Every accounting vendor now claims AI, which makes the claim useless as a filter. The useful question is narrower and harder to answer from a website: is the intelligence in the ledger, or beside it?
A system built after this technology existed puts an agent inside the record, acting under an identity, inside the same policies and audit trail as a person. A system built before it adds a copilot next to the record, which answers questions and drafts things a human then does.
Both demo well. Only one changes the headcount arithmetic. Here is how to tell them apart.
of bookkeeping automated at Dreamdata using the platform plus custom agents
ARR at customers that have completed real audits with agents posting to the ledger
less bookkeeping time at All Gravy's partner, while the company doubled
1. Does it act, or does it suggest
The dividing line in the category. A copilot proposes a coding and a person accepts it, which saves keystrokes and leaves the review step intact. An agent codes the bill, routes it by policy and posts, and the person reviews the exceptions.
Ask for the demo where nobody clicks approve. At Tillo, 750 to 900 invoices a month go out over the API with no triage in the UI. At All Gravy, bills arrive by email, get read and coded on arrival, and nobody opens them.
2. Does the agent have an identity
This is the question that separates real architecture from a wrapper, and it takes 1 minute to ask.
Every action an agent takes should be attributable to that agent rather than to a shared service account, with a creation and update log tying it back to the human who set it up and anyone who has changed it since. If agent activity lands in the ledger under a generic system user, the audit trail has a hole in it and no amount of model quality closes it.
3. Who checks the agents
A vendor with agents that act and nothing watching them has shipped half a system.
Look for a separate read-only layer that verifies outputs against policy and, more importantly, verifies the instructions the working agents run on. The failure mode that catches companies out is not a bad output. It is a well-executed bad instruction: an admin edits an accrual threshold to save time, every entry the agent produces is correct, and output-only checking passes it while entries go missing. Checking intent catches that. Checking artefacts does not.
4. Has anyone been audited on it
The strongest available evidence, and easy to ask for. Not whether the vendor is SOC compliant, which is about the vendor. Whether a customer has completed a statutory audit with agents posting to the ledger throughout, and what the auditor did about it.
Customers at $500M ARR have completed audits on Light, with the auditor pulling 100% of the journal population through the API rather than sampling it. That is a different conversation from a roadmap commitment.
Bolted on
A copilot beside the ledger
Answers questions, drafts entries, suggests codings. A person still performs the work, and the review step stays exactly where it was.
Built in
Agents inside the record
Named agents act under their own identity, within policy and approval limits, and a read-only layer verifies both their output and their instructions.
The tell
What happens with nobody watching
Ask what runs overnight and what is waiting in the morning. A copilot does nothing. An agent has already done the work and logged it.
5. Can your team extend it
No vendor will ever model 1 company's chart of accounts as well as that company does. A platform optimises for everyone, and a finance team optimises for itself, and that asymmetry does not resolve by the vendor working harder.
So the question is whether the finance team can close the gap without engineering. In practice that means writing an instruction in plain language, setting a schedule, and having it run under its own identity with every run visible afterwards. Dreamdata's controller reached 95% of bookkeeping automated this way, and the last stretch was work only somebody inside the company could specify.
"We've automated 95% of bookkeeping with Light and some simple custom agents. This thing is powerful."
Andreas Jørgensen, Controller, Dreamdata
6. Does the architecture allow it at all
The least visible criterion and the one that constrains everything else. An agent has to see the whole transaction. If accounts payable lives in 1 tool, expenses in another and the ledger in a third, no agent can match, chase and post across the gaps, and the automation stops at every system boundary.
This is why the strongest numbers in this category come from companies that consolidated rather than integrated. All Gravy replaced 7 systems with 1 and its bookkeeping partner now spends 30 to 40% less time on the books while the company has doubled. Oper Credits went from 8 systems to 3. Tillo collapsed 8 ledgers into 1 and took its close from 12 days to 5.
"When I started in February, we were at an inflection point with AI. Systems built in the last century cannot integrate it the same way, because it was never organic to how they were built."
Sebastian Sandorff Jacobsen, Head of Finance and Ops, All Gravy
The short version
Ask what runs while nobody is watching, whether it has a name, who checks it, whether an auditor has already looked at it, whether your own team can write the next one, and whether the ledger is whole enough for any of it to work. 6 questions, and most of the category answers 2 of them.
See the agents that ship with Light, the Custom Agents any admin can write, or book a demo.