Using Claude for Finance Work: Four Use Cases That Hold Up

The rule that makes any of this usable

In finance, a plausible number is worse than no number. Someone will act on it, and the error surfaces in a board pack, a filing, or a covenant calculation.

So the division of labor is strict. The model produces structure: layout, chart selection, narrative, formatting, the code that renders a view. Your systems produce values: the ledger, the data warehouse, the FP&A platform, the source of record for whatever is being discussed. The model never types a figure it was not handed.

Enforce this mechanically rather than by instruction. Generated views should read from a data file or query result, so the code contains references rather than literals. Anyone reviewing can check that every number traces to a source, and no one has to trust that the model behaved.

That single constraint is what separates an AI finance workflow that survives audit from a demo that impresses in a meeting and is never used again.

Four use cases that hold up

These are the ones where the work is genuinely repetitive and the judgment stays with a person.

Variance dashboards. Actual against budget or forecast is mechanical to compute and tedious to present. The agent builds the view: the breakdown by cost centre, the waterfall from prior period, the drill-down into the largest movers, refreshed each close from the same query. What it must not do is explain why a variance happened, because that requires knowing about the hiring freeze and the delayed contract, which is exactly what your finance team knows and the model does not.

Board and investor memo drafts. The structure is stable quarter to quarter: performance against plan, key drivers, risks, asks. An agent that has last quarter's memo and this quarter's figures produces a complete first draft in minutes. You rewrite the judgment paragraphs and keep the hours you would have spent on structure and formatting.

Model documentation. Every material model is supposed to be documented and almost none are current. An agent can read a workbook or the code behind a model and produce a description of inputs, assumptions, calculation flow, and outputs. It is derived from the artifact itself, so it is accurate about mechanics, and a reviewer only needs to confirm intent.

Scenario comparison views. Finance teams run cases constantly and then present them badly, in a wall of numbers. Given the outputs of each case, an agent builds a comparison view showing what changes and what does not, with the assumption deltas stated plainly. The cases come from your model. The presentation comes from the agent.

Why the first attempt usually disappoints

The initial output looks polished and is subtly wrong for your organization. It uses a cost centre grouping you retired, it applies a materiality threshold nobody agreed to, it labels a line with a term your CFO does not use, and it formats currency in a way your board pack never does.

None of that is a capability failure. The model does not know your chart of accounts, your reporting calendar, the definition of adjusted margin your company actually uses, or the fact that one business unit reports on a different cycle. Those definitions live in a policy document, a finance handbook, and the heads of two people in the team.

So you correct it in the session, get a good result, and lose all of it when the session ends. Next quarter the same corrections happen again, which is why so many finance AI pilots stall after the first impressive demo.

Ground the work in your own definitions

The durable fix is to make your finance definitions retrievable. RDK indexes your policy documents, chart of accounts, close calendar, metric definitions, and last quarter's approved outputs as encrypted private chunks on the RDK network. Encryption matters here more than in most domains: this is unreleased financial information, and it cannot sit in a shared prompt or a vendor's training pipeline.

Before the agent builds a view, it retrieves the definitions that apply. It uses your adjusted margin definition rather than a textbook one. It groups by the structure you actually report on. It formats the way your board pack does, because last quarter's approved pack is indexed and it can see the convention.

The review burden drops accordingly. A first draft that already respects your definitions is edited, not rebuilt, and that difference is where the hours are.

The cost side follows the same mechanism. With stacked retrieval, a private index answers 40 to 65 percent of queries, the public network of published chunks adds 15 to 20 percent, and the LLM handles the remaining 5 to 10 percent as fallback, so token spend drops 80 to 90 percent. Close periods are exactly when usage spikes, which is when that ratio is worth the most.