Navigate cross-functional conflict to deliver outcomes
Company: Capital One
Role: Data Scientist
Category: Behavioral & Leadership
Difficulty: medium
Interview Round: Technical Screen
Describe a specific project where you had to coordinate among at least three departments (e.g., Product, Legal, Risk) with conflicting goals. What was the concrete business objective, the precise points of disagreement, and what decision-making framework did you use to break deadlocks? Detail the artifacts you produced (e.g., decision log, RACI), how you negotiated scope/time/quality trade-offs, and the measurable result versus baseline. If you had to do it again with one constraint changed (e.g., Legal veto power), what would you do differently and why?
Quick Answer: This question evaluates cross-functional leadership, stakeholder management, decision-making under conflict, regulatory risk awareness, and the ability to document trade-offs and deliver measurable outcomes in a Data Scientist role.
Solution
# Sample high-signal answer (STAR + frameworks)
## Situation
As the lead data scientist, I was asked to launch a machine-learning–driven pre‑selection system for a new credit card acquisition campaign in 8 weeks. The environment was highly regulated, and we had to align Product (growth), Risk/Underwriting (losses), and Legal/Compliance (fair lending and marketing language). Marketing and ML Engineering were also involved.
## Task (Business objective and guardrails)
- Objective: Increase booked accounts by ≥8% versus baseline campaign while keeping first‑year charge‑off rate within ±10 bps of baseline and meeting fair lending standards.
- Key guardrails (co-created in week 1):
- Loss guardrail: ΔCharge‑off ≤ +10 bps vs baseline.
- Fairness guardrail: Adverse Impact Ratio (AIR) ≥ 0.90 on approvals across protected classes.
- Timeline: MVP ready for a seasonal campaign in 8 weeks.
We translated the goal into an expected value formula to ground trade‑offs:
- Expected NPV per prospect = Conversion × LTV − LossRate × LGD − CAC
- Example values for planning: LTV = $380, LGD = 0.85, CAC = $120.
## Stakeholders and precise points of disagreement
- Product: Push for broad targeting to hit the 8% uplift; willing to accept slightly higher losses.
- Risk/Underwriting: Tighten thresholds; require challenger validation; concerned about a spike in delinquencies.
- Legal/Compliance: Restrict feature set (no proxies like ZIP granularity), require bias testing and documentation; objected to “pre‑approved” language (UDAAP risk), preferred “pre‑selected.”
- Marketing: Wanted “pre‑approved” language for higher response; favored direct mail and digital in parallel.
- ML Engineering: Concerned about feature store readiness; flagged risk in deploying both batch (mail) and real‑time (digital) within 8 weeks.
Conflicts crystallized into three decision areas:
1) Targeting breadth vs loss risk.
2) Marketing language and channels ("pre‑approved" vs "pre‑selected", mail + digital vs digital‑first).
3) Feature set and fairness (use of borderline features vs compliance risk and model performance).
## Decision-making framework to break deadlocks
I used a combination of DACI, RACI, decision log with typed decisions, a weighted decision matrix, and guardrails.
1) DACI + RACI
- Driver (DACI): Me (Data Science) for the MVP design and experiment plan.
- Approvers: Head of Risk (loss guardrail) and Compliance Director (language and fair lending). Product GM co‑approver on scope.
- Contributors: Product, Marketing, ML Eng, Model Risk Management (MRM).
- Informed: Operations, Finance.
2) Guardrails and one‑way/two‑way doors
- Declared one‑way doors: mailed language/creative (irreversible this cycle), feature exclusions required by Legal.
- Two‑way doors: threshold tuning, segment inclusion, digital channel pacing.
3) Weighted decision matrix (example)
We compared three options on normalized 0–1 scales with weights: w_NPV = 0.5, w_ComplianceRisk = −0.3 (negative because risk reduces score), w_Fairness = 0.2.
- Option A (Broad, "pre‑approved", mail + digital):
- NPV = 0.95, Compliance risk = 0.8, Fairness (AIR) = 0.82 → Score ≈ 0.5×0.95 − 0.3×0.8 + 0.2×0.82 = 0.475 − 0.24 + 0.164 = 0.399.
- Option B (Moderate, "pre‑selected", digital‑first):
- NPV = 0.80, Compliance risk = 0.3, Fairness (AIR) = 0.93 → Score ≈ 0.5×0.80 − 0.3×0.3 + 0.2×0.93 = 0.4 − 0.09 + 0.186 = 0.496.
- Option C (Status quo): Score lower due to minimal uplift.
We selected Option B. The matrix, guardrails, and one‑way door classification made the trade‑offs explicit and unblocked the Legal/Product standoff.
4) Decision log and ADRs
- Logged each decision in Confluence with owner, date, status, rationale, evidence (e.g., bias tests), and whether reversible.
- Architecture Decision Records (ADRs) documented feature exclusions and monitoring design for MRM review.
## Scope/time/quality trade-offs
- Scope: Reduced initial audience to mid‑risk segments (P50–P80 score band) and deferred high‑risk edges to a phase‑2 test. Swapped “pre‑approved” for “pre‑selected.”
- Time: Shipped digital‑only in MVP to avoid creative, print, and mail QA lead times; reused the governed feature store; limited model to monotonic GBM with calibrated thresholds to speed up validation.
- Quality: Removed ZIP‑granularity features and complex interaction terms flagged by Legal/MRM; recovered ~1.2 AUC points by adding safer transactional recency features and isotonic calibration. Added fairness regularization in thresholding to keep AIR ≥ 0.90.
## Artifacts produced
- PRD and Experiment Plan (success metrics, guardrails, sample sizes, stopping rules).
- RACI and DACI charts shared in kick‑off, updated at each milestone.
- Decision Log (Confluence) with typed decisions and links to evidence.
- Model Card (training data, features, performance by segment, fairness metrics, intended use, limitations).
- Bias/Fair Lending Report (AIR, TPR/FPR parity, sensitivity analyses).
- ADRs for model choices, feature exclusions, and monitoring.
- Monitoring Dashboards (leading indicators: data drift, PSI; lagging: loss, AIR drift).
## Outcome vs baseline (A/B test on 200k prospects)
- Booked accounts: +9.8% uplift vs baseline (p < 0.01).
- First‑year charge‑off: +6 bps vs baseline (within +10 bps guardrail).
- CAC: −11% due to better targeting and digital‑first.
- Fair lending: AIR improved from 0.84 (baseline) to 0.92 (treatment).
- Manual review rate: Reduced from 18% to 7% through clearer thresholds and exception rules.
- Financial impact: Using Expected NPV = Conversion × LTV − LossRate × LGD − CAC, we estimated +$3.2M annualized incremental NPV at campaign scale.
Validation/guardrails
- Pre‑specified stopping rules: pause if ΔCharge‑off > +10 bps or AIR < 0.90 for two consecutive weeks.
- Shadow deployment for 1 week before exposure to traffic; fairness and drift checks green before ramp.
## If Legal had explicit veto power (what I'd change and why)
- Compliance‑by‑design backlog: Codify Legal requirements as acceptance criteria in user stories (e.g., banned feature list, approved language library, fairness thresholds) so issues surface in sprint grooming, not at sign‑off.
- Stage‑gated approvals: Insert formal Legal gates at PRD freeze, model card draft, and creative mock review with SLAs; pre‑wire contentious areas via weekly office hours with Compliance.
- Technical constraints early: Train with only pre‑cleared features; apply fairness‑aware thresholding from the first prototype; include a “compliance risk score” in the decision matrix with higher weight.
- Two‑track plan: Ship "pre‑selected" digital MVP under low‑risk channel, and run a separate workstream for mailers pending Legal creative approval, to protect timeline.
Why: Moves the highest‑risk, one‑way decisions earlier, reduces late surprises, and gives Legal durable control without derailing delivery.
## Pitfalls and tips
- Pitfall: Optimizing solely on AUC. Remedy: Align on business and risk guardrails and use expected value and fairness metrics.
- Pitfall: Late Legal review. Remedy: Compliance‑by‑design and staged gates.
- Pitfall: Overbroad pilot. Remedy: Start with mid‑risk segments and expand behind monitoring.
- Pitfall: Ambiguous ownership. Remedy: Publish RACI/DACI and a decision log with approvers and escalation paths.
