Every "AI for finance" piece in the last year talked about capabilities: reconciliation agents, expense categorization, treasury copilots. None of them explained how a finance team is supposed to budget for software that bills per token. Or how a controller is supposed to sign off on an automated journal entry when the audit log says the human did it. The capability conversation skipped the two questions finance actually has to answer before any of this ships: can I defend it under audit, and can I forecast what it costs.
Dock is built around those two questions.
Three finance workflows where Dock fits
Month-end close. A shared workspace with a typed table per ledger account. Reconciler agents read source-of-truth balances, write attributed rows for variances, and flag anomalies for human review. Because every agent is a first-class principal with its own identity, the row written by your reconciler agent carries the agent id AND the owning controller's id. The close package an auditor sees later is the table itself, not a reconstruction from chat logs.
Expense compliance. A doc surface holds the policy. A table surface holds receipts. A categorization agent reads incoming receipts, proposes a GL code, and writes a draft row. Anything above the policy threshold gets routed through a consent gate before posting. The agent cannot post the entry on its own; it surfaces a confirm token, the controller approves, the agent re-calls. The same dangerous-ops contract Dock uses for plan upgrades extends naturally to any operation that touches the ledger.
Budget tracking. A shared dashboard table with one row per department. Agents update actuals on a cadence. Finance sets the columns and the shape. The agent cannot invent a new column or skip a required one because every write goes through the same shape cap the human writes do. Drift is impossible at the schema level.
The compliance angle
PCAOB AS 2201 expects an integrated audit of internal controls over financial reporting, and the controls only count if they are documented, tested, and attributable (PCAOB AS 2201). When an automated control writes to the ledger, "the system did it" is not an acceptable attribution.
Dock's dual-keyed audit log records the agent principal id AND the owning user id on every privileged write. A SOX walkthrough becomes a query. An internal audit ticket asking "which agent posted this entry and who is accountable" is two columns away. A regulator inquiry into AI-assisted journal entries gets a coherent timeline instead of a forensic reconstruction.
The cost-containment angle
a16z's 2025 study of 100 trillion tokens across OpenRouter found agentic inference is the fastest-growing usage shape: long sequences of tool calls, retrieval, and iteration, with cost scaling per step (a16z, State of AI). The implication for finance is direct. A reconciler agent stuck in a retrieval loop on a per-token plan does not stop on its own. It bills until someone notices.
Dock's pricing is flat: Free $0, Pro $19/month, Scale $49/month, per organization. Spawn three agents on Free, ten on Pro, thirty on Scale. The bill does not change when an agent loops, because Dock does not charge for the loop (why we kept flat pricing). The substrate has the cap built in: shape caps on writes, consent gates on money-moving operations, and a monthly API ceiling that 402s a runaway before it can drain the workspace.
Finance teams pick Dock because the worst-case agent is bounded.