Attributes missing or inconsistent
Colour in the title on one product, in a tag on the next and missing on the third. Search and recommendations can only be as good as the attributes underneath them.
AI by industry · Retail and eCommerce
AI for UK Retail and eCommerce
Search that understands what shoppers mean, recommendations that respect stock and margin, catalogue data cleaned at scale and service assistants that know where an order is — each measured with experiments on your own traffic, not promised in advance.
- A/B tested against your current set-up
- Merchandising rules stay under your control
- Generated product data checked before it goes live
Why retail AI disappoints
Retail AI underdelivers when the catalogue is a mess.
Most search and recommendation problems start in the product data, and most AI claims are never tested against a control group. Both can be fixed.
Where AI tends to pay back in retail
The work that moves the numbers is usually the unglamorous kind:
- Filling and standardising product attributes across the catalogue
- Search that copes with synonyms, misspellings and descriptive queries
- Recommendations constrained by stock, margin and returns
- Service answers grounded in live order and courier data
- Experiments that show what each change actually did
Queries the shop does not understand
Shoppers type “navy dress for a summer wedding” or a misspelt brand and get nothing, or the wrong thing, because search matches words rather than meaning.
Suggestions that ignore stock and margin
Widgets push out-of-stock lines, deeply discounted items or products that usually come back, because nobody told the model what the business cares about.
Uplift that was never measured
A new AI widget goes live at the start of a sale, sales rise, and nobody can say how much was the widget. Without a holdout group, the effect is a guess.
What we build
Retail AI, built on your product data.
Every capability is evaluated offline on your own queries and orders, then tested live against a control group before it reaches every shopper.
Hybrid product search
Keyword and vector search combined, with query understanding for sizes, colours and occasions, and your boosts, buries and pinned products applied on top.
Discuss this
Recommendation models
Collaborative filtering and embedding models for similar items, frequently bought together and personal picks, with business rules for stock, margin and returns.
Predictive analytics
Catalogue attribute enrichment
Language and vision models that read supplier feeds and product photos to fill missing attributes, standardise values and flag conflicts for your merchandisers.
Generative AI
Visual search and image tagging
Image embeddings for “shop the look” and search by photo, plus automatic tagging of colour, pattern and style on new product shots.
Computer vision
Order-aware service assistants
Assistants on web chat and messaging that answer delivery, returns and stock questions from live order and courier data, and hand refunds and complaints to your team.
Chatbots and assistants
Replenishment and markdown signals
Forecasts by product and store or channel that suggest reorder quantities and markdown candidates for your buyers to accept, change or ignore.
Predictive analytics
How we work
From a query log to a tested rollout.
Our usual seven stages, adapted for retail AI: offline evaluation on your own queries and orders, a live experiment against a control group, and work planned around your trading calendar.
- 01
Discovery
We pull a sample of real search queries, sessions and orders and measure today’s baseline: zero-result searches, search exits, click-through and attribute coverage.
A baseline from your own traffic
- 02
Strategy
We decide what to fix first — often the product data — and agree the metric each change will be judged on and how long the experiment needs to run.
A prioritised plan and test design
- 03
UX & architecture
Where AI shows up for shoppers and staff: results pages, filters, recommendation slots, merchandiser controls and review queues for generated content.
Placements and merchandiser controls
- 04
Development
Indexes, models and prompts built against a judged set of your queries and products, so relevance is scored before any shopper sees a change.
A candidate ready for a live test
- 05
Testing
Offline relevance scoring, load tests at peak volumes and checks for embarrassing results — offensive matches, wrong sizes, out-of-stock lines at the top.
Offline results and a peak-load check
- 06
Launch
An A/B test against your current search or widget on a share of traffic, away from peak weeks, read against the agreed metric before a full rollout.
A measured experiment result
- 07
Optimisation & support
Retuning as ranges, seasons and shopper language change, a regular look at the queries that still fail, and the next experiment in the queue.
An experiment roadmap
Technology
What retail AI runs on.
We use your commerce platform and search provider where they are good enough, and add custom models only where they earn their keep.
Search
Elasticsearch and OpenSearchVector searchHybrid rankingSpelling and synonym handling
Recommendations
Collaborative filteringEmbedding modelsLearning to rankBusiness rule layers
Language and vision models
OpenAI GPT modelsAnthropic ClaudeGoogle GeminiImage embedding models
Commerce platforms
ShopifyWooCommerceMagento (Adobe Commerce)Headless and custom stores
Measurement
A/B testingOffline relevance scoringEvent trackingExperiment dashboards
Engineering
PythonNode.js and TypeScriptNext.jsQueues and batch jobs
We do not quote uplift figures before a test. What a change is worth depends on your range, traffic and current search — the experiment tells you.
Use cases
Retail tasks AI handles well.
Jobs with a measurable before-and-after, from the supplier feed to the returns desk.
Zero-result query recovery
Find the searches that return nothing, map them to the right products or categories with synonyms and spelling correction, and track how many recover.
Size and fit guidance
Use order, return and review data to suggest whether a product runs small or large, shown as guidance on the product page once the team has checked it.
Supplier feed mapping
Map supplier spreadsheets and feeds into your attribute schema, standardise units and values, and route anything the model is unsure about to a person.
Product picks for campaigns
Choose products for newsletters and triggered emails by customer segment and stock position, with the team approving each send.
Order fraud and abuse signals
Score orders for signs of payment fraud, reseller bots or return abuse so the team reviews the risky few rather than every order. The model never cancels an order on its own.
Staff product lookup
Let staff ask about specifications, compatibility and stock in plain English, with answers drawn from your product data and linked to the listing.
Why Techsleight
What working with us is actually like.
No inflated numbers — just how we run projects, and what you can hold us to.
Product engineering, not ticket-taking
We ask what the software is for before we estimate it — and we will tell you when something should not be built, or should be bought instead.
AI where it earns its place
LLM features, retrieval and automation built with evaluation, guardrails and cost controls, and plain software where that is the better answer.
Full-stack under one roof
Design, frontend, backend, mobile, cloud and QA in one team, so nothing falls between suppliers.
UK-focused delivery
UK business hours, estimates in pounds, and a contract with a UK company. Our engineers are based in the UK and India.
Flexible engagement
A fixed-scope project, dedicated developers or a monthly retainer — and you can move between them as the work changes.
You own everything
Code, IP, cloud accounts and documentation are yours from day one. We sign an NDA before discovery if you need one.
Support after launch
We stay on for fixes, upgrades and new features, or hand over cleanly to your in-house team with the documentation to match.
FAQs
Questions we get asked.
Straight answers on scope, cost, timelines and how we work. If yours is not here, ask us directly.
Will AI search beat our platform’s built-in search?
Sometimes, not always. Platform search has improved and may be enough for a small, well-tagged catalogue. We test your real queries against it first — zero-result rate and relevance of the top results — and recommend custom or third-party search only when the gap is large enough to be worth paying for.
How do you measure whether recommendations actually help?
With a controlled experiment: part of your traffic sees the new recommendations, part sees the current ones, and we compare the agreed metric — revenue per visit, add-to-basket or margin — over a fixed period. We do not quote uplift figures in advance, because the result depends on your range, your traffic and what you have today.
How much product and order data do we need?
Less than people expect for search, which mainly needs a clean catalogue and a log of queries. Personalised recommendations need enough order and browsing history to find patterns; for a small or new store, simpler approaches — similar items, bestsellers by category — often perform just as well, and we will say so.
Can AI fill in missing product attributes?
Yes. Language and vision models can read titles, descriptions, supplier data and photos to suggest colour, material, fit and other attributes, and standardise inconsistent values. Suggestions below a confidence threshold go to a person, and anything that could become a product claim — materials, safety, sizing — is checked before it is published.
Does personalisation work with our cookie consent set-up?
It has to. Personalisation that relies on tracking should only use data from shoppers who have agreed to it under PECR and UK GDPR, with non-personalised results for everyone else. We design that fallback so the shop still works well without consent; your team or advisers decide what your consent covers.
Can we keep control of what appears at the top of search?
Yes. The model ranks by relevance, and your merchandisers set boosts, buries, pinned products and campaign rules on top. We also build rules the model cannot override, such as never ranking out-of-stock or age-restricted products where they should not appear.
Will AI features slow our store down?
They should not if they are built carefully. Embeddings and recommendations are computed ahead of time where possible, search responses are cached, and model calls sit off the critical page-load path. We load-test before peak and keep a plain fallback in case a model or provider is slow.
Which platforms can you add AI to?
Shopify, WooCommerce, Magento (Adobe Commerce) and custom or headless stores, through their APIs, apps and data exports. Where a platform limits what can change on the storefront, we explain the options during discovery before anything is built.
What does a retail AI pilot cost?
A discovery sprint to audit your search, catalogue data and analytics starts from £2,000. A pilot — for example, search on one category or a single recommendation slot — is then quoted at a fixed price, including the experiment. Search-service and model API fees are paid to those providers, and we estimate them from your traffic.
Start a project
Want to know what AI could do for your store?
Tell us your platform, roughly how many products and orders you handle, and where shoppers get stuck. We will look at the real data before suggesting anything.
What happens next
- 1A senior engineer reads your brief within one working day, and replies with questions or a first view.
- 2A 30-minute call to understand the goal, constraints and what good looks like — no sales script.
- 3A written proposal with scope, milestones, team and a GBP estimate you can take to your board.
Techsleight Labs is a trading name of Krapton IT Consultancy.
Explore