy, removes the accountability that boards, clients and regulators expect.
A third model comes from long-standing internal-control practice. One agent plans, another builds, an independent auditor tests the result against evidence, and a policy gate blocks a short list of actions that never run automatically. People keep authority over objectives, exceptions and irreversible decisions. The principle is old: ISO/IEC 27001:2022 control 5.3 requires conflicting duties to be separated [22][24].
The payoff is a pilot hypothesis, not a result: materially fewer human review minutes per completed task, at an equal or better rate of detected errors than manual review today. The architecture is conceptual, and no controlled baseline-to-outcome case of a full multi-agent consulting engagement was found. The decision this article asks for is a 90-day, single-workflow pilot, governed with ISO/IEC 42001 and ISO/IEC 27001 as the management scaffolding.
Introduction
Consider a consulting team that deploys agents to gather client documents, draft analyses and prepare a benchmark deck. Within a week the governance question surfaces. Approving every step caps throughput at the speed of an inbox, and oversight can degrade into rubber-stamping when the person is reduced to a confirmation step [7].
Why now? Agents are moving from drafting text to taking actions: sending messages, changing configurations, touching client data. The cost of an error rises with each tool an agent can call. Regulators are formalizing expectations at the same time. The EU AI Act requires that high-risk systems be designed so natural persons can effectively oversee them [31], and ISO/IEC 42001 gives organizations a structured set of policies, processes and controls for governing AI [32].
The business problem is precise. How can a firm grant agents more autonomy, and capture the speed that justifies the investment, without losing the control that makes the work defensible? The answer owes less to new AI techniques than to a mature idea from finance and security: no single actor should be able both to create and to conceal an error.
Main Body

The control logic is older than the technology
Practitioner summaries reduce the major control frameworks to one triangle: authorization, execution or custody, and recording or reconciliation belong with different people [27]. COSO names monitoring as one of the five components of effective internal control [25], and the Sarbanes-Oxley Act asks management of listed companies to maintain an adequate internal control structure for financial reporting [29].
ISO/IEC 27001:2022 states the principle in operational terms. Control 5.3 requires that conflicting duties and areas of responsibility be separated [22][23]. One guide offers two blunt examples: whoever initiates a transaction should not also approve it, and whoever writes code should not push it to production without a review [22]. Transferred to agents, the design rules nearly write themselves. The planner cannot execute. The implementer cannot approve its own output. The auditor cannot edit the deliverable. No agent can bypass the gate.
One caution outweighs the rest. Four agents do not create separation of duties. If they share a credential, the same transcript and the ability to alter evidence, the firm has staged a play with four actors and one set of keys. Real separation needs separate identities, separate permissions, append-only logs and enforcement outside the language model. Where full separation is impractical, substitute safeguards are accepted; one published self-assessment shows a sole-operator structure offset by an append-only audit log [18].
Autonomy is assigned by function, not by system
Early research proposed ten levels of human-computer decision-making, from fully manual to fully autonomous [7][10]. A later model argued for setting automation separately for each type of function, rather than as one dial for a whole system [13][14]. For knowledge work, that means four decisions, not one.
| Function | Default automation | Control | Consulting example |
|---|---|---|---|
| Gathering information | High | Approved sources, access scope, data classification | Agent searches authorized client repositories |
| Analyzing information | Medium to high | Traceability to sources, validation tests | Agent drafts a benchmark and cites each figure |
| Choosing an option | Medium | Human-set objectives and thresholds | Agent ranks alternatives; a partner chooses |
| Acting | Low for irreversible, high for reversible | Hard gate, rollback requirement | Agent may update a draft model; it may not email the client |
The SAE J3016 driving levels, six steps from no automation to full automation [1][11], offer a useful analogy: a system is autonomous only inside defined conditions. For a consulting agent those conditions are the client, the data classification, the tools and the ability to roll back.
A worked scenario: the client benchmark deck
An agent team prepares a competitor benchmark for a client. Gathering and analysis run autonomously inside authorized repositories, and every number carries a source link. The implementer assembles the deck in a sandbox, which is reversible by design. The auditor then checks each figure against the source documents, not against the implementer’s summary, and re-runs the calculations. The gate stops three things no agent may do alone: send the deck to the client, change access rights on the client folder, and edit the checklist that governs all this. A partner reviews only what the auditor flagged and the release itself. Most routine steps move without a human click, while the decisions that carry consequences still reach one.
A short hard-stop list beats case-by-case approval
Routine approval fails, so a short, explicit list of actions that never happen automatically should replace it. A defensible list covers production changes, destructive deletion, credential and privilege changes, external communication, financial or contractual commitments, and any change to the controls themselves. The CISO, together with risk, should own the list and review it at least quarterly.
Case-by-case approval yields inconsistent decisions, blurred accountability and a queue of low-value prompts. A hard list is predictable, testable and enforceable by a deterministic gate. It should describe the outcome, not the command: a rule against “database deletion” is too narrow if an agent can reach the same result by shortening a retention policy. Everything outside the list, provided it is reversible and bounded, proceeds autonomously with logging and sampled review.
Independent review must rest on evidence
The auditor is the component most often weakened in practice. An auditor handed the implementer’s reasoning and confident summary will tend to agree with it, simply because the summary frames the question. A credible design gives the auditor the task specification, acceptance criteria, final artifact, execution logs and independent access to source evidence, and withholds the implementer’s narrative until after a first independent pass. Evidence ranks in a clear order: external system state and deterministic tests, then signed source documents, then reproducible calculations, and far below them the explanation written by the party being checked.
Independence has a price. Published analysis catalogues recurring multi-agent failures [556], so add only the minimum independent control needed to catch the failure that matters.
The opposite failure: stopping too often
Governance can also fail in the other direction. Agents that treat every uncertainty as a reason to stop push routine decisions back to the human, and the result is alert fatigue and rubber-stamping aimed at the exception queue. Measure both directions: blocked hard-stop actions and seeded-error detection, along with unnecessary escalations, false blocks on known-safe work and review minutes per completed task. A seeded error is a mistake planted deliberately to test whether the auditor catches it. If review burden approaches the manual baseline, the architecture has failed even if no incident occurs.
What this costs, and when to stop
Build cost sits in four places: separate identities and permissions, immutable logging, the policy gate, and the auditor’s access to source evidence. Run costs add model usage for every extra agent and the time of control owners. No study in the library quantifies these costs, so break-even must come from your own baseline: reviewer hours saved per deliverable, multiplied by their loaded hourly rate, must exceed added model and control-ownership cost, with rework held flat. Set a stop-loss before the pilot starts. If review effort stays near the manual baseline, or seeded errors slip past the auditor, pause promotion and fix the design.
What regulators actually require
Article 14 of the EU AI Act requires that high-risk systems be designed for effective oversight by natural persons, with measures commensurate with risk, autonomy and context [31][45]. Nothing in that text asks a person to approve each intermediate step. What counts is practical capability: understanding the system’s limits, staying alert to over-trusting machine output, interpreting results, overriding or reversing them, and stopping the system. Application dates for the high-risk obligations have been shifting in legislation, so counsel should confirm the current date for your use case. Build the controls now regardless, because they are needed on any timeline.
ISO/IEC 42001 supplies the management layer: a structured set of policies, processes and controls governing how AI is designed, developed, deployed and used [32], including risk and impact assessment, transparency and human oversight across the lifecycle [33]. Certification evidences a management system. It says nothing about agent accuracy or return on investment.
ISO Alignment (Management Perspective)
ISO/IEC 42001: governing the agents
Management intent. It turns oversight from individual heroics into an auditable routine with named owners and continual improvement.
Minimum practices:
– Keep an inventory of agents, tools and autonomy levels.
– Name an oversight owner per workflow, with authority to pause and reverse.
– Assess risk before promoting an agent to a higher autonomy level.
Evidence and artifacts. An agent register, an authorization matrix, a signed hard-stop list and promotion records.
KPI. Seeded-error detection rate and the share of attempted prohibited actions blocked.
Risk and mitigation. Autonomy expands informally and nobody can say what an agent was allowed to do. A promotion gate that demands evidence before any increase mitigates this.
ISO/IEC 27001: enforcing the separation
Management intent. Control 5.3 gives the design a recognized anchor and a language auditors and clients already use [22][24].
Minimum practices:
– Issue separate identities and least-privilege permissions per role.
– Write immutable logs no operating agent can alter.
– Define substitute safeguards where separation is impractical.
Evidence and artifacts. Role-to-permission mapping, access reviews, log-integrity records.
KPI. Agents holding self-approval permissions, targeted at zero.
Risk and mitigation. Shared credentials silently collapse the separation; periodic automated conflict checks catch it.
ISO/IEC 20700 and ISO/IEC 21500 bear on engagement quality and delivery governance but are not applied here, to keep the compliance load readable.
Implications for the C-Suite
The decision to take now. Approve a 90-day, single-workflow pilot with a named executive owner, a measured baseline and explicit go/no-go thresholds.
A phased outline.
– Days 1–30: Security and risk write and sign the hard-stop list; delivery selects one workflow with a clear reversible/irreversible boundary; the team records manual cycle time, rework, quality and review effort.
– Days 31–60: Run shadow mode. Agents plan and audit while people execute, so you can compare machine and human decisions. IT provisions separate identities and append-only logging.
– Days 61–90: Allow reversible execution under the gate, with seeded errors to test the auditor.
Promotion rule. Promote autonomy when seeded-error detection holds and review minutes fall against baseline. Hold when results are mixed. Roll back on any unblocked hard-stop action.
State your assumptions. The pilot presumes a crisp reversible/irreversible boundary, working identity and logging controls, and client consent for agent involvement. If one fails, narrow the scope or fix the foundation first.
Manage the human shift. Consultants and reviewers move from approving steps to owning policies, exceptions and sampling. Track their review time and feed them seeded errors so the exception queue does not recreate rubber-stamping, and let skeptical partners inspect the logs themselves.
Fund independence, not agent count. If the answer to “how is the auditor different?” is “a different prompt”, separation does not yet exist.
Conclusion
Control does not come from asking a human more questions. It comes from designing a system in which no single actor can both make and hide a mistake. Finance and information security have relied on that idea for a long time [22][27].
For agents, the business value is practical. Routine, reversible work moves at machine speed, people spend their attention on decisions with irreversible consequences, and auditors and regulators see role-level evidence in place of reflexive approvals. The EU AI Act asks for effective oversight, and constant interruption is not part of that [31][45]. ISO/IEC 42001 makes the oversight repeatable [32].
The limits deserve equal weight. The architecture is a conceptual framework, the failure literature is a sober reminder that added agents add added risk [556], and no controlled production case of a full consulting engagement was found. The sensible path is a narrow pilot with a measured baseline, a short hard-stop list and a genuinely independent auditor. If the results hold, widen the domain. If they do not, the evidence will show where the design must change.
References
[1] edgepedia.org (2026). SAE J3016 — Edgepedia. https://www.edgepedia.org/sae-j3016
[7] votito.com. Levels of Automation. https://www.votito.com/methods/levels-of-automation/
[10] slideshare.net (2025). Automation and CNS-ATM-Human Technology Interface-ATSEP Competency-NOU-ng.pdf. https://www.slideshare.net/slideshow/automation-and-cns-atm-human-technology-interface-atsep-competency-nou-ng-pdf/284240902
[11] drivingresearch.com (2026). SAE J3016 — DrivingResearch. https://www.drivingresearch.com/standards/sae-j3016/
[13] onlinelibrary.wiley.com (2026). onlinelibrary.wiley.com · doi · pdfHuman Interaction with Intelligent Automation – Advances in Human. https://onlinelibrary.wiley.com/doi/pdf/10.1002/9781394266401.ch1
[14] onlinelibrary.wiley.com (2026). Is Our Future Colleague Even Human? Advancing Human–AI. https://onlinelibrary.wiley.com/doi/10.1002/job.70125?af=R
[18] ethicsportal.eu (2022). ISO/IEC 27001:2022 self-assessment | EthicsPortal. https://ethicsportal.eu/iso-27001/
[22] normwijzer.nl (2022). Annex A of ISO 27001: 93 controls (2022). https://normwijzer.nl/en/iso-27001/annex-a
[23] quldex.com (2026). ISO 27001 Requirements, Controls & Certification. https://quldex.com/frameworks/iso-27001
[24] isms.online (2025). How to Implement ISO 27001 5.3: Segregation of Duties That Pass Audits. https://www.isms.online/iso-27001/annex-a-2022/how-to-implement-iso-27001-2022-annex-a-control-5-3-segregation-of-duties
[25] coso.org (2026). Internal Control | COSO. https://www.coso.org/guidance-on-ic/pages/internal-control—integrated-framework.aspx
[27] gartsolutions.com (2026). ISO 27001 vs. SOC 2: Which Access Controls Do You Actually. https://gartsolutions.com/iso-27001-vs-soc-2-access-controls/
[29] manageengine.com (2025). SOX 404: Management’s guide to internal controls assessment. https://www.manageengine.com/compliance-manager/sox-compliance/sox-compliance-sec-404.html
[31] datenschutzgesetze.eu (2026). AI Act – Article 14 – Human oversight – DSG. https://datenschutzgesetze.eu/ai-act-14/
[32] iso.org (2026). ISO – ISO/IEC 42001 explained. https://www.iso.org/cms/live/live/en/sites/isoorg/home/insights-news/resources/iso-42001-explained-what-it-is.html
[33] elastic.co. Elastic achieves ISO/IEC 42001 certification. https://www.elastic.co/blog/elastic-achieves-iso-iec-42001-certification
[45] prighter.com (2026). Human oversight – High-risk AI systems. https://prighter.com/resources/laws/ai-act/articles/article-14-human-oversight
[556] Mert Cemri, Melissa Z. Pan, Shuyi Yang, Lakshya A. Agrawal, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya Parameswaran, Dan Klein, Kannan Ramchandran, & et al (2025). Why Do Multi-Agent LLM Systems Fail?. arxiv.org. https://arxiv.org/html/2503.13657
