AI agents only make finance better when the foundations are built for them: clean data, explicit policies, clear thresholds and people who review. Here is how we build that, step by step.
Most of this happens in your first three weeks. The shadow close runs alongside your existing process, so nothing changes until the agents have proven themselves on your numbers.
We map every entity, bank, system and money flow, and find where data breaks today.
Output: systems map and data-gap listWe redesign the chart of accounts, dimensions and close calendar so every transaction has one right home.
Output: chart of accounts and close calendarYour accounting policies, approval rules and checklists become standing instructions that every agent reads before it acts.
Output: policy library for agentsWe set materiality thresholds, segregation of duties and review queues, so it's clear what posts automatically and what waits for a person.
Output: control matrix and review queuesAgents run a full close next to your current process. We compare line by line and tune instructions until they match.
Output: shadow-close reportAgents run the work. Your controller and CFO review it. Every month we refine the instructions from what the review catches.
Output: a five-day close, every monthEvery judgment area, such as prepaids, revenue recognition, capitalisation or accruals, is written down once as a standing instruction. Agents follow it, reviewers check against it, and auditors can read it.
policy: prepaid-expenses owner: controller applies_to: [software, insurance, rent] capitalize_if: amount: ">= 5,000" service_term_months: "> 1" amortize: straight_line_monthly evidence: [invoice, contract] auto_post_below: 25,000 review: above_25k: controller above_100k: cfo # every agent reads this before it touches a prepaid
Every agent reports to a named person. Nothing material reaches your books, your board or a buyer without a human signing it.
Anthropic's October 2026 study of its own workforce is the clearest picture yet of AI-first work. These are the findings we design around.
Agents worked six times as many hours as the people directing them by August 2026, after passing human hours in March.
of interactive AI session time was idle, waiting on a human.
"That 20% that was strategic becomes 90% of the job," as one employee put it.
standing instructions for agents a year after there were a few hundred.
Meta-work, such as building agent tools, writing specs and monitoring pipelines, rose to 22% of interactive sessions.
code commits per engineer versus six months earlier, with no difference in revert rates between agent-written and human-written code.
Findings from Wilmers, Catipon, Massenkoff and McCrory, "How employment is changing at Anthropic", Anthropic, October 2026. The figures describe Anthropic's own workforce, and its authors note Anthropic is an unusual case. The "what it means" lines are how we apply these findings to finance. They are not results from the study.
Read-only access in. A written review of your books, cash and top risks out, within seven days.