Customer-relevant product discovery that improves commercial performance while keeping availability, eligibility and brand rules in business control.
Reference design target spanning eligibility, availability, category and merchandising.
Reference design target spanning relevance, diversity and commercial quality.
Reference design target for explaining why an item appeared.
Reference design target: recommendations remain subject to business rules.
Bring recommendation logic, merchandising authority and journey measurement into one operating layer that teams can test, explain and improve.
A shared model of customer context, catalogue eligibility and commercial priorities for each recommendation surface.
Explicit availability, margin, category, eligibility and brand controls that remain visible to business teams.
Balanced measures for relevance, diversity, conversion and downstream commercial quality.
Define the customer, catalogue and merchandising rules that shape an appropriate recommendation.
Connect ranking, constraints and explanation so changes remain visible and manageable.
Run controlled experiments that improve the journey without weakening commercial or customer safeguards.
Personalisation work needs to improve relevance without letting a model override merchandising rules, eligibility constraints or customer trust. A good recommendation is only useful when the business can explain why it appeared and why other options did not.
Personalisation had to improve discovery without allowing a recommendation model to override merchandising rules, eligibility or customer expectations. The useful question was not simply whether a product was likely to be clicked, but whether the recommendation was appropriate for the customer, the catalogue and the commercial context in which it appeared.
Personalisation had to improve discovery without allowing a recommendation model to override merchandising rules, eligibility or customer expectations. The useful question was not simply whether a product was likely to be clicked, but whether the recommendation was appropriate for the customer, the catalogue and the commercial context in which it appeared. The design also needed a clear response when availability, promotion or category constraints made an otherwise relevant item inappropriate.
The experience also needed to protect the customer from recommendations that were technically relevant but operationally wrong. Items can be unavailable, unsuitable for a segment, incompatible with a promotion or over-represented because the model has learned a narrow preference. Those conditions need to be handled as part of the experience, not discovered after launch through a falling conversion rate.
The bottleneck was control. A recommendation engine that cannot respect merchandising and eligibility rules becomes another channel the business cannot confidently operate.
The delivery treats personalisation as a governed journey layer. Models rank relevance, deterministic controls enforce business rules, and experiments measure whether the journey improved rather than only whether a click occurred.
Recommendation quality is treated as a shared outcome between model, catalogue and experience design. This prevents a narrow optimisation target from quietly degrading trust, availability or the commercial intent of the page.
The build pattern combines customer-context signals, content-aware recommendations, business-rule exclusions, experiment design and journey-level measurement so product discovery can improve without losing control.
The delivery combines behavioural and content signals with explicit exclusions, catalogue constraints and measurement views for relevance, diversity and commercial performance. Merchandising teams can inspect and adjust the rules around a recommendation, while the experience records enough context to explain why an item appeared and which constraints shaped the result.
The delivery combines behavioural and content signals with explicit exclusions, catalogue constraints and measurement views for relevance, diversity and commercial performance. Merchandising teams can inspect and adjust the rules around a recommendation, while the experience records enough context to explain why an item appeared and which constraints shaped the result. This creates a controlled handoff between model output and the customer-facing experience, so a change in ranking does not silently change the commercial policy of a page.
The design makes the controls and measures visible to the teams operating the catalogue. A recommendation can be traced to its signals and constraints, while the page-level view shows whether relevance is improving without sacrificing variety or trust. That shared view gives product, merchandising and engineering a common basis for changing the experience.
The result is a personalisation capability that can be operated by more than the model team. Merchandising can change an exclusion, product can adjust the page experience, and engineering can inspect the event and recommendation context when behaviour looks wrong. The shared operating view makes trade-offs explicit and helps the organisation improve discovery without losing the judgement that makes a catalogue commercially useful.
The operating model also creates a better path for experimentation. A change to ranking, catalogue rules or page placement can be evaluated against the same set of customer and commercial measures, with exclusions and availability kept visible. Teams can then learn whether a recommendation improves the experience in context rather than optimising one interaction while weakening the rest of the journey.
This creates a more durable experience than a narrow recommendation experiment. The page can explain and measure the recommendation in the context of catalogue availability, customer intent and merchandising policy, while the operating teams retain the ability to intervene when those conditions change. The result is a controlled system for improving discovery, not a black box that silently reshapes the catalogue. It also gives future experiments a consistent baseline, because teams can compare customer and commercial quality using the same constraints and measures rather than changing the definition of success with each test.
The capability can therefore mature with the organisation. It starts with controlled recommendations and measurable constraints, then gives teams the evidence needed to expand the experience without losing catalogue, customer or commercial control.
The capability is consequently more than a ranking algorithm. It is a controlled experience layer with clear rules, explainable context and measures that account for customer and commercial quality. Teams can improve the recommendation logic, catalogue constraints or page treatment independently while preserving a shared view of what changed and why. That makes future experimentation safer and keeps the experience aligned with the way the business actually operates.
The capability is consequently more than a ranking algorithm. It is a controlled experience layer with clear rules, explainable context and measures that account for customer and commercial quality. Teams can improve the recommendation logic, catalogue constraints or page treatment independently while preserving a shared view of what changed and why. That makes future experimentation safer and keeps the experience aligned with the way the business actually operates. It also gives the catalogue team a clearer basis for explaining trade-offs when relevance, availability and commercial priorities do not point in the same direction.
Scroll diagram horizontally
Behavioural and content signals are used to rank likely relevance without exposing sensitive attributes or unsupported inferences.
Availability, margin, eligibility and brand rules remain explicit constraints outside the model.
Recommendation changes are tested against defined journey measures rather than shipped on intuition.
Search, browse, purchase and rejection signals feed the next iteration without treating every click as success.
Applies eligibility, availability, category and merchandising rules before recommendations reach the customer.
Separates relevance, diversity, conversion and downstream customer signals so optimisation stays balanced.
The hard part is resisting the easy metric. A click can mean relevance, curiosity or confusion. The useful measure is whether the journey helped the customer reach the right product with less effort.
Industry: Retail & E-Commerce
Demand intelligence that helps inventory teams protect availability and working capital while keeping replenishment decisions accountable.
One governed toolset for an enterprise platform’s management surface, designed and operated by Humint Labs to make authority, tenant isolation and auditability executable.
Humint Labs delivery for one accountable policy across voice, web chat and messaging, with every intervention traceable, tested and owned.
Executive AI guide series
Start with Design Thinking and choose the right combination of generative AI, workflows, bounded agents and agentic services. The four-pillar framework connects practical service outcomes, authority, evidence and human oversight.
Read part 1: Exec Guide to Generative AI, LLMs & Agentic AIAn executive guide from Design Thinking to accountable enterprise AI delivery: departmental use cases, investment ranges, service design, decision rights, architecture, pilot selection, rollout, adoption and operating measures.
Read part 2: Executive Guide to Enterprise AI Implementation