The Knowledge Gap in AI Finance Tools

There are two ways to make an AI useful for financial work: teach it the steps, or teach it the domain. Most teams building AI finance tools are doing one. Almost nobody is doing both.

That gap is about to become the most consequential design decision in the space.

Anthropic recently published financial-services-plugins, their official Claude Code skill plugins for financial services. I’ve been building my own finance skills repo. Both projects aim to make Claude useful for financial work — but they represent fundamentally different theories about what “useful” means.

Workflow vs. Knowledge: Two Theories of Usefulness

The shorthand: Anthropic teaches Claude what steps to follow. I teach Claude how to think about finance.

Anthropic’s skills are workflow templates. A skill like portfolio-rebalance walks Claude through a structured sequence: gather current holdings, compare to targets, propose trades, confirm with the user. It reads like a standard operating procedure — reliable, repeatable, well-designed for the common case.

My skills are knowledge modules. My rebalancing skill explains tax-loss harvesting strategies across account types and the tradeoffs between calendar-based and threshold-based approaches. It includes a workflow section, but the center of gravity is the domain itself.

Neither approach is wrong. They solve different problems. And the difference creates two distinct failure modes that matter enormously in regulated financial services.

Two Failure Modes

Failure mode one: the AI doesn’t know the process. It skips steps, misses compliance checkpoints, forgets to document suitability rationale. Anthropic’s workflow skills address this directly — they give Claude a reliable playbook.

Failure mode two: the AI doesn’t understand the domain. It calculates a Sharpe ratio but doesn’t recognize a skewed return distribution that makes Sharpe misleading. It recommends municipal bonds for tax efficiency but doesn’t flag AMT implications for a high-income client. It processes a corporate action but doesn’t understand the downstream impact on cost basis tracking.

If your AI advisor can optimize a portfolio but can’t flag a suitability violation, you don’t have a finance tool. You have a liability.

Where the Numbers Tell the Story

	My Repo	Anthropic’s Repo
Total skills	84	~41
Wealth management	32 skills	6 skills
Compliance / regulatory	16 skills	0
Operations	27 skills (3 plugins)	0
Python reference implementations	Yes	No
MCP data connectors	No	11 (Daloopa, Morningstar, S&P Capital IQ, FactSet, etc.)
Slash commands / hooks	No	Yes

The comparison reveals where each approach invests its energy.

Anthropic built 11 MCP data connectors that wire Claude directly into institutional data platforms, plus slash commands and hooks that make workflows accessible in practice. These are real advantages — skills are only as useful as the data they operate on, and usability features determine adoption.

I invested in depth, covering from front to back office: 16 compliance skills KYC/AML, Reg BI, GIPS, fiduciary duty, and privacy requirements with specific rule citations. 27 operations skills spanning trade settlement, margin requirements, corporate actions, reconciliation, and straight-through processing. Python reference implementations that give Claude verifiable ground truth for every quantitative formula. And cross-references that link rebalancing to tax-loss harvesting to compliance considerations to client reporting — the way an experienced advisor’s mind connects domains. I’m just one person, so there’s definitely room for improvement - please contribute! Let me know what you think if you find them helpful!

Operations may be unglamorous, but it’s where real money gets lost when AI hallucinates a settlement date. Compliance isn’t optional — it’s the difference between a tool and a liability.

What This Tells Us About the Market

The skill plugin ecosystem for AI in financial services is still remarkably early. Anthropic — with the best AI lab on the planet — shipped a wealth management plugin with 6 skills. Not because of a lack of resources, but because getting financial domain content right is genuinely hard. Every sentence in a compliance skill has to be precise. Every formula has to be correct. Every cross-reference has to be accurate. LLMs can’t reliably generate this content about themselves.

And that’s the structural insight: domain knowledge is a moat, not a feature. The firms that invest in teaching AI the why behind financial work — not just the what — will compound that advantage every quarter. Workflow automation is table stakes. Domain comprehension is the differentiator.

The Business Implication

Consider what happens when a wealth management firm evaluates these tools. Workflow automation solves the efficiency problem — advisors move faster through standard processes. But domain knowledge solves the trust problem — the AI’s recommendations are defensible, its compliance awareness is verifiable, its quantitative work has ground truth.

Efficiency without trust doesn’t survive the first regulatory exam. Trust without efficiency doesn’t survive the quarterly business review. Firms need both layers, and right now, almost no one is building them together.

These approaches will converge. But the teams that figure out how to combine deep domain comprehension with reliable process execution first won’t just have better AI. They’ll have a structural advantage that compounds every quarter.

---

Joel Lewis builds at the intersection of AI and financial services. The finance skills repo is at github.com/joelelewis/finance-skills. Anthropic’s repo is at github.com/anthropics/financial-services-plugins.