Article

Digital asset management and AI: enriching assets with human review

How I built AI enrichment in a DAM: images, text, PDFs and videos with audio, confidence scores and human review before enabling semantic search.

Published 12 November 20265 min readAIDAMSemantic enrichment

For BaseCamp Studio I developed the AI enrichment part of the DAM, the system that organises digital content. I worked with images, text, PDFs and videos with audio. An extraction result does not immediately become asset data: a person reviews it first. Here I explain that boundary, because an AI-generated suggestion is not yet verified information.

Digital asset management and AI: start with the source

I developed an algorithm that extracts information from the different source types and categorises it. I do not treat every file as the same input. The type matters when understanding where information comes from and what needs checking. The common goal is to produce semantic enrichment information: information about the content that must be assessed before it can be used in search.

  • Images: the source to examine is visual content. Extraction proposes information to check against the image, not a truth to accept automatically.
  • Text: the starting point is written content. Even a plausible categorisation needs to be compared with what the text actually says.
  • PDFs: these are among the sources I handled. The file format alone does not establish whether extracted information correctly describes the document.
  • Videos with audio: the work also covers this source. I mean audio contained in videos, not support for standalone audio files.

What gets extracted and what still needs deciding

The output is a set of information and categorisations to enrich the asset semantically. It is not just a new filename. Its purpose is to describe content in a way searches can use. I do not promise a universal list of fields for every archive: first you need to establish which information matters and whether the source lets you verify it. Extracting something and being able to use it are two separate steps.

The confidence percentage does not approve the information

Each piece of information receives a confidence percentage. In the workflow I built, that indicator supports human review. I do not describe it as a certified measure of the probability that the information is true. A percentage does not remove the need to compare the suggestion with its source: a description can sound convincing and still be wrong. I am not claiming a threshold above which the system approves information by itself.

A person checks whether the proposed information is true, using the percentage as supporting context, and decides whether to merge it. By merging I mean integrating that information into the asset. Only after this step is the asset enriched and that information enabled for semantic search. If a suggestion is not approved, it must not be treated as enrichment already available for searching. Review is not a correction to postpone until later.

I also built the system with a vector database that supports searching by meaning. The point here is which information we allow it to use, not the details of how a search runs. For that part, see my case on semantic search in the DAM. A result found by meaning is only useful if the information making it searchable is trustworthy.

When to consider this workflow for your company

I would consider it if you have different kinds of content to find, useful information inside files and people who can verify it. Before choosing models, I would ask which searches you cannot perform today, which source lets you check a suggestion and who decides to accept it. If nobody can verify the information, automating extraction does not fill that gap. I do not promise savings or accuracy I have not measured.

If your main problem is scattered product information, start with my article on PIM for small businesses: managing product records and enriching assets are connected but different responsibilities. To explore where AI can fit into your process, see my AI consulting service. You can contact me with an example of your content and the search you cannot do: I would start there, not with a promise about a model.

Let's talk

Tell me what's blocking you.

30 minutes, free. Afterwards I'll send you, in writing, the first 3 things I'd do. Or email me: the person doing the work reads and replies.