Rules-and-Exceptions Personalization Engine
Turn expert judgement into scalable recommendations
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 5
- Confidence
- 97%
Capture the attributes an expert uses to make a recommendation, then apply the same vocabulary to both customers and products. Trinny London's system considered combinations of skin, hair, and eye characteristics, while products were allocated across labels such as cool, neutral, and warm. A customer profile received an overview label, which was matched back to product labels. Rules handled the common cases; exceptions and precedence resolved conflicts and edge cases. Real-user testing informed the underlying classifications. This structure turns a specialist's judgement into a database-driven result while retaining nuance. The customer receives a manageable, relevant set of choices, and the company can refine the system as more combinations and exceptions become visible.
Origin
Extracted from The Foundr Podcast
Core principles
- 01Personalization begins with explicit attributes, not vague intuition
- 02Domain experts and technologists must teach each other
- 03Broad labels narrow the candidate set before exceptions refine it
- 04The experience should feel personal even when many users share recommendations
How to run it
- 1
Expose expert judgement
Ask domain experts to explain the attributes they notice and how those attributes change a recommendation. Convert implicit judgement into explicit variables.
Pro tip Use real customer sessions to reveal criteria that experts apply automatically.
Watch out Do not let technical convenience define attributes that have no domain meaning.
- 2
Create a shared classification
Label customer profiles and products with a compatible vocabulary. Keep the first classification simple enough that both experts and engineers can audit it.
Pro tip Start with broad categories before adding fine distinctions.
Watch out Inconsistent labels make later rules impossible to reason about.
- 3
Match the base case
Assign each customer an overview label and return products whose allocations match it. This creates the default recommendation path.
Pro tip Show a concise recommendation rather than exposing the full choice set.
Watch out A base match should not pretend that every combination is identical.
- 4
Layer rules and exceptions
Add precedence rules for cases where variables conflict, then document exceptions. Make the order of evaluation explicit to the technical team.
Pro tip Record why each exception exists so it can be tested later.
Watch out Unordered exceptions can create contradictory recommendations.
- 5
Validate with users
Test recommendations on real customers and compare their reactions with the expected result. Refine classifications, rules, and exceptions when repeated evidence disagrees.
Pro tip Separate a one-off preference from a recurring classification problem.
Watch out Do not treat the initial expert model as permanently correct.
In the wild
Woodall described testing makeup on about 500 women and defining combinations across skin, hair, and eye characteristics. Products received cool, neutral, or warm allocations. The system assigned an overview label to a customer, matched that label to products, and then layered rules and exceptions to produce a recommendation.
→ Tacit makeup expertise became an online matching experience designed to reduce the customer's choice burden.
Common mistakes
Making personalization feel mechanical
A questionnaire that merely returns a result can feel unemotional and disengaging. The experience must communicate relevance in language the customer understands.
Skipping cross-functional translation
Experts must explain the domain, while engineers must explain rule precedence and implementation constraints. Without both directions, the encoded model will drift from the intended judgement.
Is it for you?
Best for
It is best for consumer products where suitability depends on several observable customer attributes.
Not ideal for
It is not ideal when products are interchangeable or there is too little domain evidence to define meaningful attributes.
From the transcript
“we look at your skin hair and eye and we give you an overview label”
“you layer the rules and exceptions on top of the database until you get to this algorithm”
“there was an education from me to the tech team to understand that side of it”
From the episode
434: Building a Beauty & Community Empire with Trinny Woodall
Trinny Woodall