Metadata vs Folders

The argument for metadata over folders is usually made abstractly and lands badly, because folders are not stupid — they are a reasonable model that fails at a specific point for a specific reason.

Understanding the reason makes the alternative obvious.

The single-path constraint#

A document lives in exactly one folder. Which means the folder structure has to encode, in one chosen order, every attribute anyone might use to navigate.

Consider a signed supplier agreement for the German subsidiary, expiring in 2027, owned by procurement. It could reasonably be filed by region, by supplier, by document type, by year, or by owning department. Whichever you choose, the other four are now unavailable as navigation.

So people compensate by nesting: Contracts / 2026 / Germany / Suppliers / Active. The tree gets deeper, the path gets longer, and every document now depends on somebody choosing the same order of attributes that the last person chose.

Why depth is the symptom#

Each level of nesting is one attribute encoded as a location. A five-level path is five attributes expressed in a fixed sequence, and it only works for someone who knows the sequence.

This is why folder trees grow faster than the content they organise, and why the same document ends up copied into two branches — not carelessness, but a rational response to a structure that can only answer one question.

What metadata changes#

Metadata makes attributes independent. The same document is simultaneously type supplier agreement, region Germany, status signed, owner procurement, expiry 2027. Nothing had to be chosen first.

That produces three concrete consequences.

Many routes to one document. Filtering by any combination of attributes reaches it, so the person who thinks in suppliers and the person who thinks in expiry dates both find it.

Rules that operate on meaning. Retention applies to a document type, not a path. Permissions apply to a confidentiality class. Neither breaks when the structure is reorganised, which folders cannot claim.

Views instead of copies. A saved filter for contracts expiring within ninety days is a view, not a folder anyone has to maintain, and it is never out of date.

Where metadata fails#

It fails when nobody fills it in.

A metadata model with twelve fields per document type produces documents with two fields completed, and partial metadata is worse than none because it looks like data and cannot be trusted for a rule. Three to five mandatory fields per type, with as many as possible derived automatically, is the design that survives.

It also fails when values are free text. Germany, germany, DE and Deutschland are four values as far as a filter is concerned. Controlled lists are not bureaucracy here; they are the difference between a filter that works and one that quietly misses records.

And it fails when the field list is designed by the people who want reports rather than by the people who file documents. Every field added for a report that could have been derived from an existing value is friction charged to the person doing the filing, and friction at the point of capture is paid every single day. The question worth asking of each proposed field is whether it could be inferred from the document type, the intake route or the content itself — and most of the time it can.

The structure that works#

Not metadata instead of folders. A shallow folder layer for permission boundaries and mental orientation — commonly one or two levels, by department or entity — with metadata carrying everything else.

The test is whether a folder level exists to express a boundary or to encode an attribute. If it is encoding an attribute, that attribute should be a field, because as a folder it can only ever answer one question and it will be reorganised eventually.

Migrating an existing tree#

The folder path is not noise. It is metadata, recorded in the only place the previous system offered.

Which means the migration mapping is usually mechanical: level one becomes department, level two becomes year, level three becomes document type. Deriving fields from paths during extraction is considerably cheaper than asking people to classify a decade of documents by hand, and it is usually more accurate, because the person who filed each document made the classification decision at the time.