Everyone agrees on how to tag design assets in theory: consistent vocabulary, a few tags per asset, applied at import. In practice, tagging is homework — the step everyone skips the week a launch slips. So the real question isn't what good tags look like; it's how to get them without asking humans to do work humans reliably won't do. The split that works: people design and police a small taxonomy, machines write the first draft of every asset's tags, and a lightweight review keeps the two honest with each other.
TL;DR
- Tagging fails as a habit, not as a concept. Design the workflow around the skipping.
- Humans own the taxonomy: 3–4 categories, 5–15 tags each, merged quarterly.
- AI drafts tags at import — content, style, even the client from a wordmark. Suggestions, not verdicts.
- One review pass per asset: approve or edit in seconds. The edits are training data.
- Never hand-tag a backlog. Batch-scan it and review the suggestions instead.
01 — The failure mode
Why tagging fails
Three ways it dies. First, the chore: tagging at import feels optional, so under deadline it doesn't happen, and untagged assets are invisible to every filter forever after. Second, the drift: without a shared vocabulary, five people invent five spellings — “social-post,” “Social Post,” “socials” — and each filter shows a fifth of the truth. Third, the flood: a zealot adds fourteen tags per asset, search returns everything, and tags stop meaning anything.
Notice all three are human-behavior problems. That's the case for splitting the work: machines are tireless about the part humans skip, and humans are judicious about the part machines get wrong.
02 — Your half
The half humans own: the taxonomy
AI can apply a vocabulary; it shouldn't invent yours. The taxonomy is a team decision, and the rules are the ones from our organizing guide: three or four categories that mirror how people ask for things (product, campaign, brand, plus one custom), five to fifteen tags per category, asset type as a first-class field rather than a tag.
Two governance details matter more with AI in the loop:
- Gate tag creation. Anyone can apply tags; few people may create them. An AI that can only suggest from the approved catalogue cannot drift it.
- Merge quarterly. Synonyms still creep in through humans. A fifteen-minute merge pass (“Q1 launch” + “Q1-2026” → one tag) keeps every filter meaning what it says.
03 — The machine's half
The half AI owns: the first draft
Vision models are genuinely good at the boring half: looking at an asset and describing it. Run at import, a scan can draft the fields a human would otherwise type:
- Content and style tags — subject, mood, palette, format, drawn from your catalogue
- Asset type — banner vs. deck vs. illustration, so imports can default to “auto-detect” and land pre-sorted
- Client or product — read from logos and wordmarks, canonicalized against names already in the system so variants collapse into one
The posture that keeps this useful: AI output is a draft with a confidence level, not a verdict. A model that says “probably a landing-page hero, dark, product X” and lets a human confirm beats one that silently writes whatever it guessed.
04 — The handshake
The review pass that keeps it honest
The review is where the two halves meet: a queue of scanned assets, each showing its suggested type and tags, and a human spending five seconds per asset — approve as-is, or edit and approve. Fast enough that it happens; deliberate enough that garbage never enters the catalogue.
Done right, the edits compound. Every correction — a tag removed, a type fixed — is a signal the system can feed back into future suggestions, so the queue gets quieter over time. You are not just cleaning output; you are teaching preferences.
And when a scan goes wrong on a specific asset, re-scanning it should be one click, superseding the old suggestions instead of stacking a second opinion on top.
05 — The backlog
Tagging the backlog you already have
The library you're migrating has eight hundred untagged assets, and the honest truth is nobody will ever hand-tag them. Don't plan for it; batch it. Run the scanner across the backlog, let it draft everything, then review in short sessions — newest first, because recent assets get searched most.
Accept imperfection at the tail: a three-year-old banner with slightly loose tags is still infinitely more findable than an untagged one. The goal is retrieval coverage, not archival perfection — the same small-team pragmatism that applies to the rest of the system.
06 — FAQ
Common questions
How many tags does a design asset need?
Three to six — one per category your team filters by (product, campaign, brand) plus one or two descriptors. Under three, the asset is invisible to filters; past six, tagging becomes a chore and the extra tags add noise, not findability.
Should AI tags apply automatically, without review?
Not while the library is young. Auto-applied tags drift the vocabulary — synonyms, near-duplicates, plausible-but-wrong labels — and cleaning up later costs more than reviewing early. A queue where a human approves or edits suggestions keeps the catalogue honest, and the edits teach the system your preferences.
What makes a bad tag?
Anything a filter cannot act on: subjective words ("nice", "modern"), one-off descriptions that will never match another asset, duplicates of the asset type field ("banner" as a tag on a banner), and synonyms of existing tags. Good tags are reusable, concrete, and drawn from a vocabulary the whole team shares.
Can AI detect the client or product in an asset?
Increasingly, yes — vision models read logos and wordmarks well. The useful implementation canonicalizes what it sees against your existing client and product names, so "ACME", "Acme Inc." and "acme" resolve to one entry instead of three, and flags genuinely new names for a human to confirm.
This split is exactly how DesignVault's tagging works when AI is enabled for your organization: imports are scanned on arrival, suggestions come only from your catalogue, client and product names are canonicalized against the ones you already use, and a review queue lets you approve or edit before anything sticks — your edits feed the next scan's hints. Re-scan an asset in one click, or batch re-scan the library from settings. AI is optional and per-organization; without it, the manual flow is the same minus the drafts.