Accounting should run itself.
We are building accounting general intelligence. Our goal is reliable, trustworthy books that close themselves.
Enterprises already have plenty of accounting software. What holds it together is people: carrying context between systems, interpreting agreements, resolving exceptions, and remembering why things are done a certain way.
Enterprise software has its own version of the xkcd standards joke: fifteen systems that don’t work together, and a sixteenth sold to tie them all together. We don’t think another AP or AR application is the answer.
We handpick our clients and go deep. We work alongside their finance and operating teams to understand how the business actually works: its contracts, accounting policies, approval paths, exceptions, and the knowledge that has never made it into a manual.
Then we embed our agents in that context, working across the systems the enterprise already uses. An agent needs to know what this company agreed to pay, who can approve a deviation, and what evidence makes an entry safe to post.
That work has taken us deep into an enterprise software provider, a global retail chain, a major ecommerce marketplace, a seafood processing company, a global cybersecurity company, and a major logistics provider.
Each is an enterprise, or part of a group, with more than $1 billion in annual revenue. Our approach has been proven across this variety of industries, geographies, and levels of operational complexity, working through the systems, accounting requirements, and exceptions specific to each.
Our agents’ work helping Big 4 auditors has been especially formative. It has taught us to approach an accounting decision from the other side of the table: can someone independently follow the evidence, question the treatment, and understand how the number was reached?
We carry that experience into cadel.ai: an approach tested in the details of very different enterprises, and shaped by the people whose job is to question the books. That is the grounding trustworthy, self-closing books require.
The living ledger
Things we’re building, testing, and learning. Open an entry to read more.
Can open-weight models run accounting workflows?
We replaced the models inside our proof-of-delivery workflow and ran the same 60 cases again. The open-weight configuration was close on accuracy and cost, but took roughly twice as long. The workflow around the model matters.
Read the researchThe engine beneath Cadel
An invoice, a contract, and a bank statement look different. The work around them often comes down to the same operations: extract, classify, match, reconcile, validate, and analyze. We wrote about building these operations once, then composing them into accounting workflows.
Read the researchDoes the same invoice get the same answer twice?
Getting an invoice right once is only the beginning. We reran the same freight-invoice cases with caching disabled to measure how often the accounting decision changed. Repeatability deserves its own test.
Read the researchCan coding agents do accounting?
Seven AI solvers, 35 synthetic freight-invoice cases, one shared scorer. We compared the accounting decisions, the evidence behind them, and the cost of getting there. Matching a total is easier than explaining every line.
Read the research
Backed by South Park Commons and Tanglin Venture Partners.