Justin Norvell, CFA

Finding where AI
actually moves the needle

Analyzing business workflows in financial services to identify where AI creates the most leverage — specifically the capabilities that solve multiple problems at once.

The Work

Opportunity Identification

Most AI initiatives fail not because the technology is wrong, but because the wrong problem was chosen. The work starts with the workflow — not the model.

⬡

Workflow analysis

Embedded observation of how investment teams actually work — where time is lost, where decisions bottleneck, and where human judgment is applied to tasks that are actually pattern-matching in disguise.

View process
⬡

Leverage scoring

Ranking opportunities by impact, feasibility, and speed-to-value. Not every AI use case is worth building. This step separates high-leverage interventions from expensive distractions.

View process
⬡

Dependency mapping

Identifying which opportunities require data infrastructure, model customization, or process redesign before AI can succeed — so roadmaps reflect reality, not wishful thinking.

View process
Once the opportunity is identified, the next step is determining the moat. Explore Model Strategy ->

Model Strategy

Intelligence Layer Strategy

Decide where commodity models are enough and where proprietary intelligence is worth building. The goal is not maximum complexity. The goal is durable leverage.

◍

Commodity first

Start with off-the-shelf models for common capabilities like extraction, summarization, and drafting. Validate impact before committing to custom builds.

View process
◍

Proprietary where it compounds

Invest in proprietary intelligence only where your firm's unique data, decisions, and workflows create defensible advantage that competitors cannot copy.

View process
◍

Knowledge layer design

Structure internal policies, prior decisions, and domain logic into a reusable intelligence layer that consistently improves outputs across multiple workflows.

View process
◍

Escalation rules

Define when to stay with commodity models, when to tune prompts, and when to move to fine-tuned or hybrid systems as error tolerance and governance demands rise.

View process
With the moat defined, the next step is execution under uncertainty. Explore Delivery Strategy ->

Delivery Strategy

Managing Probabilistic Roadmaps

AI systems are probabilistic, not deterministic. Roadmaps must account for drift, validation cycles, and continuous adaptation instead of assuming stable behavior.

◎

Evaluation-first milestones

Define success with benchmark tasks and quality thresholds before shipping. Each phase advances only when model behavior meets measurable standards.

View process
◎

Validation loops in production

Build human review, exception handling, and feedback capture into live workflows so outputs can be corrected and learned from continuously.

View process
◎

Drift monitoring and alerts

Track model performance over time, detect regression early, and trigger rollback or retraining when behavior diverges from agreed risk tolerances.

View process
◎

Retraining governance

Treat retraining as a product process with version controls, approval gates, and audit trails to preserve trust, compliance, and reproducibility.

View process
Execution only works if trust, controls, and oversight are embedded. Explore Responsible AI ->

Responsible AI

Ethical Governance

Governance is a product discipline, not a legal afterthought. The strongest AI systems in finance are transparent, auditable, and designed for oversight from day one.

◌

Transparency by design

Every output should be explainable: what inputs were used, what assumptions were applied, and how conclusions were generated for internal and external review.

View process
◌

Bias detection and controls

Build structured bias testing into release cycles so decisions are checked for disparate impact across counterparties, sectors, and portfolio exposures.

View process
◌

Regulatory alignment

Align product behavior with current standards, including EU AI Act obligations, model risk management expectations, and record-keeping requirements.

View process
◌

Human accountability loops

Define when humans must approve, override, or escalate AI recommendations, preserving accountable decision ownership for high-consequence outcomes.

View process
Governance defines how AI should behave. Next, see what AI can actually do. Explore AI Capabilities ->

AI Capabilities in Finance

What AI can actually do in investment workflows

Prediction

Forecasting outcomes from historical patterns — credit performance, portfolio risk, client churn, market signals.

Extraction

Pulling structured data from unstructured documents — financials, agreements, research, DDQs, IC memos.

Classification

Categorizing inputs at scale — routing, flagging exceptions, screening for compliance criteria.

Generation

Drafting documents, summaries, and recommendations from structured inputs with traceable rationale.

Orchestration

Coordinating multiple AI capabilities into end-to-end workflows — the highest-leverage capability for complex institutional processes.

About

Justin Norvell, CFA

20+ years building and optimizing workflows for financial services teams at Charles Schwab and TD Ameritrade — gathering requirements from sales teams, translating regulatory mandates into operational workflows, and delivering systems that practitioners actually use.

Now applying that embedded, workflow-first lens to AI opportunity identification for investment banks, asset managers, and institutional teams.