A retailer with 30,000 SKUs sits down to refresh its product descriptions. The choice looks simple at first glance: hire copywriters, plug in an AI tool, or some combination of the two. Six months later the project has stalled, the AI output reads like AI output, the copywriters are buried in attributes they cannot verify, and nobody can agree whether the original descriptions were better than what replaced them. This is the most common pattern in product description automation right now, and it sits at the heart of every catalogue refresh project.
The question is no longer whether AI product descriptions are good enough to use commercially. They are. The question is which of three approaches fits the catalogue, the team, and the buyer.
Three approaches to product description automation
Most teams pick between three options when scaling product description work. Each one has a different cost structure, a different quality ceiling, and a different point at which it stops scaling.
Manual writing
This means human copywriters research each product, draft descriptions, and review them before they go live. This is what most brands did up until 2022 and what many still do for hero SKUs.
AI-only generation
This means feeding source data into a large language model and accepting the output, sometimes with light QA on top. This is what most ecommerce platforms now offer as a built-in feature, and what most low-budget agencies have switched to.
Hybrid
This means AI-generated drafts with structured human review, or human-written templates that AI populates with attribute data, or attribute extraction combined with description generation in one workflow. Most catalogues at scale end up here once the limits of the other two approaches become obvious.
| Factor | Manual | AI-only | Hybrid |
| Cost per description | High | Very low | Moderate |
| Speed for 1,000 SKUs | 6 to 12 weeks | Hours | 2 to 4 weeks |
| Brand voice fidelity | High | Variable | High |
| Attribute accuracy | Depends on the copywriter | Depends on source data | High when source data is structured |
| SEO performance | Inconsistent across writers | Inconsistent unless prompted | Strong when templates and keywords are managed |
| Scales past 50,000 SKUs | No | Yes | Yes |
| Risk of generic output | Low | High | Low to moderate |
Manual product descriptions: where they still work
There’s nothing wrong with paying a person to write product descriptions. For hero SKUs, configurable products, and anything where the description is genuinely the deciding factor in the buying decision, a human copywriter with category expertise still produces the best output. The work is consultative. It involves asking questions about use cases, comparing alternatives, and structuring the description around what the buyer actually needs to know.
The point at which manual writing breaks is roughly 5,000 SKUs, give or take. Up to that size, a small in-house team or a copywriting agency can keep pace with launches, refreshes, and seasonal updates. Past that point, three problems compound:
- The cost per SKU stays flat while volume rises
- Quality drifts across writers and over time
- New products outpace the team’s capacity, so the catalogue ages faster than it can be refreshed.
Manual writing also struggles when product data is the actual bottleneck. If the source attributes are incomplete or inconsistent, the writer spends most of the day chasing facts rather than writing. This is what most teams discover when they audit the time logs from a copywriting project: the writing itself takes thirty per cent of the effort and the data gathering takes the rest.
That is the problem content enrichment was built to solve, and it’s a different problem from copywriting.
AI product descriptions: where they work, where they fail
Off-the-shelf AI tools generate product descriptions at a per-SKU cost that rounds to zero. For long-tail SKUs that nobody is searching for, products that compete primarily on price and availability, and catalogues where description quality is genuinely not the deciding factor, this is the right answer. A 200,000-SKU industrial distributor selling low-value consumables does not need bespoke copy for every part. They need accurate, consistent, scannable descriptions, and AI handles that competently.
The failure mode for AI-only product description automation shows up in three places:
1. Brand voice
Generic AI output reads generic. Two retailers in the same category using the same underlying model end up with descriptions that are difficult to tell apart, which erodes the differentiation a brand has spent years building.
2. Factual accuracy
Large language models do not know whether a product is waterproof, fire-rated, or compatible with a specific socket type unless that fact appears in the source data. When the source data is thin, the model fills the gap with plausible-sounding language that is sometimes wrong. For regulated categories like electrical or construction, this is a serious problem and a real liability.
3. SEO
Pure AI output without structured prompting tends to repeat phrases and patterns across thousands of SKUs, which dilutes ranking signals rather than strengthening them. Product pages compete with each other in the index, and Google notices.
Tools like Descriptionwise address these failure modes by combining generation with structured input, brand voice training, and per-category templates. That moves the work out of pure AI and into hybrid territory.
Hybrid: the approach most catalogues end up needing
Hybrid is not a single method. It is a category. The three patterns that show up most often in client work are worth naming separately:
1. AI-generated drafts with human review
Source data feeds the model, the model produces a description, and a human editor reviews the top tier of SKUs in detail and the long tail at a faster pace. This works when the source data is good and the editorial team has a clear voice guide.
2. Templated generation with AI filling specific slots
Each product category has a description template, with placeholders for product name, key attributes, use cases, and a benefits section, and AI populates the slots using product attribute data. This pattern produces consistent, predictable output and scales to hundreds of thousands of SKUs without sounding mechanical, provided the templates are properly designed and reviewed.
3. Description generation paired with simultaneous attribute extraction from source documents
Datasheets, supplier PDFs, and manufacturer specs get parsed into structured attributes, and the descriptions are generated from those attributes rather than from the raw documents. This is what SKULaunch was built for, particularly in B2B distribution where supplier feeds arrive as 700-column XML files or unstructured PDFs that no copywriter could realistically work through.
The trade-off with hybrid is operational complexity. Someone has to design the templates, train the model on the brand voice, set up the QA process, and maintain the system as the catalogue grows. For organisations without a clear owner of product data services, this is the point at which projects stall. The technology works. The operating model to run it does not yet exist.
Which approach fits your catalogue?
The decision turns on five variables:
- SKU volume
- Attribute completeness
- Brand voice importance
- Internal capacity
- The cost of getting it wrong
Catalogues under 5,000 SKUs with strong category expertise in-house and a tight brand voice usually do best with manual writing for the hero range and AI assistance for variants and long-tail items. The cost stays manageable and the quality stays consistent.
Catalogues between 5,000 and 50,000 SKUs almost always need a hybrid approach. Manual writing cannot keep pace and pure AI generation does not hold quality. The variant within hybrid that works best depends on whether the bottleneck is attribute data or editorial capacity. If attributes are incomplete, the first job is fixing the data, not generating more copy. A separate piece on cleaning versus enrichment covers that decision in detail.
Catalogues above 50,000 SKUs only work with automation at the core. The question is no longer manual versus AI. It is which hybrid pattern, with what data sources, and what governance process around it.
Regulated categories (electrical, automotive, construction, healthcare) add a constraint that pushes everyone toward hybrid regardless of size. Pure AI output cannot be trusted on safety-critical or compliance-related attributes without human review. Manual writing is too slow to keep pace with regulatory changes. The hybrid pattern is the only one that scales while staying accurate.
When each approach wins
A few quick scenarios from client work make the choice more concrete:
1. A 2,500-SKU specialist apparel brand with a distinctive editorial voice is best served by keeping descriptions manual. The catalogue is small enough, the voice is the differentiator, and the brand is paying its writers for category expertise that AI cannot replicate at the current state of the technology.
2. A 200,000-SKU industrial distributor selling fasteners, fittings, and consumables is best served by AI-only generation paired with strong taxonomy and complete attribute data. Nobody buys a stainless-steel M8 bolt because the description was poetic. They buy on availability, price, and accurate specifications.
3. A 40,000-SKU electrical distributor with a mix of branded products and own-label lines almost always needs hybrid. The branded lines pull descriptions from manufacturer data with AI cleanup. The own-label lines need editorial work to differentiate them in the market. Regulated attributes (IP ratings, compliance standards, certifications) need verified extraction rather than generated language.
4. A 70,000-SKU homewares retailer planning to relaunch on a new ecommerce platform needs hybrid weighted toward template-driven generation, because the bottleneck is timeline rather than quality at the top of the range.
Key takeaways
- Manual writing scales to roughly 5,000 SKUs before cost and consistency break down.
- AI-only generation handles volume but fails on brand voice, factual accuracy, and SEO when source data is thin.
- Hybrid is not one method. It is three patterns: AI drafts with human review, templated generation with AI slot-filling, and attribute extraction paired with description generation.
- The choice depends on SKU volume, attribute completeness, brand voice importance, internal capacity, and regulatory exposure.
- Most catalogues above 5,000 SKUs end up at some form of hybrid. The question is which variant and who owns it inside the business.
Working out which pattern fits a specific catalogue is the conversation that needs to happen before any tool is bought or any project starts. If you are weighing manual, AI, or hybrid for your own catalogue, book a content enrichment scoping call with the Start with Data product data services team. The first session is thirty minutes and produces a clear view of which approach makes sense for your SKUs, your data, and your buyer.