Building an enrichment workflow - from manual to automated
Everything in this academy has pointed at the same dependency: filters, feeds, SEO, AI visibility, markets - all of it consumes product data richer than what catalogs naturally contain. Enrichment is the work of closing that gap: taking products from "exists with basics" to "complete against your data model." This lesson is about running that work as a workflow with maturity stages - because most brands do enrichment as heroic one-off projects, and the projects don't stick (Phase 2 said why: new products arrive incomplete, and decay outruns heroics).
The loop, whatever the tooling
Enrichment at any maturity is the same five-step loop. Detect: which products are incomplete, against which attributes (your audits and QA rules - the gap report). Prioritize: by exposure, not alphabet - revenue products, filter-load-bearing attributes, channel-blocking gaps first (the traffic-weighted triage that's run through this whole academy). Generate: produce the missing values - from supplier documents, product knowledge, images, existing prose (the fact hiding in a description, promoted to a field), research. Review: human judgment on what was generated, weighted by risk - compliance-relevant fields get real review; a season tag doesn't. Publish and verify: into the catalog through your governed write paths (last lesson), with downstream verification (the rendered check, the feed re-crawl). The loop is the invariant; maturity is about what executes each step.
The maturity stages, honestly
Stage one - spreadsheet enrichment: export gaps, humans fill cells, import back. Genuinely fine at hundreds of products; its failure mode is arithmetic - 2,000 products × 8 attributes is 16,000 human decisions, so coverage plateaus exactly where attention runs out (always the long tail, which is precisely the inventory AI search retrieves - Phase 6's irony). Stage two - assisted: generation gets tooling (extraction from existing content, supplier-doc parsing, model-assisted drafting), humans move from writing values to reviewing them - a 5–10× throughput change that shifts the bottleneck to review quality and prioritization. Stage three - platform: detection, prioritization, generation, and review live in one system (this is the enrichment-platform category from Phase 2's tooling lesson) running continuously - new products enter the loop on arrival, rules route review by risk, and enrichment becomes catalog metabolism rather than a project. The stage-three property that matters most isn't speed - it's that completeness becomes a maintained state: the QA gates from lesson two and the enrichment loop merge, and the gap report trends to zero and stays there.
What doesn't change with automation
The judgment layer stays human, and pretending otherwise is how automated enrichment produces confident garbage at scale. Specifically retained: the data model itself (what attributes exist, what values are legal - automation fills the model; it must never quietly extend it), brand voice and claims (generated descriptions reviewed against your TOV and against substantiation - "waterproof" is a claim with consequences, not a plausible token), and the escalation sense - generated values that look fine and contradict a physical fact are the failure mode reviewers exist for, which is why review is sampled by risk rather than rubber-stamped by throughput metrics. The honest framing for the automation decision: stage three doesn't remove the human from enrichment - it spends human attention only where human judgment is the actual value-add. That's also, not coincidentally, the only version of automation that survives contact with a compliance question or a brand audit.