A €4,300 design-tool invoice arrives at a Light customer. The Bill Agent codes it to 642000 and routes it for approval, exactly as that company's policy directs. Nobody typed the account. Nobody vetted a suggestion before it posted. An agent holding its own identity did the work, inside the same approval limits a person would work inside, and wrote every step to the audit trail on the way past.
"AI agents for finance" is a phrase doing a lot of imprecise work in this market, so it is worth being exact about what it means here. Not a chatbot answering questions about the ledger. Named agents, each responsible for 1 job end to end. The Bill Agent codes bills. The Contract Agent manages contracts. The Accrual Agent books month-end entries. The Bank Rec Agent clears the bank. Each carries a principal of its own and its own line in the record, the way an employee does.
ARR at customers that have completed real audits on Light, with agents posting to the ledger throughout
average time saved on repetitive admin work at teams running a Custom Agent written for their own workflow
of the journal population an auditor pulls through the API, not a sample, because every action is already attributable
The second layer is the interesting one
Agents that act are the easy half. Agents that check the agents are what makes the first half safe to run.
A separate class of audit agents is read-only by design. It inspects everything and changes nothing, verifying outputs against policy and, more usefully, verifying the instructions the working agents run on. Each month it produces a control report: what was checked, what passed, what was flagged.
The example that shows why the second part matters: an admin edited the Accrual Agent's instructions to skip anything under €20,000, to save time at close. Output-checking would have found nothing wrong, because every entry the agent produced was correct. The audit agent compared the instruction to policy, caught the mismatch, and flagged it before the missing entries became a restatement. Astra, the always-on analyst, watches everything agents and humans post across the ledger and raises what needs attention.
This is why auditors at customers with $500M ARR have completed engagements with agents posting throughout. The record was assembled at the moment of the transaction rather than reconstructed for the request list.
What it feels like from inside the team
At Alva Labs, the finance manager describes the change without reaching for the word automation. "It's like having a team of employees, but it's actually just agents doing the repetitive and manual work for me, and the structured analysis too," says Thobias Cogrell.
"It's like having a team of employees, but it's actually just agents doing the repetitive and manual work for me."
Thobias Cogrell, Finance Manager, Alva Labs
At Famly it reads as relief rather than efficiency: 1 ledger, agents on the routine work, and a company that no longer dreads expense day. At KeyShot the agent has become the company's shared reference point for its own numbers. "The whole company can post questions in Teams, and the Light agent answers," says CFO Jeppe Bygholm.
When the job does not have an agent yet
The 4 shipped agents cover work every finance team has. The work specific to 1 company does not fit a shipped agent's job description: the vendor whose spend needs watching, the cards that need freezing the day someone leaves, the receipts nobody has time to chase. That work gets written instead. An admin describes the job in plain language, sets a schedule, and a Custom Agent runs it from then on under its own identity, delivering into Slack or into Light. Teams running these save 80% of their time on average on exactly that category of work.
At Dreamdata the number is further along than the average.
"We've automated 95% of bookkeeping with Light and some simple custom agents. This thing is powerful."
Andreas Jørgensen, Controller, Dreamdata
Peter Egehoved, the company's COO and CFO, describes what the time becomes: "They allow me time back to do more financial planning and strategy."
Shipped
4 agents, 1 job each
Bill, Contract, Accrual and Bank Rec Agents work end to end inside the roles and approval limits a company sets. Every action lands in the audit trail.
Verified
Read-only, watching everything
Audit agents check outputs against policy and check the instructions behind them, producing a monthly control report of what passed and what was flagged.
Written
Custom Agents for the rest
An admin writes the instruction, sets the schedule, and the agent runs under its own identity. Average time saved on that work: 80%.
The role this creates
A rules engine does what it was configured to do and stays quiet about everything else, including whether the configuration still matches policy 6 months later. An agent holds an identity, gets checked by a layer that reads intent as well as output, and leaves a record built for an auditor rather than assembled for one.
The more consequential change is to the person. Andreas at Dreamdata does not operate a bookkeeping process. He writes the instructions that run it and reviews what came back. That is a different job from the one the title used to describe, and it is the job the best finance people are moving into: less operating, more building. The teams furthest ahead here are not the ones who bought the most automation. They are the ones who started writing it.
The open question was never whether an agent can do the task. It is whether the task has a name yet, and if it does not, how long before somebody on the team writes 1.
See the full workforce, the Custom Agents any admin can write, or book a demo.