Claude Code Changed Engineering. Finance Is Next.

It is easy to look at Anthropic’s new financial services agents and treat them like another AI product launch.

That misses the point.

The real story is not that Claude can help build pitchbooks, screen KYC files, review earnings transcripts, reconcile general ledger accounts, or close the books at month-end. Anthropic did announce all of that on May 5, with ten ready-to-run finance agent templates available across Claude Cowork, Claude Code, and Claude Managed Agents. [1]

But the deeper story is this:

Financial work is starting to move from assistant mode to operating system mode.

Not chatbots.

Not better search.

Not a slightly faster way to summarize PDFs.

A new layer of work execution.

The reason this matters is because we have already seen the pattern play out in software engineering.

Claude Code started as a tool for developers. Then it became something more interesting. It could read a codebase, make changes across files, run tests, fix failures, and commit working code. Anthropic says Claude Code was made generally available in May 2025 and had already reached more than $2.5 billion in run-rate revenue by early 2026. Weekly active users doubled after January 1. Business subscriptions quadrupled. [2]

That is not casual adoption.

That is behavior change.

Anthropic also says the majority of its own code is now written by Claude Code, with engineers spending more time on architecture, product thinking, and agent orchestration. Stripe, Ramp, Wiz, and Rakuten are listed as examples of teams using Claude Code for migrations, incident response, large refactors, and faster delivery cycles. [3]

Here is the pattern:

First, AI helps with individual tasks.

Then it starts handling full workflows.

Then the human role shifts from doing every step to directing, reviewing, correcting, and deciding what ships.

That is what happened in coding.

Now Anthropic is making the same bet on finance.

Finance looks sophisticated from the outside.

On the inside, a shocking amount of the work is still document-heavy, spreadsheet-heavy, deadline-heavy, and context-heavy.

Analysts live in Excel.

Bankers live in PowerPoint.

Compliance teams live in files, checklists, approvals, and escalations.

Investment teams live across filings, transcripts, market data, broker research, CRM notes, internal memos, and models that someone inherited from someone else.

That is why these agents matter.

The first wave of AI in finance was mostly about retrieval and summarization. Ask a question. Get an answer. Summarize a document. Pull a few citations.

Useful, but limited.

The second wave is different. It is about agents that can carry work across systems. Anthropic’s finance agents are built around skills, connectors, and subagents. In plain English, that means the agent has task-specific instructions, access to governed data sources, and the ability to call smaller agents for focused work like selecting comparables or checking methodology. [1]

That is a different category.

A pitch builder does not just summarize a company.

It creates target lists, runs comparables, drafts a pitchbook, and can produce outputs across Excel, PowerPoint, and Outlook. [1]

A KYC screener does not just read documents.

It assembles entity files, reviews source materials, and packages escalations for compliance review. [1]

A month-end closer does not just answer accounting questions.

It runs the close checklist, prepares journal entries, and produces close reports. [1]

That is the difference between “ask Claude” and “put Claude on the work.”

I think adoption in financial services will happen in three layers.

Call it the Finance Agent Stack.

This is where most firms will start.

Meeting prep. Earnings reviews. Market research. Model updates. Pitchbook drafts. Statement review. Valuation checks.

The analyst still owns the judgment.

The advisor still owns the client relationship.

The compliance officer still owns the decision.

But the agent absorbs the first-pass work that usually eats the day.

This is where Claude can become dangerous in the best way. Not because it replaces a smart analyst. But because it reduces the penalty for being thorough.

More filings reviewed.

More transcripts checked.

More portfolio notes prepared.

More client-specific context pulled before a meeting.

Better preparation stops being reserved for the largest accounts or the most important deal.

That matters.

Finance work breaks when context breaks.

The model is in Excel. The narrative is in PowerPoint. The client history is in the CRM. The research is in FactSet or Morningstar. The email thread is in Outlook. The policy is in an internal folder no one remembers exists.

Anthropic is clearly trying to attack that problem.

Claude now works across Excel, PowerPoint, and Word through Microsoft 365 add-ins, with Outlook support coming soon. Anthropic also says Claude connects to financial data and research providers including FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, Chronograph, LSEG, and Daloopa, along with internal systems like data warehouses, research repositories, and CRMs. [1]

This is where the product gets more serious.

Because the value is not just intelligence.

The value is continuity.

A finance agent that cannot see the relevant data is a toy. A finance agent that can access approved systems, respect permissions, generate work in the tools people already use, and leave behind an audit trail starts to become infrastructure.

That is the line.

This is the part that will make people uncomfortable.

Anthropic says the same finance templates can run as Managed Agents on the Claude Platform for work that spans a full book of deals or runs on a nightly schedule. The cookbooks include long-running sessions, tool permissions, credential vaults, and audit logs where compliance and engineering teams can inspect tool calls and decisions. [1]

That is not “help me write a memo.”

That is “run this process every night and show me what changed.”

Think about what that could mean.

A market researcher watches issuers, sectors, filings, news, and broker research, then flags changes for credit and risk review.

A GL reconciler runs NAV calculations against books of record.

A statement auditor checks financial statements for consistency and audit-readiness.

A KYC agent assembles files and escalates the cases that need human attention.

This is where the work changes shape.

The employee becomes less of a processor and more of a reviewer, exception handler, and judgment layer.

That sounds clean in theory.

It will be messy in practice.

Financial services is not software engineering.

The tolerance for error is lower. The data is more sensitive. The supervision burden is real. The client impact is direct.

FINRA’s 2026 regulatory report is clear that existing rules still apply when firms use GenAI, including supervision, communications, recordkeeping, and fair dealing. FINRA also calls out risks around hallucinations, bias, cybersecurity, governance, testing, monitoring, prompt logs, model versions, and human review. [4]

The agent-specific risks are even sharper.

Autonomy. Scope. Authority. Auditability. Data sensitivity. Domain knowledge. Misaligned incentives. [4]

That list should slow people down.

Not stop them.

There is a difference.

The firms that win will not be the ones that blindly hand everything to agents. They will be the ones that map the workflow, define the permissions, constrain the agent, monitor the output, and keep humans in the right places.

Not everywhere.

Not nowhere.

The right places.

The Federal Reserve has already acknowledged that AI in banking requires a balanced approach, supporting safe adoption while recognizing that old supervisory frameworks may not fit generative and agentic AI cleanly. [5]

That is the tension.

Finance firms cannot ignore this. But they also cannot treat it like a toy.

Claude Code taught engineers a simple lesson:

If you can describe the system clearly enough, you can hand off more of the work than you thought.

Financial services may be next.

The first-order use cases are obvious. Better pitchbooks. Faster meeting prep. Cleaner close processes. More complete KYC files. Stronger research coverage. Faster model updates.

The second-order effect is bigger.

Smaller teams start operating with the coverage of larger teams.

Advisors prepare like institutional desks.

Compliance teams review more without drowning.

Analysts spend less time assembling information and more time asking whether the information actually changes the decision.

That is the real promise.

Not replacing finance professionals.

Not removing judgment.

Not pretending models should make every decision.

But compressing the work between decisions.

Finance has always rewarded people who can process more information, faster, with better judgment.

AI agents attack the first two.

The third still belongs to the human.

For now, that is the job.

[1] Anthropic, “Agents for financial services,” May 5, 2026. Used for details on the ten finance agent templates, Claude Cowork, Claude Code, Claude Managed Agents, Microsoft 365 add-ins, connectors, agent architecture, use cases, data providers, audit logs, and availability.

[2] Anthropic, “Anthropic raises $30 billion in Series G funding at $380 billion post-money valuation,” February 12, 2026. Used for Claude Code adoption figures, including more than $2.5 billion in run-rate revenue, doubled weekly active users, quadrupled business subscriptions, and enterprise revenue mix.

[3] Anthropic, “Claude Code by Anthropic.” Used for Claude Code’s product description, Anthropic’s internal coding claim, and examples from Stripe, Ramp, Wiz, and Rakuten.

[4] FINRA, “GenAI: Continuing and Emerging Trends,” 2026 Annual Regulatory Oversight Report. Used for regulatory obligations, common GenAI use cases, supervision, governance, testing, monitoring, and agent-specific risks.

[5] Federal Reserve Board, Michelle W. Bowman, “Artificial Intelligence in the Financial System,” published May 1, 2026. Used for the Fed’s discussion of AI adoption, supervisory balance, third-party risk, and generative or agentic AI considerations.