Generative AI is having a moment in retail, and most of the noise is about chatbots. At Rabbit, the Egyptian quick-commerce company delivering groceries in minutes from neighbourhood dark stores, we take a more boring and more useful view. Generative AI in retail is the use of large language and multimodal models to produce or transform content — catalogue copy, search interpretations, draft support replies, internal analysis — wherever generating text or understanding messy human input is the bottleneck. The interesting question is never “can we use a model here?” but “is generation actually the right tool, or is a classical model cheaper, more accurate, and easier to trust?” This is a pragmatic tour of where the answer is yes.
Key takeaways
- Generative AI earns its place in quick commerce where the problem is language and content — catalogue enrichment, query understanding, support drafting, and internal productivity — not where the problem is prediction or optimisation.
- Forecasting demand, routing riders, and ranking products are still better served by classical machine learning and operations research. Generative models complement those systems; they do not replace them.
- Every generative use case ships with guardrails: human review for anything customer-facing or money-touching, grounding in our own data, and a hard eye on cost and latency.
- The biggest near-term wins are unglamorous — cleaner catalogue data and faster internal workflows — not a flashy assistant.
Catalogue enrichment and clean-up: the quiet workhorse
A grocery catalogue is a swamp of inconsistency. The same product arrives from different suppliers with different names, missing attributes, no description, the wrong category, or a transliterated Arabic spelling that no one will ever search for. In a country where customers shop in Arabic, English, and Franco-Arabic all at once, this is not a cosmetic problem — it directly determines whether a product is findable.
Where generation helps
- Drafting descriptions and attributes: given a product name, brand, and image, a model can propose a clean description, infer attributes like size or pack count, and suggest a category. This turns a blank field into an editable draft.
- Normalisation and deduplication: models are good at recognising that two differently-spelled entries are the same item, and at standardising units and naming so the catalogue reads consistently.
- Bilingual content: generating parallel Arabic and English copy, and bridging Franco-Arabic spellings, so search has more surface area to match against.
The trade-off
Models hallucinate confidently. A wrong ingredient list or a fabricated allergen claim is worse than a blank field — it is a safety and trust issue. So the pattern is generation as a draft, grounded in supplier data, with human review on anything sensitive and automated checks (does the inferred weight match the pack image? is the category in our taxonomy?) before anything goes live. The win is throughput: people review and correct instead of authoring from scratch.
Search and query understanding: meeting customers in their own words
Customers do not search like a database expects. They type “something for a headache,” “stuff for koshary,” a misspelled brand, or a half-Arabic half-English phrase. Classical search handles exact and fuzzy matches well; it struggles with intent. This is where language models add genuine lift, and we go deeper on the full stack in our piece on search and recommendations.
Where generation helps
- Query interpretation: expanding “headache” into relevant analgesic categories, or mapping a recipe name to its ingredients, so the search engine has something concrete to retrieve.
- Multilingual and typo tolerance: understanding code-switched and misspelled queries without maintaining endless hand-written synonym lists.
- Embeddings for semantic matching: using model-generated vector representations so “baby milk” can surface formula even when the words don’t overlap.
The trade-off
You do not want a generative model improvising the result list in real time — it is slow, costly per query, and unpredictable. The durable design uses generation offline or at the edges: to build embeddings, expand queries, and enrich the index, while a fast, deterministic retrieval and ranking layer serves the actual results. Generation shapes understanding; classical systems serve the response.
Customer support: assist, don’t autopilot
Support is the most obvious place to reach for a chatbot and the easiest place to get it wrong. A quick-commerce order is time-sensitive and emotional — a missing item or a late delivery is a real frustration. Getting an answer wrong, or sounding robotic, erodes trust fast.
Where generation helps
- Agent assistance: drafting replies for a human agent to approve, summarising long conversation threads, and surfacing the right policy so the agent answers faster and more consistently.
- Triage and routing: classifying incoming messages by intent and urgency so they reach the right queue.
- Self-service for the simple cases: answering “where is my order” or “what are your hours” directly, grounded in order and policy data.
The trade-off
Anything that touches money — refunds, credits, account changes — stays behind a human or a deterministic rule. We ground answers in our own knowledge base rather than the model’s open-ended memory, and we measure deflection carefully so we are not trading a small cost saving for a large trust cost. The honest framing is assistance: make agents faster and more consistent, automate only the unambiguous cases.
Internal productivity: the underrated win
The use case with the cleanest cost-benefit is often the one customers never see. Generative coding tools, query assistants, and document drafting help engineering, operations, and commercial teams move faster. An analyst can ask a question in plain language and get a first-pass query; an engineer gets boilerplate written for them; an operations lead gets a draft of a supplier brief.
Where generation helps
- Helping teams build: code generation and review assistance that shortens the loop from idea to working feature.
- Helping teams analyse: turning natural-language questions into draft queries and summarising results, lowering the barrier to data.
- Drafting and summarising: internal documents, meeting notes, and reports that would otherwise eat hours.
The trade-off
The risk here is silent error — a plausible-looking query that subtly counts the wrong thing. So generated code is reviewed and tested like any other code, generated analysis is validated against known numbers, and nothing sensitive leaves controlled environments. The benefit is leverage: the same team ships more, provided it keeps its hand on the wheel.
Where generative AI is the wrong tool
Being candid about limits is what makes the rest credible. Several of the hardest problems in quick commerce are not generative problems, and pretending otherwise wastes money and accuracy.
- Demand forecasting is a prediction problem with strong signal in history, weather, and local events. Purpose-built models do this far more accurately and cheaply than any language model — see our deep dive on demand forecasting.
- Routing, slotting, and inventory placement are optimisation problems for operations research and classical machine learning, where constraints and objectives are explicit.
- Ranking and personalisation lean on embeddings and learned ranking models, not on generation, for the live response.
Generative AI sits alongside these systems — enriching the catalogue they index, interpreting the queries they serve, helping the teams that build them. It is one tool in a portfolio, and the discipline is matching the tool to the job. For the wider picture, our overview of how Rabbit uses AI across the stack shows how these pieces fit together.
Frequently asked questions
Does Rabbit use generative AI to set prices or forecast demand?
No. Pricing, demand forecasting, and similar prediction problems are handled by classical machine learning and optimisation models, which are more accurate, cheaper, and more auditable for those tasks. Generative AI is used for language and content problems — catalogue copy, query understanding, support drafting, and internal productivity.
How do you stop generative AI from giving customers wrong information?
Three habits: we ground models in our own verified data rather than their open-ended training, we keep a human in the loop for anything customer-facing or money-touching, and we run automated checks before generated content goes live. Generation produces drafts and interpretations; it does not get the final word unsupervised.
Is the main benefit a customer-facing chatbot?
Not really. The most reliable wins so far are behind the scenes — cleaner, more findable catalogue data and faster internal workflows. A support assistant helps, but as an aid to human agents rather than a full replacement. We favour the unglamorous, high-confidence uses over the flashy ones.
Want to see how these systems come together in a real quick-commerce operation? Discover how Rabbit works.
