Imagine a company that gives three departments their own AI agents. Sales configures its agent to improve conversions and authorises it to offer selected concessions. Finance configures another to protect margins and reject spending outside agreed limits. Customer support uses a third to resolve requests within a fixed service allowance.A customer accepts an offer that includes extra onboarding. Finance rejects the additional delivery cost. Support cannot provide the promised assistance within its allowance. The customer receives three responses from what was supposed to be one company.In this hypothetical example, each agent follows a locally reasonable instruction. Together, they expose an argument the leadership team has never settled: who can make a promise that another department must fulfil?This is how office politics can enter an automated workflow. Conflicting priorities become instructions, permissions and targets. Once those rules govern action, the disagreement can repeat across every transaction that encounters it.Every objective contains a choiceAn instruction such as “reduce support costs” leaves considerable room for interpretation. Should the system resolve a problem thoroughly, shorten the conversation or make escalation harder? Those approaches have different consequences for customers and employees.Choosing a metric gives one interpretation an advantage. If the team is rewarded for closing requests quickly, time spent understanding a complicated problem becomes harder to justify. The measurement helps determine which behaviour looks successful.The same applies to instructions given to agents. A leader who approves an objective is also approving a set of priorities, whether those priorities have been discussed openly or left implicit.I think this is an important limit of the idea that automation makes management objective. A system can apply a rule consistently while preserving a choice that deserves to be challenged.Cooperation includes the exceptionsReturn to the onboarding example. A sales manager might previously have called a colleague before making the offer. The colleague might have agreed to help because the customer had unusual needs, because capacity was available or because sales had supported their team during an earlier problem.That arrangement contains knowledge about timing, relationships and reciprocal obligations. Some of it can be documented. Some requires a conversation about the particular situation.There are good reasons to replace informal arrangements with clearer rules. Personal favours can create unfairness, hide costs and leave decisions dependent on who knows whom. But removing the conversation without replacing its function creates another problem: the company loses a way to negotiate competing demands.For a team combining people and AI, cooperation needs an explicit place in the design. Someone must be able to recognise a conflict, bring the affected parties together and agree what happens next.Test the disagreement before deploymentBefore connecting departmental agents, leaders should examine a shared situation from several perspectives. Use a customer request that is commercially attractive but difficult to deliver. Ask each team what its agent would be instructed and permitted to do.Then compare the proposed actions. Does one agent create work that another rejects? Can a discount trigger a service commitment? Does anyone have authority to make an exception, and will the affected team know about it?These questions reveal where the organisation needs an agreement before it needs another automated step.For the onboarding case, the answer might be a service allowance that sales can commit without further approval. Larger promises could require confirmation from delivery before reaching the customer. The important work is deciding the boundary together and ensuring that each workflow respects it.Repeated exceptions should prompt a review of the agreement. Otherwise, the person handling escalations becomes responsible for repairing a contradiction that leadership has left in place.Letting agents negotiate with one another still requires a human agreement about acceptable outcomes. A system that finds a compromise needs boundaries around what it can concede, who bears the cost, and when consent is required first.Give people a role in setting the rulesEmployees who handle difficult cases should help shape these decisions. They can explain where a target encourages the wrong response, which exceptions recur and what customers understand differently from the company.Their involvement also changes the meaning of accountability. Asking someone to supervise an agent is more credible when they can question its instructions and propose changes. Responsibility becomes difficult to sustain when a person must defend outcomes produced by rules they cannot influence.This is where human capability matters beyond checking whether an AI answer is accurate. A strong colleague can recognise another team’s constraint, explain a competing interest and negotiate an outcome that others can support. Those abilities help an organisation decide what it is trying to achieve together.A company’s use of AI will reflect its choices about authority, trust and whose experience deserves attention. Before delegating more work, leaders need to make those choices visible enough to discuss and revise. The rules that agents follow will become part of the working environment people must live with every day.