How DoorDash built a world-class centralized data org

Executive overview

Most analytics teams are reactive service functions that answer questions on demand. Jessica Lachs built DoorDash's data org as a proactive business partner with a seat at the table alongside product, engineering, and ops.

The core structural choice: a centralized model where analysts report through one org but are embedded in cross-functional pods aligned to partner teams. This preserves collaboration and ownership while maintaining consistent talent bars, shared methodologies, and growth paths.

Analytics should find opportunities and shape decisions — not just answer questions.

The case for a central analytics org

• Central reporting preserves a consistent talent bar; embedded silos produce uneven quality
• Analysts can move between pods (e.g., marketing to merchant) — prevents stagnation and aids retention
• Shared metrics and methodologies across the org; no competing definitions of the same KPI
• One churn prediction model built collaboratively beats six siloed versions
• Shared team culture and identity matters for recruiting and belonging — DoorDash calls it the "A team"
• Central visibility lets leadership spot cross-team patterns and get ahead of scaling problems

How pods work in practice

• Pods map directly to how product, engineering, ops, and marketing are structured
• Analysts share the same goals as their partner teams — incentives are fully aligned
• Partner teams feel the analyst is "theirs"; analysts feel part of both their pod and the broader data org
• Business leaders retain influence over priorities through shared goals, not through org control

Balancing reactive requests and proactive deep dives

• Exploratory work is the first thing cut under inbound pressure — protect it deliberately
• Set explicit team goals around self-directed insight work; make it accountable, not optional
• Hackathons carve out dedicated days for deep dives; many roadmap-shaping insights come from these
• When a new request arrives, surface the trade-off explicitly: "If I do this, which of these three things drops?"
• Shared goals make prioritization conversations natural — both sides want the analyst on the highest-impact work
• Quick wins still matter: occasionally just knock out a five-minute ask to build goodwill

Example: referral channel fraud discovery

• Referral channel showed below-average consumer quality and payback — standard read would have cut spend
• Hackathon deep dive revealed a bimodal distribution: great referrers driving great consumers, and a second group gaming the system with online promo codes
• Removing the second group would have revealed referral as a top-performing channel
• Outcome: fraud caps introduced, referral rules tightened, channel spend re-evaluated
• Lesson: averages mislead; always look at distributions and break down what you're seeing

Hiring for curiosity and ownership

• Technical bar is table stakes — screen for it early, but it's not what separates great from good
• Curiosity is the most valuable non-teachable trait: the instinct to pull a thread even when you could call it done
• Test for it in case interviews by embedding something slightly off — see if the candidate notices unprompted
• Look for candidates who can make a call with incomplete information; life rarely gives perfect data
• How someone reacts to being told they're wrong is a strong signal — assess this during cases
• Extreme ownership over outcomes matters more than staying in job-description bounds; data scientists should be willing to call customers, write product specs, or do ops work if that's what unblocks the team

Picking and designing good metrics

• Find proxy metrics: short-term measurables that reliably predict the long-term outcome you actually want
• Retention is a poor goal metric — too slow to move; find its leading inputs instead
• Keep metrics simple: composite scores with coefficients that no one understands drive no real behavior
• A simpler metric that everyone intuits and can act on beats a theoretically perfect one no one uses
• Three clear metrics beat one opaque composite — prioritize by impact and work through them sequentially
• Translate everything into common currency (at DoorDash: gross order value and order volume) so trade-offs across teams become comparable; e.g., "what does lowering delivery time by one minute buy us vs. a $1 price cut?"
• Switching a team's metric every quarter is inefficient — let them develop deep expertise on one metric before rotating

Tracking fail states, not just averages

• Rare but severe failures (DoorDash calls them "never delivered") don't show up in average quality metrics
• These events drive outsized churn; the full impact is invisible because lost customers don't appear in future data
• Some failures (e.g., login errors) mean affected users never enter your dataset at all — the data has a blind spot
• Set explicit goals to eradicate these edge cases; assign a team whose sole job is to drive them toward zero
• Root-cause analysis requires both quantitative and qualitative research — data scientists should make customer calls when the data runs out

Building team culture and retaining talent

• DoorDash is a net importer of talent into analytics — people join from ops, engineering, finance, and marketing
• Diverse backgrounds (startups vs. large companies, technical vs. operational) make the team stronger and more self-teaching
• Analysts who rotate across pods can discover interest in product or ops — some have made that transition; that's healthy
• Extreme ownership is a cultural expectation, not a personality trait — it's set at the top and reinforced in how goals are structured
• Four times a year, all DoorDash employees go dashing or do customer support (WeDash) — builds product empathy and catches bugs

AI and internal tooling

• DoorDash built Ask Data AI — an internal chatbot that helps non-technical employees write and adjust SQL queries without pulling analyst bandwidth
• Named after the existing Slack channel (#ask-data) where teams submitted data questions
• Goal: empower business teams to self-serve on common data tasks; free analysts for higher-leverage work
• Office hours (running for eight years) remain valuable for thought partnership and review — AI handles the mechanical lift

Global data orgs

• Similarities across markets outweigh differences; many problems feel like sitting an exam with the answer key
• New complexity: multi-currency, multi-language, and EU vs. non-EU regulation
• Local variations are worth testing even when the instinct is "we've seen this before" — occasionally the answer differs
• Treat international as an opportunity to discover what works in one market but not another, and vice versa