Is an AI agent for financial services worth building in 2026?
Rogo's valuation nearly tripled to $2B in sixteen weeks selling AI agents to investment banks. Anthropic ships eleven of its own for free, wired into FactSet and Moody's, at 37,300 stars. The number that decides this isn't either raise. It's the disclaimer sitting at the bottom of Anthropic's own repo.
Not the horizontal co-pilot. Anthropic's own financial-services reference repository gives away 11 named agents and 30-plus skills, comps, DCFs, LBO models, CIMs, KYC screening, under Apache 2.0, wired directly into data providers like FactSet, Moody's, and LSEG, and it sits at 37,300 stars. In the same window Rogo raised $160 million at a $2 billion valuation and Hebbia has raised $161 million at $700 million, both selling a version of the same idea to the same buyers. The open wedge is a firm's own proprietary deal data and the verification layer that catches AI errors before they compound, neither of which a reference architecture or a generic co-pilot can supply.
What Anthropic actually gave away
maybe worth building read the repository directly, not the announcement. It ships 11 named agents, a Pitch Agent that builds comps, precedents, and an LBO into a branded deck, a KYC Screener that parses onboarding documents against a rules engine, a GL Reconciler, a Month-End Closer, a Model Builder that writes DCF and three-statement models straight into Excel, and more, plus over 30 slash-command skills covering investment banking (/cim, /teaser, /buyer-list, /merger-model), equity research (/earnings, /initiate, /sector), and private equity (/source, /screen-deal, /ic-memo). It connects to 11 external data providers out of the box, Daloopa, Morningstar, S&P Global, FactSet, Moody's, LSEG, PitchBook, and more, deployable as a Cowork plugin, a headless Managed Agents API, or a Microsoft 365 add-in. It's Apache 2.0, at 37,300 stars and 5,400 forks.
The two companies racing to sell the same thing
Rogo raised a $160 million Series D led by Kleiner Perkins in 2026 at a $2 billion valuation, up from $750 million just 16 weeks earlier at its Series C, a roughly 2.7x jump in under four months. It's an AI co-pilot inside Excel, PowerPoint, and Word for investment bankers, serving over 250 institutions including Truist Securities, Nomura, and Baird, with more than 100 analysts running over 10,000 workflows a week at 95% engagement. Hebbia has raised $161 million total, a $130 million Series B led by Andreessen Horowitz among them, at a $700 million valuation, building AI research tools for finance and legal work. Both are strong, real businesses with paying enterprise customers. Both are also now competing with a free reference architecture from the model vendor they're built on top of.
Why it isn't already solved
Here's the mechanism that explains why none of this, Anthropic's repo included, replaces an analyst outright. Anyscale co-founder Robert Nishihara named it directly: "Large language models do well when you collect a lot of data, and we don't have nearly as much data for real-world tasks." Poetry and code were scraped from a public internet built over decades. Investment banking workflows, specific data entry conventions, firm-specific underwriting judgment, the exact way a CIM gets built at one shop versus another, were never public and never scraped. Small errors in that gap compound across a multi-step model build until the whole thing goes off the rails, which is the technical reason a 95%-engagement co-pilot still isn't an autonomous one.
The tell in Anthropic's own repo
The clearest evidence of that gap is sitting in Anthropic's own disclaimer, not a critic's. The financial-services repository states plainly that its agents "draft analyst work product for review by a qualified professional. They do not make investment recommendations, execute transactions, bind risk, post to a ledger, or approve onboarding." The company that trained the model won't ship it as an unsupervised financial agent. That's the honest version of the exact hallucination risk Nishihara is describing, coming from the one party with the least incentive to admit it.
When this is worth building
- You own data a reference architecture can't see. A firm's historical deal flow, proprietary underwriting criteria, or a narrow regulatory workflow in one jurisdiction isn't in FactSet or Moody's, and it isn't in Anthropic's connector list either.
- You build the verification layer, not the drafting layer. Software that catches a compounding error before it reaches a client deck is a narrower, harder problem than generating the deck, and it's the one gap both the free repo and the funded co-pilots leave explicitly open.
- You sell the review workflow, not the automation. Anthropic's own framing, "draft... for review," is the honest shape of this market right now. Software that makes that review faster and more reliable is a real product; software that pretends the review step is gone isn't.
When it isn't
Skip a general-purpose finance co-pilot competing head-on with Rogo, Hebbia, or the free agents Anthropic already ships. That's a market where the model vendor undercuts your price to zero and two well-capitalized startups already have hundreds of institutional customers between them. The category isn't short on horizontal tooling. It's short on the narrow, defensible layer around it.
Related: Is it worth building a vertical AI agent in 2026? covers the general build-vs-buy logic this is a specific case of. See also is an AI ROI-measurement tool worth building for the adjacent trust-layer problem, and is it worth building an AI accounting tool for the closest sibling vertical.
Frequently asked questions
Is an AI agent for financial services worth building in 2026?
Skip the horizontal co-pilot. Anthropic's own reference repository gives away 11 agents and 30-plus skills for free under Apache 2.0, wired into real data providers, at 37,300 stars, while Rogo ($2B valuation) and Hebbia ($700M valuation) already sell funded versions of the same idea. The wedge is proprietary data and the verification layer.
What exactly did Anthropic give away for free?
Eleven named agents for pitch decks, KYC screening, GL reconciliation, model building, and more, plus 30-plus slash-command skills and 11 external data connectors including FactSet, Moody's, and LSEG, all Apache 2.0.
How much has Rogo raised and what does it actually do?
A $160M Series D at a $2B valuation, up from $750M sixteen weeks earlier. It's an AI co-pilot in Excel, PowerPoint, and Word serving 250-plus institutions with 10,000-plus workflows run weekly at 95% engagement.
Why hasn't AI already replaced financial analysts if the tools are this good?
Data scarcity. Anyscale co-founder Robert Nishihara: "Large language models do well when you collect a lot of data, and we don't have nearly as much data for real-world tasks." Specialized workflows were never scraped from the public internet, so small errors compound across multi-step tasks.
Does Anthropic's own repo trust its agents to work unsupervised?
No. Its own disclaimer says the agents "draft analyst work product for review by a qualified professional" and do not make recommendations, execute transactions, or approve onboarding on their own.
What would actually make a financial-services AI agent worth building in 2026?
Proprietary data a horizontal agent can't see, or the verification layer that catches compounding errors before they reach a client, both narrower and less crowded than the co-pilot itself.
The free pack: 100 AI ideas actually worth building, each with the receipts and a clear verdict. No fake MRR screenshots.