Designing Decision Governance in Regulated AI Systems

Wells Fargo · AI Governance and Decision Systems Lead · 2019 · 6-month engagement

---

Context

Wells Fargo’s AI Enterprise Solutions group was tasked with advancing machine learning across fraud detection, claims processing, and operational systems. Despite significant technical capability, adoption lagged.

The barrier was not technology. It was trust.

Business units operated in a deeply risk-averse environment, where experimentation was often interpreted as exposure. AI initiatives stalled due to regulatory uncertainty, unclear ownership, and lack of alignment between legal, data science, and product teams.

---

The Problem

AI efforts failed not because models were ineffective, but because no system existed to make their outputs accountable.

Teams encountered:

• inability to frame AI initiatives in compliance-safe terms
• misalignment between technical capability and regulatory expectations
• lack of shared language across legal, product, and data teams
• absence of structured pathways from idea → experiment → approval

The organization had AI ambition, but no reliable way to test it safely inside a regulated environment.

---

Intervention

I led design strategy to create decision governance frameworks that made AI experimentation structured, observable, and compliant.

The focus was not on building models, but on designing the governance system around them.

Early conversations revealed that legal and risk stakeholders often viewed experimentation itself as a liability. Initial approaches framed around innovation velocity failed to gain traction. The work shifted toward controlled evaluation instead: defining decision boundaries, escalation paths, and accountability before any model could move into testing.

These frameworks introduced friction intentionally, requiring teams to define risk, ownership, and decision boundaries before experimentation could proceed. This shifted AI work from informal exploration to accountable system design.

Key changes included:

• Decision framing templates
  Standardized how teams defined assumptions, risks, data requirements, escalation ownership, and expected outcomes. Each template required teams to document where human override was mandatory and what conditions would trigger review or rollback.

• Compliance-aligned experiment canvases
  Created structured review paths for machine learning proposals, allowing legal, compliance, and product teams to evaluate experiments using shared governance criteria rather than competing interpretations of risk.

• Cross-functional labs
  Created controlled working sessions where risk officers, data scientists, and product leaders reviewed live scenarios together instead of handing requirements across organizational silos.

• Human-in-the-loop checkpoints
  Embedded validation, escalation, and override checkpoints into early-stage workflows so automated recommendations could be challenged, paused, or rejected before downstream impact occurred.

• Regulatory trust layer
  Introduced interpretability thresholds, audit traceability, and decision lineage requirements so stakeholders could understand how outputs were generated, reviewed, and approved.

---

Constraints

The organization operated under significant regulatory scrutiny, with little tolerance for ambiguity around accountability, auditability, or operational risk.

AI proposals often stalled not because teams lacked technical capability, but because ownership fragmented the moment automated decision-making entered the conversation. Legal, compliance, product, and data teams frequently operated with incompatible definitions of acceptable risk.

Any approach to AI had to be credible to legal and risk stakeholders while remaining actionable for product and engineering teams.

Much of the work involved reframing experimentation itself from uncontrolled exposure into a governed process with visible accountability.

---

Outcome

AI shifted from a theoretical capability to a structured, testable discipline.

Observed outcomes included:

• new governance pathways established for evaluating machine learning proposals inside regulated environments
• reduced friction between legal, product, and data science teams, allowing decisions to move through review cycles more consistently
• cultural shift from blanket risk avoidance to controlled experimentation, particularly in teams that had previously avoided AI proposals altogether
• 4 previously stalled AI initiatives moved into pilot phases within 3 months across fraud and compliance workflows, validating the framework through real operational use

Teams that previously could not advance AI proposals due to regulatory ambiguity were now able to frame, test, and evaluate them within defined governance boundaries.

---

Long-Term Impact

The engagement helped shift AI from an abstract innovation initiative into a governed operational capability.

The most important outcome was not the number of pilots launched. It was the creation of a shared decision framework that allowed legal, product, risk, and data teams to evaluate machine learning proposals using common governance language.

The conversation shifted from:

“Is this allowed?”

To:

“How do we test this safely, and what evidence would justify expansion?”

That shift created a foundation for future AI adoption rooted in accountability, traceability, and institutional trust rather than informal experimentation.

---

What I’d Do Differently

Looking back, I would have pushed earlier for a shared governance vocabulary across legal, product, and data science teams.

The technical barriers were often overstated. The deeper challenge was semantic. Different groups were using the same words to describe fundamentally different forms of risk, accountability, and operational exposure.

Once those definitions became visible, alignment accelerated considerably.

---

Due to the regulated and confidential nature of the engagement, internal governance artifacts, workflows, and review materials are not shown directly. Descriptions have been generalized while preserving the operational structure of the work.