AI Strategy for Finance & Operations: My Point of View

As a Data & AI Customer Success Engineer at IBM, I've had the opportunity to help financial services organization (across commercial banking, asset & wealth management, insurance) adopt AI. That work has given me a perspective on where AI can create the most meaningful enterprise value.

This post brings some ideas around AI strategy, use-case prioritization, and the data foundations required to support them.

If you're building AI solutions, developing an AI strategy, or thinking about how AI can reshape operations and decision-making, I hope these frameworks are useful.

CSE Point of View

An introduction:

I am a Data and AI Customer Success Engineer at IBM focusing on financial service clients, helping my clients on enterprise data modernization, and adopt AI responsibly. I am also leading our market's agentic center of excellence as a technical lead, driving agentic solutions across IBM's data and AI products and enabling adoptions across 300+ paractioners. I am experienced in adopting emerging AI capabilities into scalable solutions that deliver measurable business impact.

Here's the one message I keep returning to as a CSE:

I bridge business problems, data, and AI. I'm not interested in AI for its own sake. I want to understand how operations teams work, identify where data/process friction limits them, and use AI to improve decision-making and operating leverage in a controlled, measurable way.

This is the difference between building solutions people use and building solutions that sit in a POC.

The flow is:

Business problem -> Required Capabilities -> DIY/Vendors -> Business value

Some reminders for CSEs:

1. Technical credibility gets you in the room. You need to know how AI works, what it can and can't do, and what the tradeoffs are (latency, accuracy, cost).

2. Business judgment makes you valuable. Understanding where AI fits into someone's actual workflow.

3. Trust make AI usable. In finance, an unexplainable recommendation is an unusable recommendation.

What to Listen For

When you're talking to investment or operations leaders, listen for friction.

Ask:

• Where do people spend too much time finding information?
• Where do they manually reconcile or consolidate data?
• Where is there low confidence in the quality or consistency of data?
• Where does institutional knowledge live in people, inboxes, spreadsheets, and meeting notes?
• Which workflows require repeated manual preparation?
• Where are decisions slowed down because information is fragmented?
• Where do people maintain shadow spreadsheets/processes?
• Which processes create operational risk or require too many controls?

The key theme:

Where is the gap between how the organization ideally works and how it actually works because of data, technology, or process constraints?

Follow-up:

If you could eliminate one bottleneck tomorrow, which would have the greatest impact?

Areas where AI can Create Value

Lead with the outcome as opposed to AI architecture. Here are four ways to think about where AI creates value:

Information and Knowledge

Goal: Create a trusted institutional knowledge system.

An investment professional shouldn't have to remember which system contains the answer. They should be able to ask a question, get the relevant information, understand where it came from, and distinguish facts from analysis.

This works when you have fragmented sources (research, manager reports, meeting notes, emails, investment memos, portfolio data, internal systems) and the time cost of manually searching is high.

Workflow Automation

AI can summarize documenents, and how does that fit in someone's workflow?

Example: Companie's quarterly finance report -> extract -> compare historically -> identify changes/anmolies -> assess invesetment implication -> prepare review package

The value is removing bottlenecks in a human workflow, with the future state of automating that workflow itself.

Decision Support

I see AI as augmenting investment professionals rather than replacing investment judgment

Surface what matters: current performance, historical performance, exposure, recent communications, prior investment thesis, relevant research, potential concerns, and sources/evidence.

Institutional Memory as Infrastructure

One of the biggest opportunities is turning institutional memory into an organizational asset rather than having it reside primarily in individual people's inboxes, spreadsheets, and experience.

This is long-term leverage: building organizational capability.

Proof-of-Concepts

If you're stepping into a role building AI for an investment organization, here's how I think about those critical early months:

1. LISTEN

Meet with: investment leadership, operations teams, technology/data teams, risk/compliance, finance teams

Ask:

• "Where do you lose the most time?"
• "Where don't you trust the data?"
• "What information is hardest to get?"
• "What requires the most manual work?"

2. MAP

Decision & Workflow -> People, Systems, Data -> Controls & Bottlenect

3. PRIORITIZE

Identify 5–10 opportunities. Not every opportunity is equally worth pursuing. I'd prioritize use cases with a clear business owner, measurable value, sufficient data, and an appropriate risk profile.

Value

• Time saved: e.g. Automating institutional client reporting, RFP parsing, ESG data aggregation, and portfolio reconciliation.
• Better decision support: e.g. Analyzing thousands of unstructured earnings transcripts, alternative data feeds, and geopolitical news in real-time to uncover hidden alpha.
• Reduced operational risk: e.g. value to things that did not happen, such as data breaches, compliance fines, or system downtime
• Faster responsiveness: e.g. Track metrics like customer service resolution times, lead response times, or product launch cycles.
• Greater scalability

Feasibility

• Is the data available? Can we access it?
• Is the workflow sufficiently standardized?: If every portfolio manager or compliance analyst runs their process completely differently based on "gut feel," you cannot automate or augment that workflow reliably
• Can we integrate with existing systems?
• Can we measure the outcome?: To prove the ROI you calculated earlier, you must have a clean baseline to measure against. If you cannot track the specific time saved or errors reduced, you will not be able to justify the model's ongoing token and hosting costs.

Risk

• What happens if the system is wrong?: This defines your blast radius. You must separate low-risk errors (e.g., a typo in an internal meeting summary) from high-risk errors
• What information can it access?: You must guarantee that the AI cannot accidentally expose Material Non-Public Information (MNPI), proprietary alpha strategies, or highly sensitive client PII (Personally Identifiable Information) to unauthorized users.
• Can we show the source/evidence?Is there an audit trail?
• Is human review required?

4. PROVE

Start with 1–2 high-value, bounded use cases. Demonstrate: Time saved, Better process, Reduced risk, Improved decision support. Then scale.

5. Governance

• Model Integrity: Hallucination mitagation, fairness evaluations, output drift detection, and accuracy/precision measurement.

• Data Reliability: Data quality assessment, feature drift monitoring.

• Risk & Compliance: Enterprise AI frameworks, access controls, and audit trails.

My Stories

Story 1: Data Foundation Matters

Situation: Client had legacy data pipelines and pervasive data-quality challenges that slowed decision-making.

What we did:

• Worked with client architects and engineers to map the end-to-end pipeline
• Implemented LLM-enabled metadata enrichment to improve data discoverability and PII data protection
• Automated data-quality remediation workflows: from reactive to active data quality management

Result:

• $1.2M direct labor savings (80% automation of identification and resolution work)
• $1M data-error reduction (reduced rework on or manual corrections)
• $2.4M data-quality improvements (faster decisions, fewer manual reconciliations)

Business takeaway: This engagement demonstrates how AI-powered data governance transforms financial services operations. By designing a unified data layer with LLM-enriched metadata and automated quality management, we developed a roadmap for the client to move from a reactive, manual-heavy approach to proactive, intelligent data stewardship.

Story 2: AI Adoption is organizational

Situation: Need to scale agentic AI adoption across financial services at a large organization.

What we did:

• Founded and scaled a Financial Services Market Agentic Center of Excellence
• Coordinated 10+ contributors across business unit to build 25+ reusable agentic patterns across 15+ products
• Reached 300+ practitioners (technical sellers, engineers, executives)

Business takeaway: "Scaling AI is an organizational problem as much as a technical one. Reusable patterns, governance, enablement, and adoption matter as much as the underlying model."

Story 3: Institutional Knowledge is an Asset

What we did:

• Designed a context-aware knowledge management framework
• Combined organizational knowledge with persistent agent memory
• Streamlined daily execution for knowledge workers
• Presented it at IBM AI Think Tank

Why it matters: Institutional knowledge is only valuable if people can actually retrieve and use it. In an investment organization, historical decisions, meeting context, research, and manager knowledge can be extremely valuable

Final Note

Keep this in your head:

I'm not trying to push AI into the organization for the sake of it. I'm trying to understand how the organization works, identify where better data and technology can create leverage, and build solutions that improve the way people make decisions and operate.

           INVESTMENT BUSINESS
                  │
        What decisions matter?
                  ↓
           BUSINESS WORKFLOW
                  │
          Where is friction?
                  ↓
                 DATA
                  │
        Can we trust/access it?
                  ↓
                  AI
                  │
        Can we augment/automate?
                  ↓
        CONTROLS + GOVERNANCE
                  │
          Can we trust it?
                  ↓
          BUSINESS VALUE

Appendix

Technical Trade-offs:

Accuracy: Large foundational models deliver deep reasoning and high precision, but achieving enterprise-grade reliability often requires multi-step agentic workflows or reflection loops that multiply processing time.

Latency: User tolerance demands fast responses (ideally under 1–2 seconds for interactive chat). Multi-agent setups can stretch processing from milliseconds to 30 seconds.

Cost: Complex reasoning chains or over-provisioned GPUs drastically raise total infrastructure overhead.

-- Semantic Caching: Store and reuse previous responses for similar queries to bypass redundant model inference.

-- Model Routing: Direct simple queries to fast, smaller models and route complex tasks to larger, premium models.

-- Quantization: Compresses model weights or cache memory into lower-bit formats to reduce memory bandwidth demands, accelerate token generation, and shrink hardware footprints.

-- Batching: Groups multiple concurrent requests into a single GPU processing step to maximize hardware utilization and drastically lower the expensive compute time required per token

-- Context Engineering: Optimizes the context using techniques like prefix caching, context compression, or dynamic retrieval.

-- Streaming Responses: Stream output text incrementally to dramatically lower perceived response delays for user.

Risk Mitigation:

1. PII Protection: Intercept prompts with a local security gateway using Named Entity Recognition (NER) to mask and vault client identifiers before hitting the cloud LLM, swapping them back only at egress for authorized users.
2. MNPI Access Control: Enforce Attribute-Based Access Control (ABAC) filters natively at the vector database layer so non-public deal context is never retrieved into the LLM context window.
3. Alpha Strategy Isolation: Exclude proprietary code and trade signals from fine-tuning datasets, use Zero Data Retention (ZDR) cloud endpoints.
4. Prompt Injection Defense: Separate untrusted user input processing from privileged data execution using a Dual-LLM architecture and restrict agent capabilities to strict, parameterized API schemas.