Claude and effect

When Amazon-backed Anthropic unveiled “Claude for Financial Services” on July 15th, it was not just adding yet another vertical to the alphabet soup of generative-AI products. The “Financial Analysis Solution” (to give its formal title) promises analysts a single chat interface that can rummage through everything from live market feeds to data stashed in Snowflake warehouses, cite its sources on demand and then crank out Python code or Monte-Carlo simulations without breaking a sweat. In short, Claude aims to be the junior associate who never sleeps – only cheaper, faster and blessed with a longer memory.

At first sight, the offer looks much like rival corporate chatbots. Yet Anthropic has sprinkled three extra ingredients. First comes a thicket of direct connectors to big data troves – FactSet, PitchBook, Morningstar and S&P Global, among them – wired in via its new “model-context protocol,” so the bot can quote chapter and verse rather than hallucinate. Second, customers get “Claude Code”, a heftier version of the firm’s Opus and Sonnet models that, according to tests by FundamentalLabs, solved five of seven Financial-Modelling-World-Cup spreadsheets and scored 83 per cent overall. Third, the package ships with expanded usage limits that matter during results season, when prompts spike, and compliance departments grow twitchy.

From hype to homework

Anthropic insists that these bells and whistles translate into genuine out-performance. On Vals AI’s new Finance-Agent benchmark, Claude’s latest models top the accuracy tables, edging past OpenAI’s corporate GPTs and Meta’s Llama brood. Early adopters supply the marketing copy. Bridgewater’s in-house AIA Labs credits Claude with producing charts and code “with the precision of a junior analyst”; Norway’s $1.6 trn oil fund, NBIM, reckons the bot already saves staff the equivalent of 2,13,000 labour hours each year, Commonwealth Bank of Australia says it is central to fraud-detection and customer service; and AIG claims underwriting turnaround times have fallen five-fold while data accuracy has risen from 75 to over 90 per cent.

Kate Jensen, Anthropic’s chief revenue officer, calls finance a “natural first-choice sector” because analysts obsess over both speed and audit trails. The company is betting that those demands will let it charge premium prices even as general-purpose chatbots get cheaper. That logic also explains why Anthropic is partnering with consultants – Deloitte, KPMG, PwC and others – to wrap the model in regulatory and integration services.

A well-funded assault

Cash for the push is plentiful. A $3.5 billion Series E in March lifted Anthropic’s valuation to $61.5 bn. Amazon, already on the hook for $8 billion, is mulling yet another multi-billion slug, a sign that the cloud giants see enterprise AI as the next battleground after consumer chatbots. The money buys compute and – critically – GPU time, both of which the firm needs to keep retraining specialised models for sectors beyond finance.

That is prudent because the field is suddenly crowded. BloombergGPT has been running inside the Terminal since April, offering subscribers a 50-billion-parameter model trained on proprietary feeds. OpenAI, meanwhile, has embedded GPT-4 throughout Morgan Stanley’s wealth-management desks; the bank boasts 98 per cent internal adoption. If Claude is to escape price wars, it will have to keep sprinting up the value chain – perhaps by automating ever chunkier swathes of modelling or by guaranteeing “write-once-audit-forever” documentation that regulators adore.

Digital clerks or pink slips?

For Wall Street’s human inhabitants, the stakes are personal. Dario Amodei, Anthropic’s chief executive, has suggested that entry-level white-collar roles may shrink by half. Ms Jensen tempers the forecast, arguing that analysts will simply be expected to “be better” – doing more creative work once Claude handles the drudgery. That is cold comfort to Excel jockeys who remember what happened to floor traders after the rise of algos.

Sceptics also note that benchmarks rarely reflect the messy data and lawyerly footnotes that litter real financial filings. Even Vals AI’s latest tests show every model, Claude included, scoring below 50 per cent on some nuanced tasks. Nor is integration painless: banks must pipe confidential data through compliance gates, decide who audits the outputs and, in some jurisdictions, convince regulators that an American unicorn will not learn sensitive European investor positions.

Yet history suggests that when a cheap tool can shave a few basis points off costs, finance grabs it. If Claude’s promises of speed, traceability and source links hold up in the wild, the industry could become the guinea pig for large-language models that finally earn their keep. Wall Street’s newest recruit does not wear a hoodie; it lives in the cloud and knows where every spreadsheet cell is buried.

Also read: The thinking investor’s advantage