Part of our series on Claude for finance teams. The surface map is in Claude Has Four Surfaces.
Treasury is the finance function with the least tolerance for "the AI got creative." Money moves, or it doesn't; it lands in the right account, or it doesn't. So let's start with the rule that makes everything else in this playbook possible:
Agents never move money. Ever. In Artifi's design, an agent can see balances, match transactions, propose a payment run, and prepare the bank file — and a named human approves and transmits. That's not a temporary limitation while the models improve. It's the design, for the same reason your bank requires a signature: responsibility doesn't delegate.
With that boundary fixed, a surprising amount of treasury's actual week becomes conversational.
Scenario 1: Cash visibility without the morning spreadsheet
Most mid-market treasury starts the day the same way: log into three or four bank portals, copy balances into a spreadsheet, eyeball the position. With bank connectors feeding one ledger — Wise, LHV, and Swedbank directly, Stripe for card receipts, and 3,000+ banks through Salt Edge — the morning ritual becomes:
You: Cash position across all entities, by currency. Claude: €412k total: €310k EUR across LHV and Wise, $84k USD on Wise, kr 142k SEK on Swedbank. Tallinn entity is down €38k week-over-week — driven by the Q2 VAT payment on Monday. Three customer receipts totaling €21k landed overnight and are matched.
Statements sync on schedule with encrypted credentials; every line lands in the ledger with full history. The point isn't the dashboard — every TMS has a dashboard. The point is the follow-up question, which a dashboard can't answer.
Scenario 2: The payment run an agent prepares — and a human sends
The weekly payment run is the clearest division of labor in this whole series:
- A payment-proposal agent runs on schedule: collects approved bills due in the window, checks balances, applies your rules (priorities, payment terms, early-payment discounts worth taking), and produces a proposed batch.
- A human reviews the batch — every line traceable to its bill and its approval — and signs it through the approval lane.
- The approved batch becomes a bank file (a Wise CSV today), and a human transmits it.
What the agent removed: assembling the run, checking due dates, cross-referencing approvals — the two hours of clerical assembly. What it didn't remove: the decision and the send. Treasury teams tend to find this division exactly right, because the assembly was never the part they wanted to own.
Scenario 3: Reconciliation as a background process
Cash application and bank reconciliation are treasury-adjacent everywhere and treasury-owned in plenty of companies. The 3-pass matching agent handles this as recurring work: exact matches confirm automatically (amount + date window), fuzzy matches queue for review with a confidence score, and true exceptions — the wire with the mangled reference — surface to a human with full context. In live runs, roughly nine in ten lines clear without a human touching them, and the lines that do reach you arrive with their candidate matches already attached. (Watch a real reconciliation run.)
Scenario 4: The questions in between
The connective tissue of treasury — when does the office lease payment hit, what's our weekly burn, what FX exposure are we carrying on USD revenue vs EUR costs, what happens to runway if collections slip 15 days — is Chat and Cowork work against the live ledger. Claude is strong at exposure visibility and scenario arithmetic, and it will happily produce the 13-week cash view as an Excel artifact you can flex.
What's real today — and what isn't
Real, shipping now: multi-bank, multi-currency, multi-entity cash visibility in one queryable book; scheduled statement sync with encrypted credentials (AES-256-GCM, org-scoped keys); 3-pass autonomous reconciliation; agent-prepared payment proposals with human-signed approval lanes and Wise CSV transmission; full per-row audit trail including everything agents did.
Not there yet — and we'd rather tell you: autonomous payment execution — deliberately absent, see the first paragraph; FX hedging — Claude can show you exposure but there's no hedge execution, no derivatives, no trading integration, and we have no plans to let a language model near one; investment and deposit management — out of scope; direct host-to-host connectivity with every bank — coverage is Wise/LHV/Swedbank native plus Salt Edge's aggregation; if your bank isn't reachable through those, statements import manually; intraday liquidity forecasting at large-enterprise depth — the 13-week view is honest mid-market tooling, not a Kyriba replacement.
How treasury teams adopt this
Visibility first — connect one bank account in bridge mode beside whatever you run today and let the morning-position conversation prove itself. Reconciliation second, because the agent's match rate is measurable within one statement cycle. The payment run last, because it touches money and deserves the most scrutiny — run it in parallel with your current process for a cycle and compare.
Real runs with screenshots in the walkthroughs; the control philosophy behind "agents propose, humans sign" is in The Thinking Behind Artifi.