Home· Work· AI & personalization

AI Personalization that moved from search to guidance

Built a recommendation engine that made a content relevant — driving 3× engagement and +11% AUM contribution through transparent, user-controlled personalization.

Experimentationt Guardrails x design Behavioral & Intent Signals
Timeline
14 months
Scope
Web/App + EMail surfaces
Teams
PM · Data Science · Eng · UX
My role
AVP of Digital Product
Increase in content engagement
11%
Lift in AUM contribution
<5% → 18%
Key tool adoption (6 months)
+40%
Average session duration
Executive summary

What we did, why it worked

The platform had excellent content and tools, but discoverability assumed users knew what to search for. We replaced “search-first” with guided discovery—using intent signals to recommend what mattered most, while earning trust through transparency and control.

Context
Platform with a large content/tool library
Problem
High-value tools underused; search didn’t match user intent
Move
Recommendation engine + guardrails + measured rollout
Proof
10% test → 3× engagement, +11% AUM contribution
Personalization succeeds when it earns trust. “Why this” explanations and user control drove adoption as much as the model did.
Context

A library no one could find

The platform had years of white papers, insights, calculators, videos, and planning tools — all high quality, with low discovery. Search returned hundreds of results sorted by recency, and categories were too broad to be helpful.

What the data said
  • Only 12% of customers engaged with educational content in a given month
  • High-value tools had <5% adoption despite prominent placement
  • Surveys asked for guidance, but users didn’t know where to start

In this domain, users often don’t know what to ask. They’re not searching for a regulation term — they’re trying to make a decisions. The system needed to reduce cognitive load and help users take the next best step.

Insight

Flip the model: from search to guidance

The breakthrough was structural, not technical: stop making customers search; start surfacing what’s relevant based on intent signals. That required defining “relevance” in business terms, earning trust, and proving impact before scaling.

Decisions

The decisions that mattered

The differentiator wasn’t “we used AI.” It was the decision system: what we optimized for, what we refused to compromise, and how we proved value.

Optimize for business outcomes, not offline model accuracy Decision rule
OptionsOptimize precision/recall vs. optimize behavior + business lift
TradeoffSlower iteration on model tuning, tighter alignment to business value
ProofHeld success to engagement lift + AUM contribution, not only offline metrics
Guardrails were product requirements Trust
Options“Max personalization” vs. transparent, user-controlled recommendations
TradeoffFewer degrees of personalization; higher adoption and lower risk
ProofDismiss controls + “why this” explanations protected trust and reduced backlash
Start with a controlled rollout Validation
OptionsFull launch vs. 10% experiment with a control group
TradeoffSlower scale; faster learning and credibility with leadership
ProofClear lift signals before expanding exposure
System design

Build a system, not a feature

The system needed to be explainable, governable, and measurable. We designed it as a pipeline with constraints, so recommendations were useful.

Recommendation system overview
Signals
Behavior + account context + user metadata
Model
Predict relevance using historical engagement + guardrails
Controls
Explain “why this,” allow dismissals, respect privacy constraints
Surfaces
Web / In-app modules + email, with measurement included
Guardrail principles we enforced
  • Transparency: users can see why something was recommended
  • Control: dismiss or mark “not relevant” to refine future recs (collaborative filtering)
  • Privacy: avoid surprising inferences; minimize intrusion
Proof

How we proved value before scaling

We ran a controlled experiment with 10% of users: half saw recommendations, half saw the existing experience. We measured leading indicators (clicks, adoption) and lagging indicators (contribution over time).

Measures we held ourselves to
  • Engagement rate on recommended modules (and dismissal rate)
  • Adoption of high-value tools (e.g., retirement planning)
  • Satisfaction signals (helpful vs intrusive)
  • Business impact over time (AUM contribution lift)

The recommendation cohort delivered 3× content engagement and +11% AUM contribution over six months. Only after that proof did we expand exposure.

Results

What changed

The metrics mattered —but the deeper win was behavior change: users found and used tools that improved decision quality and increased contribution.

High-value tools got adopted

A flagship retirement tool moved from <5% adoption to 18% within six months by surfacing it to users approaching retirement—without requiring the right search query.

Engagement became durable

Average session duration increased 40%, and engaged users returned 2.5× more frequently—translating into measurable business lift.

A repeatable blueprint for responsible AI

The playbook became standard: problem-first, guardrails defined early, measurement in the product, and experiments before scale.

What I learned

AI wins when the product does

Model sophistication is not the win. Adoption, trust, and measurement is. This worked because we built a product system that made AI useful, governable, and proved its value.

  • Trust compounds: transparency + control improved adoption more than “smarter” recommendations
  • Outcomes over accuracy: engagement and contribution were the north stars
  • Start small: controlled rollout created learning without reputational risk
Artifacts

Templates I use to ship AI responsibly

These are the kinds of lightweight artifacts that made the work repeatable. (Link to redacted versions if you want to share publicly.)