This episode of The Experimentation Edge explores how A/B testing, feature flags, and user research transformed Atlassian's talent product after it failed with its first users. Andrew Willingham — 11 years at Amazon, now Head of Legal and People Products at Atlassian — shares how product experimentation works when you can't test at scale, why your customer and your user are not the same person, and how the metrics you choose decide which experiments you can even run.

Summary
Andrew Willingham, Head of Legal and People Products at Atlassian, spent 11 years at Amazon before joining Atlassian a year ago. His path from running A/B tests on millions of Amazon shoppers to building talent management software for a few hundred thousand employees forced a fundamental shift: when you can't run tests at scale, you have to sit with your actual users and watch them fail. He shares how building a talent review product for Amazon's HR specialists completely flopped when handed to HRBPs — and why that failure taught him more than any winning experiment. Now at Atlassian, he's applying that same rigor to reimagining hiring processes with AI, testing everything from recruiter screens to interview sequences that the industry has run the same way for decades.

Timestamps
03:09 From marketing Amazon's mobile app to building HR software for 1.5 million associates 
08:19 Why a talent review product loved by IO psych experts flopped with actual HRBPs 
11:11 How A/B testing helps product managers escape opinion-based politics 
15:25 Testing copy that changes behavior: "We'll generate that status report for you" 
17:20 The two North Star metrics Andrew optimizes: efficiency and quality 
19:05 Khan Academy's metric trap: measuring cognitive engagement, not just completion 
21:10 Why product managers resist experimentation — and what changes when you admit you don't know 

Takeaways
- Your customer and your user may not be the same person — building for HR specialists instead of the HRBPs who actually run talent reviews resulted in a feature nobody could use. 
- When you can't test at scale, desk rides replace A/B tests — sitting with users and watching them struggle reveals failures faster than any dashboard. 
- Experimentation short-circuits political debates by removing opinion from product decisions. 
- Test metrics before you test features — usage time could signal engagement or just mean your product takes too long to do its job. 
- The experiments that fail deliver the most valuable learnings, especially when you expected a slam dunk. 

Connect with the guest
Andrew Willingham on LinkedIn: https://www.linkedin.com/in/andrewwillingham/
Learn more about Atlassian: https://www.atlassian.com/

Sponsor
Growthbook helps you ship features with confidence by bringing experimentation and feature flagging into one open-source platform. No more guessing whether that new checkout flow actually moved the needle, waiting weeks for data team bandwidth, or flying blind on rollouts.

Growthbook gives you a single place to run A/B tests, manage feature flags, and analyze results against your existing data warehouse.

With powerful stats built in, it takes the complexity out of experimentation, helps you catch regressions before they hit every user, and makes it easy to test ideas that keep your product improving and your metrics moving in the right direction.

See a demo at https://www.growthbook.io/

Topics: A/B testing, product experimentation, feature flags, user research, talent management, qualitative research, metric design, experimentation at scale, growth experimentation.

• (03:09) - From marketing Amazon's mobile app to building HR software for 1.5 million associates

This episode of The Experimentation Edge unpacks how DoorDash's experimentation platform runs 12,000+ A/B tests per year across 42 million monthly active users — and now powers merchant-led testing on menu pricing and promotions. Ilya Izrailevsky, Senior Engineering Manager leading the platform, explains how feature flags, marketplace experimentation, and CEO-level experiment reviews built a multi-million-dollar experimentation culture across consumers, dashers, and merchants.

Summary
Most companies struggle to scale experimentation beyond engineering teams. DoorDash runs over 12,000 experiments per year across 42 million monthly active users — and now they're enabling restaurant owners to run their own A/B tests on menu pricing and promotions. Ilya Izrailevsky, Senior Engineering Manager leading DoorDash's experimentation platform, shares how the company built a three-sided marketplace testing program that balances consumers, dashers, and merchants across 40+ countries. From his time scaling search at Amazon (where offline model evaluation narrowed hundreds of candidates down to 10 for live testing) to preventing DashPass churn at DoorDash, Ilya reveals what happens when experimentation scales beyond product teams — and why CEO-level experiment review emails drive cultural change faster than any training program.

One standout learning: expanding delivery radius to 11+ miles increased grocery orders but tanked retail conversions. The lesson wasn't about distance — it was that one metric approach breaks in multi-dimensional marketplaces. DoorDash now segments experimentation by vertical, behavior pattern, and regional market, using AI agents to mine institutional knowledge from past tests and auto-generate experiment summaries that ship company-wide within hours of readout.

Timestamps
00:40 From building Wasabi (Intuit's open-source platform) to running ML at Amazon and Uber 
03:04 Why product velocity without experimentation creates feature bloat, not impact 
05:32 Scaling search at Amazon: billions of products, 10 visible results, 25% win rate 
08:22 Offline evaluation as a filter — golden data sets cut model candidates before live traffic 
10:23 DoorDash's three-sided marketplace: 300 million feature flag evaluations per second 
12:38 CEO Tony Xu reads every experiment email and replies with alternative hypotheses 
13:33 Democratization at scale: enabling merchants to A/B test menu pricing and promotions 
17:05 DashPass churn experiment uncovered value perception gap — became a full product area 
22:03 Why expanding delivery radius killed retail orders but boosted grocery conversions 
24:16 No single North Star metric — balancing consumer quality, dasher earnings, merchant mix 
27:29 Four-dimensional scale: democratization, global expansion, new verticals, AI agents 
31:03 Agentic experimentation: AI mines past tests to generate hypotheses and debug imbalance

Takeaways
- Win rate matters less than learnings per test — DoorDash ships company-wide experiment summaries (win or lose) that the CEO actively reads and responds to, creating cultural accountability around testing rigor.
- Offline evaluation acts as a pre-filter for model velocity — Amazon's search team used golden data sets to cut hundreds of ML candidates down to 10 for live A/B testing, preventing wasted experiment slots.
- One-size metrics break in multi-dimensional marketplaces — DoorDash balances consumer retention, dasher utilization, and merchant inventory mix across verticals because optimizing one side degrades the ecosystem.
- Democratization requires opinionated templates, not open-ended tools — enabling non-technical users to run tests means embedding success metrics and guardrails into pre-built experiment configs.
- AI scales institutional knowledge, not just analysis speed — mining past experiment readouts to auto-generate new hypotheses turns your testing history into a compounding advantage.

Connect with the guest
LinkedIn: https://www.linkedin.com/in/ilyaizrailevsky/
Learn more about DoorDash: https://www.doordash.com/

Sponsor
Growthbook helps you ship features with confidence by bringing experimentation and feature flagging into one open-source platform. No more guessing whether that new checkout flow actually moved the needle, waiting weeks for data team bandwidth, or flying blind on rollouts.

Growthbook gives you a single place to run A/B tests, manage feature flags, and analyze results against your existing data warehouse.

With powerful stats built in, it takes the complexity out of experimentation, helps you catch regressions before they hit every user, and makes it easy to test ideas that keep your product improving and your metrics moving in the right direction.

See a demo at https://www.growthbook.io/

Topics: A/B testing, experimentation platform, feature flags, marketplace experimentation, machine learning, growth experimentation, statistical significance, experimentation culture, agentic AI workflows.

• (00:40) - From building Wasabi (Intuit's open-source platform) to running ML at Amazon and Uber

This episode of The Experimentation Edge shows how UPS's A/B testing program drove $500M+ in incremental revenue across 80+ customer-facing applications. Dave Massey — head of the J.E.D.I. team (Journey Experience and Design Innovation) — walks through the first test that proved UX could move revenue, how he defended counterintuitive results to skeptical execs, and how a small experimentation team can override opinion with data at enterprise scale.

Summary
Dave Massey walked into UPS in 2016 and immediately got pulled into a meeting about AB testing tools. By the end of the day, he owned the platform—and the problem: UPS hadn't run a single meaningful experiment. Three years later, senior leadership gave him a hard number to hit. Prove UX could move revenue, or the pilot dies. His first test—removing navigation from the checkout flow—delivered $35 million in incremental revenue. Senior leaders didn't believe it. They made him defend the results upside down and sideways. When the dust settled, the data held. Today, Massey's team has driven over half a billion dollars in incremental revenue by treating UPS.com like the e-commerce business it actually is.

Massey's approach is simple: test everything, especially what senior leaders think will work. His team, Journey Experience and Design Innovation (nicknamed J.E.D.I.), has built a reputation for saying no with data, not opinion. When a business unit demanded required recipient emails to capture customer data, J.E.D.I. ran the test in 24 hours and killed it. Conversion tanked. Two years later, the international team asked for the same feature—but framed it as a customs solution. That test passed. Same feature, different reason, different outcome. That's the edge Massey's team delivers: rigorous hypothesis design, a UX research team embedded in the experimentation workflow, and zero tolerance for untested ideas.

Timestamps
03:09 Dave's first day at UPS: inheriting an AB testing tool with no program 
05:59 Senior leadership's ultimatum: prove UX ROI or kill the pilot 
08:38 First test result: $35M from removing navigation in checkout 
09:48 Defending the numbers: how Massey's team survived scrutiny 
11:07 Why a data-driven engineering culture made experimentation inevitable 
16:12 Team size: 80 people supporting almost 80 customer-facing applications 
19:08 The 24-hour test: when required email fields killed conversion 
22:28 Why Massey embeds UX research inside the experimentation team 
24:41 AI at UPS: treating it as a tool, not a replacement 

Takeaways
- Massey's first test removed navigation from UPS's shipping checkout flow and delivered $35 million in incremental revenue—proving e-commerce best practices apply even when customers think "this is just a tool, not e-commerce." 
- J.E.D.I.'s win rate stays high because UX research and experimentation teams operate under the same leader, giving the program both behavioral metrics and voice-of-customer insight before tests ever launch. 
- When senior leaders push ideas, Massey's team tests them instead of arguing—then delivers results that either validate the idea or identify three better alternatives the data actually supports. 
- The same feature (required recipient email) failed for customer data capture but passed for international customs—proof that framing and customer benefit matter more than the feature itself. 
- UPS runs everything centrally now, but the real win is that demand for testing has decentralized—business units across the company now come to J.E.D.I. asking to test their ideas.

Connect with the guest
LinkedIn: https://www.linkedin.com/in/masseycreates/
Learn more about UPS: https://www.ups.com

Sponsor
Growthbook helps you ship features with confidence by bringing experimentation and feature flagging into one open-source platform. No more guessing whether that new checkout flow actually moved the needle, waiting weeks for data team bandwidth, or flying blind on rollouts.

Growthbook gives you a single place to run A/B tests, manage feature flags, and analyze results against your existing data warehouse.

With powerful stats built in, it takes the complexity out of experimentation, helps you catch regressions before they hit every user, and makes it easy to test ideas that keep your product improving and your metrics moving in the right direction.

See a demo at https://www.growthbook.io/

Topics: A/B testing, experimentation, conversion rate optimization, feature flags, UX research, e-commerce experimentation, statistical significance, experimentation team building, growth experimentation, sequential testing.

• (03:09) - Dave's first day at UPS: inheriting an AB testing tool with no program

Summary

Most e-commerce companies test a handful of features each month. Fanatics runs nearly 100 experiments monthly and delivers a big portion of the company's total annual growth through experimentation alone. Medha Umarji, VP of Growth and Experimentation at the multi-billion dollar sports merchandising retailer, explains how she built a program that scales from 10 tests per month to 100—and maintains enough rigor to spot false positives before they become costly decisions.

The difference isn't tooling or headcount. It's culture. When your CEO reads Excel spreadsheets for fun and actively wants data to prove him wrong, you stop debating whether to test and start debating how to test smarter. Medha shares the frameworks Fanatics uses to balance speed with rigor: a "do no harm" track for brand plays that won't show up in conversion metrics, a small-sample framework for teams that can't hit statistical significance thresholds, and an experimentation Wiki that feeds a continuous iteration flywheel. One surprising test on ad removal initially showed 95% statistical significance—until they replicated it and found the result was a false positive. The lesson: even at scale, you need to double-click on causality.

Timestamps

03:09 How Fanatics scaled from 10 to 100 experiments per month over 10 years

05:25 Why some leadership teams embrace experimentation and others resist it

07:06 How experimentation consistently delivers a big portion of Fanatics' annual growth

08:20 What happens when your CEO consumes Excel spreadsheets and questions everything

10:35 How top-down humility shapes an entire company's testing culture

12:10 The ad removal test that looked like a 95% win—then failed replication

15:55 How Fanatics built an experimentation Wiki that powers their growth engine

22:45 The "do no harm" framework for features that don't measure cleanly in A/B tests

25:20 Why lowering barriers to adoption matters more than statistical perfection early on

26:27 Your odds of winning at experimentation are worse than roulette

Takeaways

• Replication catches false positives: A 95% confidence level still means 1 in 20 results are noise—if a critical test outcome can't be explained through micro-metrics, run it again before committing resources.
• Top-down buy-in shifts the conversation from "why test?" to "how do we test?": When leadership treats data as the tiebreaker, teams stop defending opinions and start building better experiments.
• Frameworks like "do no harm" and "small sample" expand who can test: Not every initiative needs 30,000 orders to ship value—lower the barrier for teams that can't hit statistical thresholds while protecting core KPIs.
• Documenting experiments in a centralized Wiki creates a growth flywheel: Fanatics' Wiki feeds their roadmap with iterations on already-built features, reducing tech dependency and accelerating velocity.
• Micro-metrics establish causality beyond top-line KPIs: If revenue moves but scroll depth, cart adds, and product views don't follow the same pattern, question the result before declaring a win.

Connect with the guest

LinkedIn: https://www.linkedin.com/in/medhaumarji/

Learn more about Fanatics

https://www.fanatics.com/

Summary

Chess.com ran its first A/B test in 2023. Two years later, the team is on track to run 1,000 experiments in a single year—and they've already shipped 195 in Q1. 

In this episode, Ashley Stirrup sits down with Nafis Shaikh, Director of Product Management at Chess.com, to get inside the experimentation engine powering one of the world's most beloved gaming products. 

Nafis brings experience from Zynga and Prodigy and a refreshingly honest take on what changes when a product built on passion suddenly has to serve a 10-million-DAU user base that spans absolute beginners to rated FIDE players. He and Ashley get into why one-size-fits-all doesn't actually fit anyone, how to measure an AI coach when you can't tell whether users have their volume on, and a game review experiment that completely upended the team's assumptions about how players want to learn. 

Nafis also shares practical advice for product managers trying to introduce experimentation culture to organizations that have never done it before—starting with a simple pre/post test rather than a fancy platform. If you lead product, care about experimentation maturity, or just want to hear how a classic product is scaling its learning loop, this one's worth your time.

Timestamps

• [00:35] – Chess.com's experimentation origin story and the 1,000-test goal
• [05:01] – Designing for a user base that spans beginners to FIDE-rated players
• [07:30] – The four metrics dimensions Nafis uses to evaluate tests
• [12:03] – How do you A/B test an AI coach when you can't tell who's listening?
• [15:49] – Embracing humility and the shift away from "we know what works"
• [20:50] – The game review test that surprised everyone: 80% of users review wins
• [24:06] – Advice for PMs introducing experimentation at a new company
• [29:15] – The onboarding debate and personalization from session zero

Takeaways

• Scale test volume to learning speed, not just shipping speed
• Build hypotheses around user psychology, not just KPI movement
• Accept that being wrong is the point—experimentation only works when leadership embraces humility
• Start simple if you're new to experimentation; a clean pre/post comparison beats a fancy platform you don't use
• Reposition features around how users actually feel, not how you assume they should feel
• Design onboarding around the shortest path to value, not the longest path to personalization

Guest LinkedIn: https://www.linkedin.com/in/nafis-shaikh-20161916/

Company website: https://www.chess.com

Summary

What happens when A/B testing stops being a tool and becomes your operating system? Suresh Teckchandani, VP of Product & Technology at Ancestry (formerly PayPal and eBay), shares how the team scaled experimentation from isolated tests to capability-building that drives roadmap and revenue. He details the “growth metering” and paywall experiments that unlocked a 5.3% lift in key engagement and improved conversions—then became platform features. Suresh explains Ancestry’s centralized experimentation platform with self-serve access for PMs and engineers, why “obvious” UX changes can be the riskiest, and how removing friction actually hurt engagement by 20–25% due to user mental models. He also breaks down a major growth lever: AI-powered storytelling that turns raw records into narratives, delivering 30%+ CTR lift and a 5x increase in story views. You’ll learn how Ancestry balances input vs. output metrics, when not to test, and why the best leaders optimize for decision quality over win counts—with clean baselines, right audiences, adequate sample sizes, and true statistical significance.

Timestamps

[00:45] – Ancestry’s experimentation maturity: metering, paywalls, and the 5.3% lift that unlocked capabilities

[03:48] – From isolated tests to a capability mindset: experimentation as an operating system

[05:34] – Balancing wins with learning: zooming out for subscription engagement and NPS

[08:49] – Operating model: centralized platform, self-serve dashboards, and baseline resets

[11:59] – Counterintuitive UX lesson: removing friction backfired (–20–25% CTR); respect mental models

[15:10] – AI storytelling as a growth lever: record comparisons into narratives, 30%+ CTR and 5x views

[18:27] – Input vs. output metrics: when to roll back and how to link short- and long-term outcomes

[30:00] – Parting advice: test what changes CX, avoid vanity testing, and optimize decision quality

Takeaways

- Build capabilities, not just tests—use experiments to unlock platform features (e.g., metering, paywalls).

- Democratize experimentation with a centralized platform and self-serve tooling; reset baselines regularly.

- Test “obvious” UX changes; preserve helpful friction and align with user mental models.

- Turn data into narratives with AI to deepen engagement and increase discovery.

- Define input and output metrics; ship only what improves core outcomes (retention, sign-ups), and roll back fast if not.

- Optimize for decision quality: right audience, sufficient sample sizes, clean baselines, and true statistical significance.

Sponsor
Growthbook helps you ship features with confidence by bringing experimentation and feature flagging into one open-source platform. No more guessing whether that new checkout flow actually moved the needle, waiting weeks for data team bandwidth, or flying blind on rollouts.

Growthbook gives you a single place to run A/B tests, manage feature flags, and analyze results against your existing data warehouse.

With powerful stats built in, it takes the complexity out of experimentation, helps you catch regressions before they hit every user, and makes it easy to test ideas that keep your product improving and your metrics moving in the right direction.

See a demo at https://www.growthbook.io/

Summary

How do you drive hypergrowth without guessing? Kameron Tanseli, Head of Growth Engineering at Fyxer—an AI assistant for your email—breaks down the experimentation playbook that helped the company scale from $1M to $35M ARR, with sights set on $100–$150M. Kameron explains how startups should think about A/B testing differently: de-risk big bets, not just button colors. He shares a risk-based approach to when to run rigorous tests vs. ship-and-measure, why a 25% win rate is a sign you’re testing ambitiously, and how PLG features should be shipped first, then rapidly iterated to drive usage. You’ll hear how Fyxer uses AI to speed the entire lifecycle—Claude, Cursor desktop cloud agents, GrowthBook, and BigQuery—plus how a Slack-first changelog and an internal “AI data scientist” democratize insights. Kameron also details turning everyday product usage into growth loops, personalizing signup paths, and measuring success by movement in global ARR, not just local metrics. He closes with candid advice for new growth engineers: expect to struggle early, be T-shaped, and adopt your customer’s language.

Timestamps

[00:34] – Startup A/B testing mindset: de-risking big bets with only a 25% win rate

[02:45] – When to A/B test vs. ship: risk appetite, funnel stage, and non-inferiority tests

[04:43] – 360 experiments with 4 people: scaling to 1,000 using AI and Cursor cloud agents

[08:22] – Separating feature impact from momentum: PLG and trial model moves ARR

[10:29] – Ship PLG features, then iterate to drive usage; measuring DAU and revenue impact

[11:40] – Habit loops to growth loops: turning product features into PLG (scheduling case study)

[16:47] – Building an experimentation culture: founder buy-in, Slack changelog, shared data

[26:50] – The modern growth stack: Claude, Cursor, GrowthBook, BigQuery, and DOT in Slack

Takeaways

- Prioritize by risk: run rigorous A/B tests where you have volume; use before/after or non-inferiority for low-risk in-product changes.

- Test big levers—not just UI: pricing models, usage limits, onboarding pathways—and judge success by ARR movement, not micro-metrics.

- Ship first, then optimize: launch PLG features and immediately run experiments to increase adoption; track daily active usage per feature.

- Build growth loops from habits: design shareable artifacts and personalized signup paths; drive users back to your domain to capture value.

- Scale experimentation with AI: use Cursor desktop/cloud agents for parallel builds and visual QA; orchestrate docs/analysis via Claude; automate cleanups and reporting.

- Make experimentation company-wide: centralize data (BigQuery), broadcast wins/losses in Slack via GrowthBook, and auto-correlate metric dips to releases.

Summary

How do you build a culture where nothing ships without evidence—and leaders actually act on the data? Makram Mansour, Head of Marketplace at ID.me and former experimentation leader at LinkedIn and Intuit, shares the systems, mindsets, and guardrails behind “experimenting everywhere.” At LinkedIn, he helped support 10,000+ annual experiments with 2,000 weekly platform users, and he explains the hard-earned lessons (like a 5px UI tweak causing a million-dollar ad loss) that led to a “test before release” mandate. At Intuit, he operationalized “fail forward,” partnering with HR to rewrite OKRs so teams are rewarded for learning, not just launching. Makram breaks down why to shift from MVP to MVT (minimum viable test), how to surface leap-of-faith assumptions with PRFAQs and “unit of one” prototypes, and where AI now unlocks faster, safer front-end testing. He also details critical guardrails—cost visibility for AI infrastructure, ethical and inclusion metrics, and the people-process-technology triad—plus practical ways to remove bottlenecks via a center of excellence. If you’re starting from scratch or scaling your program, you’ll learn how to personalize responsibly at the top of the funnel, define your North Star and signposts, and stack early wins while building influence across the org.

Timestamps

[00:45] – Makram’s path: running experimentation at LinkedIn and Intuit, and why nothing ships without an A/B test

[02:15] – Costly lessons: 5px banner change, algorithm tweaks, and the case for rigorous guardrails

[06:40] – Leadership discipline: killing features (voice meetups, LinkedIn Stories) and changing OKRs to reward learning

[11:05] – People, process, technology: top-down and bottom-up tracks, and embedding “fail forward”

[13:40] – From MVP to MVT: validating leap-of-faith assumptions, PRFAQ, and rapid “unit of one” prototypes

[15:55] – Bottlenecks and unlocks: engineering/data science capacity, centers of excellence, and AI for fast front-end tests

[22:45] – Personalization at the top of funnel: avoid waste, design reviews, and right-size testing before building

[25:45] – Guardrail metrics that matter: AI infra costs, ethics/compliance, and fairness-by-design

[29:45] – ID.me now: zero-to-one builds, vision-to-values, North Star and leading indicators

[33:30] – How to start at a new org: crawl-walk-run, small wins, relationships, and over-communication

Takeaways

- Shift from MVP to MVT: list leap-of-faith assumptions and design minimum viable tests before you build.

- Institutionalize learning: align OKRs with “fail forward,” and be willing to kill low-performing features quickly.

- Build the triad: pair an easy-to-use platform with training, top-down sponsorship, and clear launch processes.

- Add real guardrails: track AI infrastructure costs, ethics/compliance, and inclusion metrics alongside growth KPIs.

- Unblock teams: create a center of excellence for data science and enable rapid variants with AI-powered tooling.

- Start small and visible: rack up quick wins, over-communicate progress, and grow influence through relationships.

How do you boost developer velocity in a highly regulated industry—without sacrificing safety or customer trust? Charles Williams, Senior Vice President and Software Engineering Director at Truist (formed from the BB&T and SunTrust merger), shares how his team elevates developer experience to ship faster and more reliably. Charles breaks down shifting quality “left” with automation, measuring success with both DORA metrics and developer sentiment, and why human-in-the-loop is non-negotiable for AI in finance. He details Truist’s governance model—steering committees, enterprise architecture, and clear guardrails—to avoid tool sprawl while building a purpose-built AI ecosystem: Microsoft Copilot for productivity, GitLab’s AI-enabled DevSecOps platform for engineering, and separate consumer-facing capabilities. Expect practical insights on starting with low-risk, high-yield use cases (unit tests, docs, security triage), tracking AI utilization, and upskilling teams in prompt engineering so developers can “manage” AI agents effectively. Charles also explores the path to personalized experiences balanced with privacy, why branches should be enhanced—not reduced—by AI, and the cultural skills leaders need now: empathy, neurodiversity awareness, and change management. He closes with where AI is driving ROI first—developer onboarding and pipeline productivity—with code quality gains following close behind.

Timestamps

[00:02] – Truist overview and Charles’s mandate: improving developer experience at scale

[00:56] – AI as a strategic priority; shifting quality left with automation to remove bottlenecks

[02:19] – Measuring success: DORA metrics plus sentiment—eliminating toil to drive happiness

[04:33] – Human-in-the-loop AI for high-stakes finance; customer and internal use cases

[07:25] – How Truist evaluates tools: personas, pain points, and starting with tests, docs, security

[08:35] – The stack: Microsoft Copilot, GitLab’s AI gateway approach, and tracking utilization

[10:36] – New skills and culture: prompt engineering, “managing” AI agents, and strong governance

[20:45] – What’s next: personalization vs privacy, fintech agility + bank stability, and where AI pays off now

Takeaways

- Shift quality left with automated checks so developers catch issues early without human gatekeeping.

- Measure DORA metrics and developer sentiment; remove mundane toil to increase speed and satisfaction.

- Keep humans in the loop for AI-assisted coding and customer answers—trust but verify in regulated contexts.

- Build an AI ecosystem with clear purposes (productivity, engineering, consumer) and a steering committee to avoid duplication.

- Start with low-risk, high-yield AI use cases—unit tests, documentation, and security triage—to build confidence and momentum.

- Upskill teams in prompt engineering and AI oversight so developers can effectively direct and review AI “agents.”

AI is moving so fast that what “good” looked like a few months ago is already outdated. 

So how do you measure value, ship safely, and scale what works? Marco Casalaina, VP of Products, Core AI and AI Futurist at Microsoft, joins to unpack how his team builds and evaluates next‑gen AI—at hyperspeed. 

Marco leads the AI Futures team and previously led Azure OpenAI, Azure Cognitive Services, Responsible AI, and AI Studio; before Microsoft, he ran Salesforce Einstein. 

He explains why enterprise value is best measured by go‑lives and real usage, how Microsoft’s Foundry equips developers with agent‑specific evals (tool call accuracy, task adherence), and why old metrics like “accepted completions” don’t fit modern dev loops. 

We dig into model routing now productized across model families, orchestration frameworks and the Copilot SDK, shared memory experiments, and the rise of self‑verifying agents that iterate to defined thresholds. 

Expect concrete examples—from rewriting docs for coding agents to Ralph loops with browser testing—and practical advice for leaders: major in evals, set acceptable error rates by use case, and get hands‑on with the tools daily.

Timestamps

[00:45] – Guest intro and Microsoft’s enterprise AI focus

[02:07] – Measuring value: go‑lives, telemetry thresholds, and token volume

[03:43] – From chatbots to agents: Foundry evals (tool calls, task adherence) and A/B testing in Microsoft 365 Copilot

[06:17] – When “good” changes monthly: model routing productized across model families

[07:23] – Orchestration and Copilot SDK: agents that create their own tools; OpenClaw and shared memory experiments

[11:45] – Engagement redefined: coding agents read your docs; writing for agents vs. humans

[14:55] – New dev loops: why accepted completions died; Ralph loops and self‑verifying builds

[17:06] – Evals in practice and guardrails: thresholds, non‑determinism, and out‑of‑domain tests; how to keep up without burning out

Takeaways

- Measure value by go‑lives and real usage (token volume), not time in portals or playgrounds.

- Evolve evals for agents: track tool call accuracy/success and task completion/adherence; A/B test models and strategies.

- Productize adaptability with model routing to match tasks to the right model family as capabilities shift.

- Build self‑verification into workflows: pair agents with automated testing (e.g., browser runners) and iterate to thresholds, not perfection.

- Write for agents as readers: tighten documentation, ship vetted code samples, and monitor bot traffic patterns.

- Guardrail open‑world agents: add out‑of‑domain evals and explicit capability limits; set acceptable error rates based on the stakes of your use case.