How we built Level
How we structure AI-assisted finance work around evidence and human review
Finance services are known for manual work, fragmented systems, and growth that depends on adding people. We built Level around a different idea: standardize the repeatable work, make every important check auditable, and preserve human judgment for the decisions that actually need it.
The constraint was not intelligence
Owners rarely lack reports. They lack a reliable operating rhythm that turns messy records into a short list of decisions. We focused on reconciliation, exception checks, margin movements, revenue leakage, spend concentration, and monthly review because those repeat every month and benefit from consistency.
Make the work inspectable
Each public skill names its inputs, calculation, possible explanations and final human decision. Synthetic examples let you inspect those boundaries before using real records. A repeatable calculation does not make an AI interpretation deterministic; ambiguous matches and incomplete source coverage still require review.
The 20 skills that survived
The public catalog covers the recurring work where structured evidence creates leverage: bank reconciliation, monthly variance, recurring subscriptions, idle-cash review, 13-week cash forecasting, AR integrity, duplicate billing, revenue leakage, margin trends, spend concentration, cost rationalization, vendor rules, remittance matching, customer margin, cost-detail footing, categorization-history health, 1099 readiness, close readiness, finance charts, and monthly CFO review.
They are not twenty prompts. Each skill defines its evidence, calculation, possible explanations, limitations, and final human decision boundary.
The surprising part was where AI belonged
AI was most useful after finance judgment had been recorded, not before. We captured how experienced reviewers investigate a variance, what evidence changes the conclusion, which explanations are plausible, and when the system must stop. AI can propose an investigation using that recorded judgment. Consistency requires testing; a finance professional checks the evidence and retains the final call.
Tools make the work easier, people make it useful
These are tools that support a finance team. AI can make a finance firm faster and easier to run, but a human with domain expertise still has to interpret the output, challenge assumptions, and talk through the decision with the owner. Software can flag an unusual margin. It cannot understand the full operating context by itself.
Now we are making the reusable layer public
We are publishing the reusable finance tools, checks, schemas, synthetic examples, and playbooks so operators and finance teams can inspect them, use them, and improve them. Start with one check, understand its assumptions, and keep a human reviewer in the loop.
Open source does not mean open client data. We publish reusable finance tools, checks, templates, and playbooks. Client records, workpapers, credentials, communications, and private configurations remain confidential.