Best Agentic AI Consulting Companies

Governing AI agents: approvals, permissions and audit

An agent that can send an invoice can also send the wrong one. Governance is the written answer to three questions, set before launch, about what each agent may do alone, what it must ask about and how anyone will know what it did.

By The editors · Updated · 5 min read

A purchasing manager on the phone initials a printed purchase order on a clipboard while a colleague waits beside his desk with a second order.
Approval thresholds for agents usually mirror the ones a company already sets for people.
Contents
  1. The accountability gap
  2. What the frameworks ask for
  3. Five controls every agent needs
  4. How the firms implement them
  5. Questions to put to any firm
  6. Setting up governance for a first agent
  7. Questions buyers ask
  8. Sources

Most companies already know how to govern delegated work. Staff have spending limits, approval chains and a record of who signed what. AI agents need the same structure, written down before they start, and many companies are letting agents act before it exists.

The accountability gap

Deloitte's survey of 3,500 senior leaders found that 40% let AI execute decisions or actions without human approval to at least a moderate degree. For 23%, accountability for a negative outcome remained unclear⁠1. That combination, autonomy without an owner, is the risk agent governance exists to close.

Agents also bring security risks that chat tools do not. The OWASP Gen AI Security Project now publishes a separate Top 10 for agentic applications, peer-reviewed by more than 100 experts, covering systems that plan, act and make decisions across workflows⁠2.

What the frameworks ask for

Four public frameworks are the usual reference points. None is written for a particular vendor.

FrameworkStatusWhat it asks for in practice
EU AI Act, Article 14Applies to high-risk systems from 2 December 2027 (Annex III) and 2 August 2028 (Annex I)⁠3People assigned to oversight can decide not to use the system, override or reverse its output, and stop it through a "stop" button or similar procedure⁠3
NIST AI Risk Management FrameworkVoluntary; released 26 January 2023 and being revised⁠4A structured way to identify and manage AI risks, with a separate generative AI profile⁠4
OWASP Top 10 for Agentic ApplicationsPublished for 2026⁠2The most critical security risks for agents, with practical mitigations
Singapore Model AI Governance Framework for Agentic AILaunched 22 January 2026⁠5A government framework written specifically for agentic AI

Most agents in a mid-sized company's back office will not be high-risk systems under the EU AI Act. Article 14 is still a useful design standard, because its requirements (override, reversal and a safe stop) are the ones any operations director would want.

Five controls every agent needs

The minimum set

  • An identity and its permissions. Each agent has its own identity and access to named tools and data, not a shared administrator account.
  • Autonomy limits. For each action, a written rule: the agent does it alone, asks first, or never does it, with thresholds in money or risk.
  • Approval routing. Requests above a limit go to a named person, in the place that person already works.
  • An action log. Every action is recorded with its inputs, sources and result, where the responsible team can read it.
  • A stop procedure. A way to pause the agent, and a defined state it returns to.

How the firms implement them

Firms describe these controls with very different levels of detail.

LeewayHertz publishes the most complete list. Its agents are configured with defined identities and permissions, access to approved tools and data, autonomy boundaries, human-approval requirements, escalation paths, logging, auditability and stop controls, and its run phase tracks policy events in production⁠6.

Deloitte treats controls as a workstream that runs alongside each Zora AI implementation, in five phases from a pre-design risk assessment to monitoring a stabilised environment, including human oversight models and alignment with internal and external auditors before cutover. It also states plainly that Zora AI does not come with SOX-related controls built in: controls must be designed for each organisation's processes⁠7.

OpenLazarus expresses limits per teammate in business terms. The company decides, for each teammate, what it handles on its own and what it asks about first. Its industry pages give examples: in wholesale, a purchase order over $15,000 or a credit over $500⁠8; in manufacturing, a material order over $20,000 or a change to a promised ship date. Every action is logged where staff can see it⁠9. Its FAQ answers the question of agents acting without asking in one line: only where the company allows it⁠10.

IBM Consulting places governance in the orchestration layer of its agentic apps, where workflows are executed, monitored and controlled⁠11. Cognizant offers Neuro AI Trust, a central view of risk across every model, agent and application⁠12. Accenture's Trusted Agent Huddle scores agent performance with a proprietary algorithm, which it describes as the basis for a future agent trust score⁠13.

Questions to put to any firm

ControlQuestionA good answer
PermissionsWhich systems and records can each agent read and write?A list per agent, enforced by the system, not by the prompt
LimitsWhere are the approval thresholds written, and who can change them?In configuration the client controls, changed by a named owner
ApprovalsWhere does an approval request appear?In the approver's normal tools, with the evidence attached
LogWhat does the log record, and who can see it?Inputs, sources, action, approver and result, readable by the business team
StopHow do we pause an agent, and what happens to work in progress?A documented procedure, tested before launch
AccountabilityWho answers for an agent's mistake?A named business owner for each agent

Setting up governance for a first agent

  1. Start from the approval matrix you already have. The spending and sign-off limits that apply to staff are the natural first draft for an agent's limits.
  2. List the agent's actions. Every action it can take on a record, one line each.
  3. Classify each action. Alone, ask first, or never, with a threshold where one applies.
  4. Name an owner. One person in the business is accountable for the agent's work, as they would be for a team member's.
  5. Test the controls. Before launch, check that a request above a limit reaches the approver, that the log captures it and that the stop procedure works.
  6. Review monthly. Read a sample of the log, check exceptions and adjust limits as confidence grows.
An external engineer walks two members of a company IT team through a tabbed run book at a conference table.
Governance only holds if the company's own team can read the log and change the limits.

When governance is working, an agent's limits read like a job description, the approver sees each exception with its evidence, the log answers "who did this and why" in a minute, and raising a limit is a decision a manager makes on evidence from the log, not a request to the vendor.

Questions buyers ask

What is AI agent governance?

AI agent governance is the set of rules and controls that decide what an AI agent may do on a company's behalf, which actions need human approval, who is accountable for the outcome, and how every action is recorded so it can be reviewed and reversed.

Do AI agents need human approval?

For consequential actions, yes. A common pattern is to let an agent act alone below a threshold, such as a purchase order under a set amount, and to require a person's approval above it. EU rules for high-risk AI systems require that people can override an output or stop the system.

What should an AI agent audit log contain?

For each action, the agent, the time, the inputs and sources it used, the tool or system it acted on, any approval given and by whom, and the resulting change to the record. The log should sit where the people responsible for the work can read it.

Sources

  1. The State of AI in the Enterprise. Deloitte US. Accessed 8 Nov 2026.
  2. OWASP Top 10 for Agentic Applications for 2026. OWASP Gen AI Security Project. Accessed 8 Nov 2026.
  3. Article 14: Human oversight. EU Artificial Intelligence Act (Future of Life Institute). Accessed 8 Nov 2026.
  4. AI Risk Management Framework. NIST. Accessed 8 Nov 2026.
  5. Singapore launches new Model AI Governance Framework for Agentic AI. Infocomm Media Development Authority, 22 Jan 2026. Accessed 8 Nov 2026.
  6. Enterprise AI agent development company. LeewayHertz. Accessed 8 Nov 2026.
  7. Zora AI: agentic AI digital workers. Deloitte US. Accessed 8 Nov 2026.
  8. OpenLazarus for wholesale and distribution. OpenLazarus. Accessed 8 Nov 2026.
  9. OpenLazarus for manufacturing. OpenLazarus. Accessed 8 Nov 2026.
  10. Questions about OpenLazarus. OpenLazarus. Accessed 8 Nov 2026.
  11. IBM introduces new AI Integration Services to help enterprises build and scale AI agents. IBM Newsroom, 20 Feb 2025. Accessed 8 Nov 2026.
  12. Build and scale agentic AI. Cognizant. Accessed 8 Nov 2026.
  13. Accenture introduces Trusted Agent Huddle for multi-system AI agent collaboration. Accenture Newsroom, 28 Apr 2025. Accessed 8 Nov 2026.

Updated 8 Nov 2026: First published.

Further reading