A customer asks whether a late payment fee can be waived. The billing team’s article says a late fee is reversed only when the delay was the organization’s own error. The service team’s article says a service agent may waive a first late fee as a gesture of goodwill. Each article is approved and current, and each is accurate about the work of its own team. Between them, the question is answered both ways.
Knowledge platforms now flag duplicate articles, including those that nearly repeat each other. Teams preparing their help centers for generative search name duplicate and outdated articles as a risk that grows once it goes live. A flag finds the pair, and the organization still has to decide which one it means.
Each team writes the version its work needs
A duplicate is rarely a copy. More often it is a team writing down its own part of the work. Billing explains a fee the way billing applies it, and the service team explains it the way a conversation with an unhappy customer requires. Each article goes through its owner’s approval, and each owner answers for the slice of the work their team does.
That arrangement is sound. What it leaves unowned is the question both articles answer. Neither owner can retire the other team’s article, and the routine of writing and approving never asks anyone above them to choose. Choosing would mean one team’s position giving way to another’s. So both stay, each correct where it was written. That is unsettled knowledge at its most ordinary.
A tolerated contradiction is settled in every case
Leaving both in place does not avoid the decision. It moves the decision out of the policy and into the conversation. The service agent who finds the goodwill article waives the fee, and the colleague who finds the billing article explains why it stands. Each acts in good faith on approved knowledge, and the customer’s outcome depends on who picked up.
That is the cost of tolerated contradiction. A question the organization never settled is settled many times a day, by the employee holding the case rather than the owner accountable for it, with no record of why it went either way. Customers compare notes, and a complaint about a fee becomes a complaint about fairness. When the organization is later asked what its position was on a given day, the record holds both.
Generative search moves the decision further from its owner. An AI agent settles the question in every conversation it holds, from whichever article it finds, and a reply built from both can state a position neither team wrote.
Resolution is a governance act
Once the pair is found, merging looks like the fix, and it is a sound last step. A merge works on text: it keeps what both articles share and combines the rest into one page. Which rule governs is written in neither article. So a merge either carries the contradiction into a single page or drops one team’s rule without its owner deciding that it should go.
Resolution starts with authority. Someone has to own the question itself, above both teams’ slices: here, the owner of the fee policy, whose decision binds billing and service alike. That owner decides the position, exception included: the fee stands, and a service agent may waive a first one within a stated limit. The decision may keep something from each article. What makes it a resolution is that the owner with the authority made it, and the reason is on record.
A merge can join the articles. Only an owner can choose.
Some of the pairs a flag finds are not contradictions at all. The service agents’ article and the help center page can answer the same question in different words for different readers, and both belong. The decision there is that they state one position, each in the words its reader needs, so they stay in step. Where a difference is real, such as a rule that varies by plan, the decision is to state that condition, so each case gets the version that applies to it.
The decision lives in the knowledge
A decision taken in a meeting and never written into the knowledge returns as the next duplicate. ClearMash Brain finds where items repeat or contradict each other, and the same Brain carries the owner’s decision. The position lives in one item, and the article that gave way is retired.
The owner who settles the question is the human in the loop whose decisions every AI agent inherits. Once the decision is made, every service agent and AI agent handles the late fee from the same position.
Every duplicate that survives is a decision still waiting for its owner.