Every BOV Deck Is Stale by Thursday. Claude in Excel Fixes That.

The workflow every investment sales and debt team runs has two artifacts with different metabolisms. The Excel model changes daily: a comp lands, the lender revises proceeds, the seller trades a concession. The deck changes when someone has an hour, because updating it means finding every figure the model moved, retyping it in three places, and hoping the sensitivity table on slide 14 was not missed. The result is a permanent, low-grade version conflict between the two documents a client sees, and everyone in the industry has watched a committee catch a slide that disagrees with the model behind it. The reason this survived a decade of automation talk is that linking Excel to PowerPoint natively has always been brittle, and the deck is not really a projection of the model: it is a designed document that quotes the model in a hundred places.

What changed is that language-model tooling now operates inside the documents rather than beside them. Anthropic announced Claude for Excel on October 27, 2025 as part of its financial services expansion: an add-in through which Claude can, in the company’s words, read, analyze, modify, and create Excel workbooks, modify a spreadsheet while preserving its structure and formula dependencies, debug cell formulas, and populate templates with new data and assumptions. The same announcement describes Microsoft 365 integration that creates and edits Excel spreadsheets and PowerPoint slides and searches files, emails, and Teams conversations for context. At announcement the Excel add-in was in beta as a research preview for Max, Enterprise, and Teams users, initially capped at a thousand users, so current availability is a thing to check against Anthropic’s own pages rather than assume. Those capability descriptions are Anthropic’s claims; PSV’s read of what they mean for a CRE desk follows.

Because the retype gap is not a formatting nuisance, it is an error surface in the documents clients price off. A BOV whose slide quotes last week’s NOI is a credibility problem when the client has the current number. And the hours are real: a team producing weekly pitch materials burns an analyst-evening per deck per revision cycle on transcription that produces nothing. The desks that close the gap get something better than saved hours, they get the ability to say yes to the Tuesday-night “can you rerun it at the new proceeds” request without a Wednesday-morning deck scramble, because the deck regenerates from the model instead of being hand-reconciled to it.

The catch the demos skip: generation from the model only helps if the model deserves to be the source of truth. A workbook with hardcoded overrides buried in formula cells, three tabs named Final, and inputs nobody can find will produce a deck that faithfully reproduces the confusion at presentation quality, which is worse than the retype gap because it looks finished. This is the same lesson as every other chain PSV builds: the automation amplifies the discipline that exists, in either direction. The prerequisite work, one canonical model file, inputs separated from calculations, assumptions on a labeled tab with owners, is exactly the discipline a good analyst wanted anyway. The tooling just made it pay twice.

Name the model as the single source of truth and demote the deck to a generated artifact. The inputs are the workbook, structured so a tool can navigate it, and a deck template whose every model-quoting field is identified. The workflow: when the model changes, the tool re-extracts the quoted figures, regenerates the deck from the template, and produces a change log, which figures moved, from what to what, on which slides. The reviewer is the deal lead, and the review is the change log first, the deck second, because the log is where a wrong number announces itself. The approval gate is unchanged from every PSV workflow: no version reaches a client, a committee, or a lender until a named person has signed the regenerated deck, and the change log is filed with it.

Sequence the adoption the boring way. Start with extraction only: the tool reads the model and produces the figure inventory, and the team confirms it found everything the deck quotes, including the sensitivity table and the footnotes. Then template the deck. Then let it regenerate drafts. A team that starts at full generation on a live mandate is testing in production on a client deliverable. And on availability: whether the firm reaches this through Claude for Excel, through Microsoft’s own tooling, or through a custom build against the model file directly is a procurement and licensing question that changes quarterly, so design the workflow around the pattern, source of truth, template, change log, human gate, and let the tool slot be replaceable. That last property matters more than any vendor’s current feature list.