AI can draft the finance deck. The footnote still has to survive.

OpenAI says the new workspace includes datasets from Daloopa, PitchBook, LSEG News and Crunchbase, along with tools for research, financial modeling and client materials. It describes granular citations meant to let a banker trace a claim or figure back to its source as the analysis develops.

That is a narrower and more useful claim than “AI does finance.” A person can spend a painful amount of time locating an earnings-transcript sentence, checking which period a figure covers, or finding the note that explains an adjustment. Less scavenger hunting can mean more time deciding whether the comparison makes sense.

The announcement is still a product announcement. It describes a system intended for eligible financial institutions and shaped with Morgan Stanley and Evercore; it is not independent evidence that a particular firm will make better investment decisions or send fewer corrections to clients.

In finance, two sources can support a sentence and still answer different questions. A company filing may tell you what management reported. An earnings call may explain a change in tone or timing. A data-provider estimate may use a different cutoff date. The citation should make that distinction easier to see, not hide it behind a blue underline.

A good AI assistant could help by carrying a few plain labels with a conclusion: reported or estimated; period covered; source date; adjustment made; and what remains uncertain. That is enough for the analyst who inherits the work to decide whether the number belongs in a client note, an internal draft, or nowhere yet.

This matters outside a bank too. A founder asking why renewals slipped, a nonprofit reviewing a budget, or a small team deciding whether to hire all have the same problem in miniature. The faster the answer arrives, the easier it is to forget that the definition of the metric might have changed underneath it.

Nobody needs an assistant to make an obviously impossible claim. The costly mistake is the reasonable-looking summary that quietly combines different reporting periods, treats an adjusted measure as ordinary profit, or turns an analyst's estimate into a company fact.

OpenAI says its product can surface a reconciliation and notes behind adjusted EBITDA. Good. The test is whether that trail remains intact after the work becomes a chart, a slide and then an email. A footnote that disappears at export is not a source trail. It is an invitation to repeat a number with more confidence than the original work earned.

Before trusting an AI-generated financial brief, pick one inconvenient figure and trace it backward. Can you open the underlying document? See the relevant date and definition? Tell whether the assistant calculated something or repeated it? If not, keep the result as a lead, not a conclusion.

Ivy Chen's test is practical: if a teammate takes over a model on Friday afternoon, can they see what changed without asking the original analyst to reconstruct the work? The answer should carry the input, the adjustment and the question that still needs a human call. Otherwise the time saved in research becomes time lost in cleanup.

Theo Marlow keeps the source boundary sharp. The release establishes what OpenAI is offering and which data providers it names. It does not establish the accuracy of every output, the completeness of a firm's connected data, or the suitability of a conclusion for a client. A citation trail can make those questions easier to inspect. It cannot answer them by itself.

Those are not competing standards. One protects the person inheriting the work; the other protects the person relying on it. Both are how a finance assistant earns the right to make the next draft faster.

Do not ask an AI finance tool for “the answer.” Ask it to show the path to an answer, then make the path survive the format change.

For one recurring report, put a short check beside each important number: source link, date, definition, calculation or adjustment, and owner for any unresolved judgment. It will feel fussy once. It is much less fussy than explaining next week why an impressive slide used the wrong quarter.

AI agents can give analysts back time when they cut the searching and formatting. The work that deserves to remain human is deciding what the number means, what it leaves out, and whether it is ready to influence someone else.