Microsoft Learning Agent

Reskilling is the product. AI is the medium.

Millions of people now have powerful AI sitting next to their work. Most have no idea what to ask it, when to trust it, or how to fold it into how they already work. My job is to change that, at Microsoft's scale.

I joined Microsoft in November 2025 as Sr. UX Skilling Lead in Workforce Acceleration, embedded in Global Learning & Skilling. My work spans four connected fronts: running and synthesizing the Learning Agent (AiLA) pilots, building the research and evaluation systems that make those pilots credible, authoring the experience standards that let AI deliver skilling at scale, and co-designing the Inside AI Signature Experience.

The through-line is a single operating principle I apply relentlessly: move AI learning from subjective opinion to evidence-based decisions, and from "training" to measurably redesigned work.

Cohort counts and scorecard percentages are measured. The 68% win rate is an early estimate (target 85%), not a final result.

Capability went up. Confidence did not.

Enterprises rolled out AI faster than people learned to use it. The tools became dramatically more capable, but most people still default to the old way of working because the new way is unfamiliar and a little intimidating.

Traditional learning makes this worse. Pulling someone out of their work into a separate course is the wrong shape for a skill that only makes sense in context. The challenge is designing learning that happens where the work happens, at the scale and standard Microsoft 365 demands.

The hard part is behavior, not features.

Three rules I lead every decision against.

Four cohorts, one decision-ready evidence base.

I designed, ran, and synthesized the UX research program behind the AiLA pilot, then converted a large, messy evidence base into a single prioritized set of engineering and product recommendations, the artifact the product team works from.

I pooled data from four cohorts, Azure Boot Camp, WFA UAT, CO+I, and Finance, into one comparable dataset. I measured nine survey dimensions on a 1-5 scale and coded every transcript across six qualitative themes. I validated survey claims against in-product behavior using 51 thumbs ratings (75% positive across 18 unique users) tracked separately from the surveys, applying a minimum three-signal rule: behavioral plus survey plus qualitative before any recommendation became actionable.

The output was seven prioritized recommendations, three P0 "must-fix," two P1 "high-impact," two P2 "refinement," each structured as observation to problem to impact to solution, with a NOW / NEXT / LATER sequencing plan for engineering. Specific, reproducible issues, like a manager-to-new-joiner link-sharing bug surfaced in the June 8 Finance focus group, came with named user evidence and were handed off directly to the product team.

The numbers, honestly labeled.

The table below is the actual AiLA program scorecard I produced, pooled top-2-box from pilot data across four cohorts. I label evidence types rather than presenting all figures as uniform results, because that is what makes the data defensible to leadership.

All figures are measured, pooled top-2-box from pilot data. Green threshold: ≥70%. Amber: below threshold, actively addressed.

Turning "why not just use Copilot?" into a number.

The hardest question the product faced was whether the Learning Agent was actually better than the general AI tool people already had. To answer it with evidence rather than opinion, I built a blind, side-by-side testing framework that converts AI response quality from subjective feedback into quantified, defensible signal.

• Identical learning scenarios sent to both AiLA and Copilot, displayed without system labels to eliminate brand bias.
• Four scoring dimensions: clarity, relevance, actionability, and trust, capturing why one response wins, not just which one, before a comparative preference decision.
• Failure-pattern tolerance set at under 10% unacceptable responses and a standalone sentiment target of at least 75% positive.

I designed the framework to be repeatable and scalable across future AI experiences beyond AiLA, so its value compounds as new products enter evaluation.

The defining decision: destination or flow.

Early in the project the team faced a fork. Build a rich learning destination people visit, or weave learning into the surfaces people already use. I prototyped both and pushed hard for the second.

The in-flow direction shaped every pattern that followed. I defined the core interaction models: contextual entry prompts that surface the right AI move at the moment it is useful, citation chips with confidence signals that make answers checkable, and a human-and-AI task split that makes the redesigned workflow legible before and after.

Decision logic for agentic skilling at scale.

I authored the standards that define the minimum experience requirements making AI-enabled and agentic skilling buildable, discoverable, and measurable. These are not guidelines. They are written as decision logic that both human designers and AI systems can follow.

The standards distinguish how they apply across upskilling, reskilling, and redeployment, making them practically usable, not just theoretically correct.

Putting Work Redesign at the center.

Within the AI Signature Experience squad, one of six squads powering Inside AI in Workforce Acceleration, I own two of the six v1 experience components and drive the blueprint that makes Work Redesign the keystone of the whole program.

Turning practice into an organizational asset.

I codified the entire pilot process into a stakeholder-ready, reusable operating model so future waves do not start from scratch. Eighteen sections, eight phases.

The core principle I institutionalized: treat the MVP as a mixed-methods learning experiment, not a feature demo, and treat content readiness as a go/no-go gate, not a soft prerequisite.

The instrument that makes design credible.

Design at this scale is only as credible as the evidence behind it. I built and maintain the pilot dashboard that brings every signal into one place: survey scores, in-product thumbs feedback, focus-group findings, usability results, and enrollment across pilot cohorts. It uses top-two-box scoring and clear RAG status thresholds so leadership can see, at a glance, what is healthy and what needs attention.

The dashboard is how design recommendations get made and defended. It turned scattered feedback from four disconnected cohorts into a single, honest read on the program, and it is what moved the work from experimentation into scaling planning.

The pilot dashboard I built and maintain. Cohort labels are anonymized and dashboard figures shown here are illustrative, the live program data is internal to Microsoft.

Evidence the program can act on.

The clearest results are the ones that moved the organization forward. Six pilots complete. The work has graduated from experimentation into scaling, with CELA and HR onboarding as new cohorts. Seven prioritized product recommendations are active in the engineering backlog, sequenced by signal strength and tractability. The evaluation framework has already shaped product direction beyond the pilot, identifying Viva Learning content gaps that led to a new partnership model.

The program scorecard shows 82% resource click-through and 71% usability, both on track. The weaker signals, 44% reuse intent and 47% value vs. Copilot, are sequenced for active attention, not buried.

82% and pilot counts are measured. 68% win rate is an early estimate, target 85%.

What I am taking from it.