From curious to a controlled, running agent
Everything here follows one rule: an agent does the work, a human signs off before anything touches a filing, a financial statement, or a material number. The four steps below take you from never-touched-an-agent to a plan a partner can approve.
Download a sample bank statement and ledger we seeded with six discrepancies, paste one ready-made prompt into Claude, and grade the agent's reconciliation against the answer key. No client data, no setup, no cost. After this, 'agent' isn't an abstraction.
Start the tutorial →Two ways: install one of Anthropic's ten prebuilt Claude Finance Agents from the library (GL reconciler, month-end closer, statement auditor…), or answer three questions (the job, the data, the boundaries) and the Agent Builder writes the spec, the guardrails, a run-today starter prompt, and a downloadable skill file.
Open the Agent Library →One agent saves hours; the plan is where the control lives. Six questions generate 2–3 architecture options: swimlanes by team, working agents with human-review gates at every filing-bound step, a conservative business case, a phased rollout, and integration readiness for your actual software. Refine it in plain language, then export a print-ready package for IT or a consultant.
Build your plan →Every plan has its own link. Send it to a partner or IT and they can review it, make their own copy to edit, and send a revised version back — the original stays untouched.
Start from a template →Agents aren't the risk. Agents without controls are. Every design here builds the sign-off in from day one, so adopting AI strengthens your control environment instead of bypassing it.