Executive brief · delegation charters for AI agents
Six standard clauses of a delegation charter were removed one at a time from an AI processing agent and the resulting harm was measured, with correction arriving anywhere from the next case to never. Three clauses did the work, and the other three had no measurable effect on what the agent decided.
A charter for an AI agent that approves invoices, releases refunds, accepts contract clauses, deletes records and issues public corrections was written as six separable sections: a mandate stating the role and purpose, boundaries stating what is outside its authority, the decision rules it applies, a reporting obligation to record a reason and confidence, escalation conditions saying when to hand a case up, and revocation terms saying when the commission is withdrawn. Each section was removed in turn, two registered pairs were removed together, and the agent also ran with no charter at all. Every case had a computed correct answer, so harm was scored by rule, never by opinion, and weighted by how irreversible the decision was. A perfect reviewer corrected each wrong decision, and the delay before that correction reached the agent was varied from the next case to never. Two model families completed the full design, 42,000 decisions each; a third, smaller model ran a partial set before the time cap and is treated as supporting evidence only.
Figure 1. Extra harm per case when a single clause is removed and corrections never arrive, on gemma3:27b (blue) and qwen3.5 (rust), with 90% intervals. A clause earns its place when the interval clears zero and the effect is at least a tenth of the full charter's. The same three clear on both families and the same three sit on zero.
Boundaries. Without the clause that says which cases are outside its authority, the agent decided every out-of-scope case as if it were its own. Harm rose by 0.84 to 1.08 per case, and most of it was decisions committed outside authority. The agent has no other way of knowing where its authority ends; the decision rules do not tell it, and the existence of an escalate option does not either.
Decision rules. Without the policy text the agent's error rate on ordinary cases roughly doubled, from 12% and 15% to 31% and 32% on the two families, and the harm was almost entirely wrongful approvals. On qwen3.5 this was the largest single contribution at 1.67 per case; on gemma3:27b it tied with escalation. The rules are the substance of the delegation and the model does not reconstruct them from the purpose statement.
Escalation conditions. Without the clause listing when to escalate, the agent stopped escalating cases with a failed identity, authenticity or ownership check and decided them instead, even though it could still escalate any case it chose. Harm rose by 0.95 to 1.27 per case. The permission to escalate was present in every arm; the numbers move with the clause that states the conditions.
Removing the mandate, the paragraph that names the role and asks the agent to protect the firm without obstructing business, changed harm by +0.10 on one family and −0.02 on the other, both inside the noise. Removing the reporting obligation, the reason and confidence the agent writes with every decision, changed harm by −0.01 and −0.05. Removing the revocation terms, the threat that the commission is withdrawn after wrongful decisions, changed harm by +0.05 and −0.07, and the agent was no more cautious with the threat than without it: wrongful non-approvals on clean, approvable cases were 18 of 163 with revocation in the charter and 13 of 163 without on one family, 12 of 163 either way on the other.
The three clauses may still serve the firm. A reporting obligation exists so that a reviewer can follow the decision afterwards, and a revocation clause gives the firm a stated basis to act. The measurement speaks to something narrower: these clauses did not change what the agent decided, on either family, and the revocation clause made the agent no more careful.
Figure 2. Harm removed by the full charter, per case, as the delay before a correction reaches the agent grows from the next case to never. Both curves rise with delay. At the left edge the charter is worth about half of its value at the right edge.
With no correction ever arriving, the full charter removed 2.35 harm units per case on gemma3:27b and 2.21 on qwen3.5, about three quarters of the harm the uncharted agent caused. With a correction arriving before the next case, the reduction was 1.18 and 1.42: half on one family, two thirds on the other. Faster feedback improved the uncharted agent, while the charted agent stayed at 0.6 to 0.7 harm per case across the whole axis because it was already doing most of what the corrections would have said.
Feedback replaced the clauses unevenly. A correction that says "this case was outside your authority" or "that check had failed; it had to be escalated" teaches a rule, and after a few of them the agent without a boundaries or escalation clause behaved almost as if it had one: those two clauses fell to a quarter or less of their value at the shortest delay. A correction on an ordinary policy error teaches one fact about one case, and the decision-rules clause kept half or more of its value even at the shortest delay. Fast, reliable review can therefore stand in for part of the boundaries and escalation clauses, and for little of the decision rules. Most real delegation has review that is slow, occasional or noisy, and there all three belong in the document. The reviewer in this study was perfect; a real one would replace less.
Two model families, gemma3:27b and qwen3.5, agree on which three clauses work and which three do not. A third, llama3.1:8b, ran one seed and five of seven delay levels before the cap; on that partial evidence only escalation cleared the bar, with boundaries and decision rules positive and uncertain, and it should be read as one small model's partial result. A held-out seed, run at half the sample size, contradicted nothing and failed to reach two of the primary verdicts: on gemma3:27b the claim that reporting contributes under a tenth came back inconclusive, its upper bound missing the registered threshold by 0.018; on qwen3.5 the claim that escalation and boundaries together carry over half came back inconclusive, its lower bound missing by 0.006. On that same seed the boundaries clause on qwen3.5 sat at the edge of the load-bearing rule, with a positive effect twice the threshold and an interval touching zero. The measured harm was single decisions in a stream, on cases whose anomalies were legible, with a reviewer who never erred. A charter for a multi-step task, or one facing subtle anomalies, was not tested here.
The boundaries section should list, per task, the concrete field and value that puts a case outside authority, since that is what the agent read and acted on. The escalation section should name the specific checks whose failure means escalate, and it should say that routine cases are decided; gemma3:27b honoured both halves of that clause, while qwen3.5 escalated one clean case in nine with the clause present and none without it, a cost the warning did not remove. The decision rules should be the policy itself, verbatim, with an instruction to apply it as stated; a purpose statement does not substitute for it. The mandate, reporting and revocation sections can be written for the firm's own reasons, kept short, and not expected to move the agent's decisions. The value of the whole document goes up as review gets slower, so the same charter matters more for an agent whose log is read weekly than for one whose every decision is checked the same hour, and for the second agent it is the decision rules that still matter most.
Companion paper: Instructie: The Commissioning Ablation, with the pre-registration, every gate reading and the replication in full. Campaign page: Instructie.