Provenance Belongs in the Product Workflow

The current debate about AI watermarks is too focused on whether generated content can be marked. The more useful product question is whether that mark helps someone make a decision.

A watermark can indicate that a model was involved. It does not explain which source material was used, what a person changed, whether the output was reviewed or who is accountable for publishing it. It can also lose its meaning when content moves between tools, gets copied into another format or is deliberately altered. We are trying to solve a workflow problem with a label.

This becomes clearer when looking at several developments together. AI companies are experimenting with visible and invisible markers while users are already looking for ways to remove or avoid them. Security incidents remind us that even a valid account does not guarantee that its owner performed an action. Product research continues to show that something can be technically useful and still fail in a real operating environment. At the same time, algorithms increasingly influence the form of the devices and interfaces through which people consume the result.

The common issue is provenance. Not provenance as a hidden technical property, but as information people need to judge whether an output belongs in their work.

I see a direct parallel with product analytics. A dashboard can show a precise number while hiding a weak event definition, a broken tracking implementation or a change in the population being measured. Adding more metrics does not create confidence. The number becomes useful when the team understands where it came from and which decision it can support.

AI output should be treated the same way. If a sales team uses an AI-generated account summary, the important information is not merely that AI was used. The user needs to know which customer records contributed to it, when those records were last updated and whether a colleague corrected the summary. If a support agent receives a suggested response, they need to see whether it is grounded in an approved policy or assembled from an old conversation. If a product manager generates a research synthesis, the team should be able to trace a conclusion back to the underlying interviews instead of trusting a polished paragraph.

This changes the product requirements. A provenance feature is not a badge in the corner. It is a chain connecting source, transformation, review and action. Different parts of that chain should appear at different moments. A reviewer may need a detailed history. An end user may only need to know that an output was checked against current approved material. An administrator may need to investigate unusual activity associated with an account. Showing everyone the same generic disclosure is easy to implement, but rarely useful.

There is also an uncomfortable growth implication. Extra context can create friction. It may make an AI feature feel less magical and expose uncertainty that a clean interface would otherwise hide. Teams under adoption pressure will be tempted to reduce provenance to a small icon so it does not interrupt the flow.

I think that is a mistake. Hiding uncertainty may improve the first interaction, but it makes repeated use harder to sustain. In B2B products especially, adoption depends on whether the output can survive review by a manager, customer, legal team or operator. A result that cannot explain itself creates work elsewhere in the organization, usually through manual checking, screenshots and side-channel approvals.

I would therefore test provenance as part of the core journey. Can users identify why a result exists? Can they inspect the relevant source without leaving the task? Can they tell what changed after generation? Can the organization establish who approved the final action? These are observable workflow outcomes, not abstract claims about trust.

Watermarks may still have a role, particularly when content leaves the product. But they are the outermost layer. Product teams should not confuse marking an output with making it accountable. The stronger product is not the one that merely announces the use of AI. It is the one that preserves enough context for people to use that AI output responsibly in real work.