Quick answerWhen two authorized users on the same account give contradictory instructions in good faith, the agent should not silently pick one, silently pick whichever came most recently, or silently split the difference. It should pause the specific action in conflict, tell both parties plainly what it was told and by whom, and require either a role-based tiebreaker rule you defined in advance (owner outranks member, for example) or explicit reconciliation between the two humans before proceeding. Everything else on the account that is not touched by the conflict should keep working normally.
This is a narrower problem than it sounds
A lot of contradiction-handling advice is written for cross-department disputes, where two different parts of a company disagree about policy, or for a customer who is deliberately trying to manufacture an exception. Neither applies here. This is two people who both have legitimate authority on the same account, acting in good faith, who happen to have told the agent to do two different things, usually without knowing the other one spoke to the agent at all.
This shows up constantly in small businesses and team accounts: a manager tells the agent to always apply a discount code, an owner later tells it never to apply that code without approval, and neither one knew the other had weighed in. The agent now holds two live, contradictory, equally authorized instructions.
Step one: detect that a conflict exists at all
The hardest part of this problem is not resolving the conflict, it is noticing one exists. Most agents process each conversation in isolation and have no mechanism to check whether an instruction from user B contradicts a standing instruction from user A given three weeks earlier in a different thread. Detecting this requires the agent to maintain instruction state per account, not just per conversation, and to check new instructions against that state before acting, not just against the current conversation's context.
Step two: do not pick a winner by default
Three defaults are tempting and all three are wrong on their own. Most-recent-instruction-wins seems reasonable until you realize it means whichever authorized user happens to talk to the agent last silently overrides everyone else, with no visibility for the person who gets overridden. Most-senior-role-wins is closer to right but only works if you defined role seniority for this purpose in advance, most account systems have roles for permissions, not for instruction-conflict tiebreaking. Splitting the difference, applying some blended version of both instructions, produces an outcome neither person actually asked for and is often worse than either original instruction.
The protocol
Pause only the specific action in conflict, not the whole account. Notify both parties in plain language: state what instruction A said, what instruction B said, and that they contradict each other on this specific point. If you have a predefined role-based tiebreaker (owner outranks member is the most common), apply it and tell both parties you did so and why, so the overridden party is not left wondering what happened. If no tiebreaker rule exists for this account, or the conflict is between two people of equal authority, hold the action and ask both parties to reconcile directly rather than the agent guessing on their behalf.
Why role-based tiebreakers need to be set up before they are needed
Waiting until the first live conflict to decide who outranks whom means the agent has nothing to fall back on in the moment it matters most. Define instruction-authority tiers as part of account setup, separately from feature-permission tiers, since the two do not always match: a finance-only role might have narrow feature access but should outrank a general member on any instruction touching billing or discounts.
How this differs from nearby problems
This is not the same as a live arbitration protocol for when two of your own AI agents contradict each other mid-conversation, that post is about machine-to-machine conflict, not two humans. It is also not what happens when a customer deliberately manufactures a contradiction to force an exception, which assumes bad faith from a single person working the system rather than two legitimate people who simply disagree.
FAQ
What if the two authorized users are the same seniority and cannot agree? The agent should not decide for them. Hold the action and require one of them to explicitly rescind or supersede the other's instruction, with that reversal logged the same way the original instruction was.
Should the agent log every detected conflict even after it is resolved? Yes. A resolved conflict is still a signal that your account's permission or instruction-authority model has a gap, and a pattern of repeated conflicts on the same account is worth a human account manager reviewing directly.
Does this apply to conflicting preferences, like tone, not just operational instructions? The same detection and pause logic applies, but the bar for pausing should be lower for anything with financial or access consequences than for a stylistic preference, where quietly defaulting to the most recent instruction and noting the change is usually low enough stakes to be acceptable.