For a great many financial institutions the first encounter with generative artificial intelligence felt almost deceptively straightforward, because a professional would open a chat window, paste in a question and read back a polished answer, only to be left working out, more or less alone, whether the result represented a genuine gain in productivity or merely an elegant distraction from the real work.
That early phase of curiosity and cautious experiment is now drawing to a close, and the question that increasingly matters is no longer whether a model can produce a convincing paragraph on demand, but whether artificial intelligence can actually operate inside the working life of financial professionals, where the real labour is spread across spreadsheets and research databases, investment memos and pitch decks, compliance files and the client reporting tools through which a firm ultimately answers for itself.
That environment is governed by formulas and permissions, by data lineage and audit trails, by professional judgement and a kind of institutional memory that no single person fully holds, and while a chatbot sitting off to one side can certainly be useful, a chatbot that lives outside the workflow tends in practice to become just another tab to keep open, another invitation to copy and paste, and on occasion another quiet layer of operational risk that someone will eventually have to account for.
The opportunity therefore lies less in the chat window itself than in what might be called AI extensions, the add-ins and connectors and skills and agents that carry intelligence into the very tools where financial work is already being done, rather than asking the work to come out and meet the machine halfway.
It is in this direction that OpenAI and Anthropic are now moving with conspicuous speed, away from general-purpose assistants and towards workflow-aware systems that can read files, manipulate spreadsheets, connect to approved data sources and carry out repeatable financial tasks; OpenAI has approached the market as a broad productivity layer, with ChatGPT increasingly embedded in spreadsheets and enterprise connectors, while Anthropic has taken a more explicitly finance-oriented path with Claude for financial services, Claude for Excel and connectors built expressly for investment research, due diligence and modelling.
What is emerging is a new layer in the financial software stack, and it is worth being precise about it: not a replacement for core banking systems, nor a substitute for portfolio management platforms, nor, at least for now, an autonomous investment analyst, but rather a flexible intelligence that narrows the distance between raw information and the informed judgement it is supposed to serve.
Financial teams need such a layer because so much of what they do remains trapped in the repetitive business of moving information from one format to another, as analysts pull figures out of PDFs, reconcile numbers that refuse to agree between systems and check, again and again, whether a number on a slide still matches the one in the latest spreadsheet, all of which is intellectually serious work and none of which is operationally efficient.
The deeper problem is not time but fragmentation, the scattering of a firm’s working knowledge across market terminals and Excel models, CRM and ERP systems, SharePoint folders and inboxes, so that even when the relevant figure can be found the context that gives it meaning is often missing, and a revenue number divorced from its assumption, a valuation multiple stripped of its peer-group logic or a client report that cannot be traced to its sources can mislead, do real damage or curdle into a reputational risk no one intended to create.
This is precisely where AI extensions become more interesting than a standalone chat, because a model embedded inside Excel can explain a formula, clean a table, build a scenario or surface an inconsistency a tired eye would miss, a connector to a financial database can spare hours of manual searching while drawing trusted sources into a single environment, and an assistant wired into a firm’s internal documents can retrieve institutional knowledge that might otherwise lie dormant in old folders and forgotten emails; the promise is not magic but compression, of the time spent searching, the labour spent formatting and the distance between source material and considered judgement.
The wider market is converging around three closely related ideas, integration, governed access to data and repeatability: integration because finance professionals have little appetite for abandoning Excel, PowerPoint, Outlook or Teams simply to please a new tool; governed access because a general model is not enough once users need permissioned, auditable access to trusted sources; and repeatability because a genuinely useful assistant must do more than answer well once, following a consistent process whether the task is a discounted cash flow model, a credit memo or a due diligence summary.
OpenAI’s strongest proposition is breadth, for ChatGPT is increasingly positioned not as a conversational interface but as a general work platform spanning documents and spreadsheets, data analysis and enterprise search, which makes it attractive to organisations that would rather install a single layer of intelligence across finance, legal, risk and operations than commit to a tool confined to one niche, and in practice it can summarise an earnings release, draft an investment note, explain the logic buried in a spreadsheet or query an approved knowledge base without ever leaving the flow of the day.
Its great advantage is flexibility, since it remains useful even when a task fits no predefined workflow, moving without complaint from a portfolio commentary to a board memo to a client email, which makes it well suited as a first layer of adoption; but the disadvantage is also flexibility, for a broad assistant in the hands of users relying on improvised prompts and unverified sources can produce strikingly inconsistent results, and in a regulated environment ChatGPT should never become the system of record or the final authority on a valuation, a compliance call or an accounting decision, which is why it is best understood as a horizontal productivity layer that demands serious investment in governance and human review.
Anthropic’s approach feels markedly more specialised, since Claude for financial services is framed explicitly around due diligence and market research, competitive benchmarking and portfolio deep dives, modelling and the memos and pitch decks that fill an analyst’s calendar, and because Claude can create and edit the Excel spreadsheets, presentations, Word documents and PDFs that are, when one is honest about it, the actual artefacts in which financial work lives and is judged; its combination of spreadsheet capability, finance-specific skills and connectors is what makes it compelling for investment teams, whose work already spans a constellation of linked documents and who need an assistant that helps structure the analytical workflow rather than merely producing elegant prose.
A credit memo, after all, is never simply a document but a chain of reasoning connecting financial data to business risk, collateral to covenants and all of it to judgement, which is what makes Claude so attractive to investment banking, private equity, asset management and credit teams, sitting closer to the analyst’s workbench than to any generic assistant, though such specialised workflows demand disciplined implementation and still depend on human validation, all the more so when outputs influence an investment, a line of credit or a decision a client will live with.
The strategic question, in the end, is not whether a firm should reach for extensions or build custom products of its own, because for almost everyone the answer is both: extensions are the natural place to begin when the objective is productivity, being quicker to deploy, easier to test and far cheaper than a full build, while custom products become necessary the moment a workflow turns out to be core or regulated, so that client onboarding, anti-money-laundering checks, credit decisioning and audit-critical reconciliations require real integration, deterministic controls and a clear line of accountability, and the rule that emerges is a simple one, to buy the productivity layer and build the control layer.
