LLMs | RAG | Automation
Retrieval-augmented assistants, catalogue content at scale and workflow automation, grounded in your own data. Certified with IBM and AcceleratorX in November 2026, and deliberately scoped to where it earns its place.
In one paragraph
The problems that are genuinely language problems: generating and translating catalogue content at a scale no copywriter can match, answering customer questions from your own product data rather than from a model's memory, semantic search that understands a description instead of matching keywords, and absorbing the operational steps that are really text moving between systems. The ones not worth doing are the ones where a model is being asked to be a database, a pricing engine or a decision-maker, which is most of what gets proposed.
What this is
Most AI proposals in commerce fail for the same reason: a language model is asked to do something that is not a language task. It gets asked to know stock levels, apply pricing rules or make a decision that has to be auditable, and it produces something confident and wrong.
The work that holds up is narrower and considerably more useful. Ground the model in your own data so it answers from the catalogue rather than from training. Use it where a human would otherwise be rewriting text. Put a review step wherever the output is customer-facing and the cost of being wrong is real.
What this page is not
This is build work. The certification is from November 2026, so the honest statement is that it formalised and extended work already underway rather than started it, and there is no published AI engagement with a measured outcome on this site.
What usually brings this in
Forty thousand products with no usable descriptions.
Search and category pages underperform, and writing them manually is a year of work nobody will fund.
Support answers the same twenty questions all day.
A team is being paid to retrieve information that is already written down somewhere.
On-site search fails unless the customer uses your exact words.
A shopper describing what they want gets nothing, and leaves for a search engine that understands them.
A chatbot was launched and had to be turned off.
It was ungrounded, so it invented answers confidently, and now the organisation distrusts the whole category.
What gets built
- Retrieval-augmented generation over your catalogue
- Chunking, embeddings and retrieval tuned against your product data, so answers come from your catalogue and cite it. This is the difference between a useful assistant and the one that had to be switched off.
- Catalogue content at scale
- Descriptions, attributes and translations generated against your tone and your attribute data, with a review queue rather than direct publication.
- Semantic search and discovery
- Search that handles a description rather than a keyword, measured against the queries currently returning nothing.
- Workflow automation with n8n
- The operational steps that are text moving between systems, automated with the human checkpoint kept wherever a wrong answer costs something.
- Evaluation before rollout
- A held-out set and a measured pass rate before anything reaches a customer. An AI feature with no evaluation is a feature nobody can defend when it is wrong.
Technologies
- Models
- GPT
- Claude
- Gemini
- Open models
- Technique
- RAG
- Embeddings
- Prompt engineering
- Evaluation
- Tool calling
- Automation
- n8n
- Webhooks
- Queues
- Human-in-the-loop review
- Delivery
- Node.js
- Next.js
- Vector stores
- Elasticsearch
How delivery runs
Is this a language problem?
The cheapest step and the one that saves the most. If the task needs to be exact, auditable or arithmetic, a model is the wrong tool and the engagement should say so.
Ground it in your data
Retrieval against your catalogue, your documentation, your policies. An ungrounded model answers from training, which is where invented answers come from.
Build an evaluation set first
Real questions with known good answers, before the feature is built. Without it there is no way to tell improvement from a change of mood.
Ship behind a review step
Customer-facing output goes through a human until the measured pass rate justifies removing them. The checkpoint comes out on evidence, not on schedule.
Measure in production
What it answered, what it refused, what a person corrected. An AI feature nobody is measuring quietly degrades.
Questions
- Will AI write our product descriptions well enough to publish?
- At volume, with good attribute data and a defined tone, yes, with a review step. The quality ceiling is set by your attribute data far more than by the model: rich structured attributes produce good copy, and a product with three fields produces plausible filler.
- How do you stop an assistant inventing answers?
- Ground it in your data with retrieval and make it cite what it used, then evaluate against a held-out set before launch. A model with no retrieval layer answers from training, which is exactly the failure that gets chatbots switched off.
- Is this proven work or is it new?
- The certification is from November 2026 and this site publishes no AI engagement with a measured outcome, so it is stated as applied capability rather than as a decade of practice. The commerce platform experience underneath it is the part with nine years behind it.
- What does an AI feature cost to run?
- Per-token inference, embedding storage, and the engineering to keep the retrieval layer current as the catalogue changes. The third is the one left out of most estimates and it is usually the largest.
Describe what is in front of you and you get a written scope and a figure, usually within a few days of a short call. No charge for either.