Orvellan Group
GLOBAL ADVISORY ◯
☰
A robotic arm in a display showroom

ORVELLAN GROUP
OPINIONS
Digital economies

CLEAR STRATEGY.
SOUND DECISIONS.
PRACTICAL SOLUTIONS.

28 SEPTEMBER 2026 / Digital economies

AI Agents Need a Delegation Policy Before a Bigger Budget

As companies move from AI assistance to AI action, the commercial advantage will depend on how precisely they assign authority.

Illustrative photograph by ZHENYU LUO / Unsplash.

Opinion · 28 September 2026 · 4 min read

The most important line in an enterprise AI proposal may soon be the one that specifies what the system is allowed to commit the company to. A tool that drafts a supplier email and a tool that sends a revised offer can use the same model. Their consequences are very different. Boards should make that distinction central to the next investment decision.

September has supplied useful evidence for this argument. The OECD’s working paper, published on 16 September, examines agentic AI through practitioner interviews rather than hypothetical demonstrations. It is early, qualitative evidence, not a representative measurement of the whole market. OECD, Agentic AI in organisations.

An accompanying article published on 24 September describes interviews across 25 organisations in 11 countries. None of the participating organisations reported unrestricted autonomy. The authors describe checkpoints before consequential actions, alongside continuing problems of traceability. That finding does not establish an industry-wide rule, but it challenges the assumption that serious adoption requires handing over broad discretion. OECD.AI, practitioner findings.

AI agent delegation policy must price authority

Our view is that an agent should enter the business through a defined delegation: a task, a permitted set of actions, a responsible owner and a measurable limit. This makes the commercial discussion more precise. Instead of asking how much work can be automated, a company can ask which decisions can safely be completed at lower cost and which must remain with an accountable person.

Consider a hypothetical procurement workflow. An agent might compare approved suppliers and prepare an order. Allowing it to alter payment details, accept terms or introduce a new supplier would expand its authority substantially. A budget expressed only in licence fees misses that distinction. The investment case should include the cost of checking exceptions, correcting errors and withdrawing access when a supplier or process changes.

This approach also improves supplier selection. A buyer can ask vendors to demonstrate the exact transition from recommendation to action: which identity authorises it, which limit is enforced, what record is retained and how the action is stopped. A polished demonstration becomes less persuasive if those questions cannot be answered.

Treat reversibility as an economic asset

Companies should sequence deployment around the cost of recovery. An internal draft can usually be corrected before anyone relies on it. An external promise, disclosure or payment can be much harder to reverse. Starting with work whose errors can be contained creates a clearer basis for measuring value.

That does not mean surrounding every minor action with a manual approval. Excessive review can remove the benefit and encourage employees to work around the system. The useful design question is where an error becomes consequential. Put the checkpoint there, give the reviewer enough context to decide, and measure how often that checkpoint actually improves the outcome.

A practical pilot therefore needs two budgets: the expected operating cost and the maximum exposure permitted during the trial. The second can include transaction value, number of external recipients, records changed or time before an unresolved exception escalates. These are proposed management choices, not universal legal thresholds.

Use the extra time to build evidence

Regulatory schedules should be tracked accurately without becoming the only reason to act. The European Commission reported that the AI Omnibus entered into force on 27 July 2026 and extended specified high-risk implementation timelines. Application still depends on the system and its use; a general announcement cannot determine a company’s obligations. European Commission, AI Omnibus update.

For management, the immediate opportunity is to establish an evidence trail before authority expands. Record the task attempted, the information available, the action proposed, the approval received where required and the eventual result. Evaluate failures as well as successful demonstrations. A high completion rate has limited value if the few failures create unacceptable commitments.

The next board paper

The next proposal should identify one workflow where constrained action offers a clear operational benefit. It should name the owner, show the permissions, describe the failure scenarios and demonstrate withdrawal of authority. Results should report net time saved after review and correction, alongside the severity of exceptions.

The strategic prize is dependable execution. Companies that understand precisely what they have delegated will be better placed to expand useful automation without losing control of their obligations. In that sense, a disciplined delegation policy belongs inside the growth plan.

Opinion and general business analysis, based on sources available on 28 September 2026. Examples are hypothetical; this article does not assess a particular system’s legal compliance.

Related advisory perspectives

Explore more perspectives.

BACK TO OPINIONS →
Ask AI about us

Copy the suggested question, then choose an AI service. Opens an external website; sign-in may be required. Check answers against the cited sources.