Anthropic's 10 Finance Agents — Claude Now Drives Excel and PowerPoint Directly

10 agents

Claude works at the desk now. On May 4, 2026, Anthropic shipped ten finance-specialized agents in a single bundle. The verb that matters: operates. The agents log into Microsoft 365 via OAuth, open Excel, type cell formulas, link sheets, and build full DCF and LBO models. They open PowerPoint, drop in slides, build charts, fill infographics. Until last quarter, Claude returned text. Now it touches the desktop. Goldman Sachs, BlackRock, and BNY Mellon are the first three beta customers.

The players — Anthropic, Microsoft Copilot Finance, beta clients

Anthropic in shorthand: $13B revenue in 2025, $2.7B in Q1 2026 (WSJ). Same week as the $1.5B PE JV with Blackstone, Goldman, and Hellman&Friedman. The finance-agents launch is effectively the first deployment proof for that JV — capital arrives in PE portfolios, agents arrive at the same desks.

Microsoft Copilot Finance is the direct competitor. Microsoft launched it in November 2025 — a finance-specialized Copilot package built on GPT-5 + Excel. Its strength: distribution into every Office 365 seat. Its weakness: domain depth still capped at advisory text. Anthropic punched into exactly that gap.

Three beta customers. Goldman Sachs (David Solomon, plus Marquee already integrating Claude since Q3 2025). BlackRock (Larry Fink, plus the Aladdin operating system getting a Claude patch). BNY Mellon (Marc Argent CIO, custody and asset-servicing workflows). All three were locked in around the same week as the Anthropic-PE JV.

The same-week timing is the story. PE JV capital arrives → finance agents deploy at the same desks. The "operations integration" model started here.

Anthropic's release groups the ten agents into four buckets: Excel modeling (4), PowerPoint automation (2), research and disclosure analysis (2), regulatory reporting (2).

Source: spoonai chart · Anthropic beta customer averages (n=12)

The 10 agents in detail

Category	Agent	Primary task	Auto-rate
Excel modeling	DCF Builder	Discounted-cash-flow modeling	85%
Excel modeling	LBO Modeler	Leveraged-buyout scenarios	80%
Excel modeling	Sensitivity Analyst	Multi-variable sensitivity	78%
Excel modeling	Portfolio Synth	Portfolio performance roll-up	75%
PPT automation	Pitch Deck Builder	M&A pitch decks	70%
PPT automation	IR Deck Synthesizer	Investor relations decks	68%
Research	10-K/10-Q Analyst	SEC filing analysis	92%
Research	News & Sentiment	News crawl + sentiment	88%
Regulatory	SEC Filing Drafter	Filing form drafting	65%
Regulatory	Basel/FRTB Reporter	Capital adequacy reporting	62%

The breakthrough is "drives Excel." Earlier GPT-4 and Claude implementations stopped at "tell me the formula." These agents log into Office 365, write to specific cells, link across sheets, and complete entire models. Roughly 70-85% of what an analyst does between 9am and 6pm collapses into a one-hour batch.

The 10-K/10-Q Analyst hits 92% automation by pulling filings from SEC EDGAR, extracting risk factors, decomposing revenue, mapping debt structure, and rendering summary tables and charts. A one-week task becomes an hourlong run.

What each side gets — Anthropic, beta customers

Anthropic gets two wins simultaneously. One: proof that a frontier lab can ship application-layer products without losing model-layer focus. With Goldman, BlackRock, and BNY all live, the domain-depth question is answered for finance.

Two: justification for the PE JV. The same-week $1.5B JV with Blackstone, Goldman, and Hellman&Friedman now has a clear use case. Capital lands in PE portfolios, Claude agents land at the same desks.

Goldman gets "Marquee 2.0." Marquee — Goldman's institutional client desktop — has integrated Claude since Q3 2025. Adding the ten agents pushes Marquee toward standard-AI-desktop-of-Wall-Street status.

BlackRock gets Aladdin reinforcement. Aladdin manages $1.4T of allocations; Claude agents patched in lift analysis throughput 5-10×. Even small fee compression can be absorbed because of cost-side savings.

BNY Mellon's win is the largest in absolute dollars. With $50T of assets under custody, even 10 basis points of operations-cost savings translates to $500M of incremental operating income — biggest measurable ROI of the three betas.

Source: spoonai chart · company announcements

Pattern matching — what worked, what didn't

Bloomberg Terminal (1981-): the standard finance desktop got there by integrating data, chat, and analysis tools. Claude finance agents redefine that integration as "AI operator on top of Excel."

Aladdin (BlackRock, 2000-): became the asset-management OS via deep operational integration. A Claude agent patched into Aladdin replicates that integration shape on the desktop layer — could match Aladdin's penetration curve in 18-24 months.

IBM Watson Wealth Advisor (2017-2020): launched with Citi and UBS betas, abandoned by 2020 because domain depth never landed. Anthropic's three-beta launch and 90%+ auto-rates on key agents are the explicit anti-pattern.

Symphony Communication (2014-): consortium chat tool from Goldman and 13 other banks. Stuck at chat, never moved to operating the desktop. Lesson: chat alone doesn't define a desktop standard — operating Excel and PowerPoint does.

Three lessons compress: a desktop standard requires data + analysis + operations; multi-customer betas are required for domain proof; auto-rates below 80% don't shift labor allocations enough to anchor revenue.

Counter-plays — Microsoft, OpenAI, Bloomberg

Microsoft Copilot Finance is the direct competitor. Distribution advantage via Office 365 is huge but auto-rates sit at 50-60%. MS will ship Copilot Finance 2.0 in the next 6 months to close the auto-rate gap; meanwhile Anthropic will scale beta from 3 to 50-100 customers — that's the contested period.

OpenAI-PwC partnership (announced May 2026) is five agents on top of PwC's consulting channel. Strength: PwC sells globally. Weakness: GPT-5 is less optimized than Claude for "directly drives Office." PwC will deploy at 70+ global clients in the next 12 months and gather domain data to counter.

BloombergGPT has unmatched data depth but weak tool integration. Strong inside the Bloomberg Terminal silo, weak as a desktop-wide automation play.

So what changes — for builders, founders, investors, end users

Builders should treat "Claude API + Office Add-in OAuth + domain RAG" as the new standard stack. The Excel/PowerPoint operation pattern proven in finance will spawn parallel domain-specific agents (medical, legal, manufacturing) over the next 6-12 months.

Founders face a moving line between "model labs" and "application startups." If Anthropic plays directly at the application layer, the application startups need narrower domains or have to specialize as orchestration on top of Anthropic's stack.

Investors should watch Anthropic's multiple re-rate. Pure model labs trade at 30-40× revenue; pure application companies at 100× ARR. A company doing both has no comp set yet — the next 2-3 quarters of revenue disclosure will define it.

End users see the bigger picture. 70-85% of Wall Street analyst desk-time moves to Claude in the next 18-24 months. The same compression pattern then propagates to accounting, legal, consulting. The framing isn't "the end of analyst jobs" — it's "the redefinition of what analyst jobs do."

Stakes

• Wins: Dario Amodei (Anthropic CEO) — application-layer entry + PE JV justification simultaneously; David Solomon (Goldman CEO) — Marquee credible as Wall Street desktop standard; Marc Argent (BNY Mellon CIO) — biggest absolute-dollar ROI among betas.
• Loses: IBM Watson successors — "AI analyst" category captured; Symphony Communication — desktop standard battle ceded on domain depth; analyst headcount roles — desk work 70-85% automated.
• Watching: Satya Nadella (Microsoft CEO) — Copilot Finance 2.0 auto-rate uplift; Sam Altman (OpenAI CEO) — PwC channel could counter via global consulting; Larry Fink (BlackRock CEO) — Aladdin integration choice between operating-OS standardization paths.

The skeptic's case — "90% in demo, 50% in prod"

Marc Andreessen (a16z) and similar critics argue that demo automation rates collapse 30-40 points in production once data cleansing, exception handling, and error recovery accumulate. Beta-stage 90% may stabilize at 50-60% — material if revenue underwriting assumes the headline numbers.

Gary Marcus (NYU professor emeritus) and similar academics flag LLM hallucinations as catastrophic in finance. A single wrong DCF assumption can swing a valuation 20-30%. Analyst sign-off can never be skipped, which caps the absolute time savings even at high auto-rates.

The skeptic case has two prongs: demo-to-production gap, and hallucination risk in finance modeling. Both check at the three beta customer outcomes over the next 6-12 months.

3-Line Summary

• Anthropic shipped 10 finance Claude agents — Excel and PowerPoint directly operated.
• Goldman, BlackRock, BNY Mellon validating 80%+ auto-rates.
• Direct collision with MS Copilot Finance — 70-85% analyst desk-time automation imminent.

Further reading

• Anthropic announcement — Financial Services Agents
• Anthropic targets financial services with Claude AI agents — PYMNTS
• Anthropic launches financial-services agents that drive Excel — TechCrunch
• Goldman, BlackRock, BNY Mellon test Claude Finance agents — WSJ
• Microsoft Copilot Finance vs Claude — Bloomberg