Agents that do the work, not chatbots that describe it.
Most finance teams are not short of software. They are short of the hours it takes to feed it. We build AI agents that sit between your documents, your bank and your ledger, and carry out the repetitive work end to end, inside controls you can inspect.
Two ways we work
We run the function for you
You send us the documents and the access. Our agents process them, our accountant reviews the exceptions, and you get clean books and answers. You do not manage any of it. This is what our bookkeeping and payables clients buy.
We build it inside your team
You already have a finance team and you want them to stop doing the manual parts. We map the process, build the agents on infrastructure you own, document the controls, and train your people to run it. You keep the asset.
What a finance agent actually does
Three processes below. These are real workflows, not concepts. Each one shows every step, including the points where a human has to make the call.
1. Supplier invoice to posted bill
A supplier invoice arrives. Historically someone opens it, reads it, finds the vendor, keys in the lines, picks an expense account and posts it. That is four to six minutes a document, and it is where most coding errors are born.
Invoice received Agent
The document is picked up from a monitored inbox, an upload, or a vendor portal.
Read and extracted Agent
The agent reads the actual document, not a template. It pulls the vendor, invoice number, date, line items, tax and total, and returns a confidence score on what it found.
Vendor matched and duplicate checked Agent
It looks the supplier up in your ledger and searches for an existing bill with the same number and amount. Paying the same invoice twice is one of the most common and most expensive errors in payables.
Coded against your chart of accounts Agent
Each line is mapped to a real account from your ledger using your own rules, not a generic guess. Tax is treated correctly for your jurisdiction, whether it is recoverable or a cost.
Checked before anything is written Agent
The lines must add back to the invoice total. If they do not, or confidence is below threshold, or the amount is over your approval limit, the agent stops and refuses to post. It never guesses its way into your ledger.
Exception reviewed Human
Anything held goes to a queue with the reason attached. A person decides, and that decision teaches the rules for next time.
Posted and logged Agent
The bill is created in the accounting system and written to an audit log: what was decided, by whom or by which agent, with what confidence, and why.
2. Bank transactions to coded ledger
The other end of the same problem. Hundreds of transactions land in the ledger every month, half of them sitting in an uncategorised holding account, and someone has to work through them line by line before anything can be reported.
Chart of accounts read first Agent
The agent pulls your live chart of accounts before it does anything else, so it can only ever suggest accounts that actually exist in your books.
Transactions pulled and triaged Agent
It fetches the period's transactions and separates the ones sitting uncategorised or in a holding account from the ones already coded.
Categorised in batches Agent
Payee, description, amount and pattern are assessed together. Each transaction gets a suggested account, a confidence score and a written reason you can read.
Your rules override the model Agent
Known vendors are pinned to fixed accounts. Suggestions outside your chart of accounts are rejected outright. Sensitive categories such as director expenses or anything with tax consequences are forced to a human regardless of how confident the model was.
Split into confident and queried Human
Confident items are ready to apply. Everything else goes to a review queue with the reason. The queue is the product: it is where the accountant's time is worth paying for.
Every decision logged Agent
Transaction, old account, suggested account, confidence, reasoning and any flags, written to an audit trail you can hand to a reviewer or an auditor.
3. Month-end close and reporting pack
The processes above feed this one. Once the inputs are clean and continuous, close stops being a two-week scramble and becomes a review.
Checks run automatically Agent
Reconciliations, unusual movements against prior months, missing documents, balances that should be zero and are not. Compiled into an exception list before anyone opens a file.
Accountant works the exceptions Human
Judgement, accruals, anything that needs a conversation with you. This is the part that should never be automated, and the reason our clients get a named professional.
Pack drafted in plain English Agent
Numbers, movements and the commentary explaining what changed and what it means, written the way a briefing is written rather than the way a trial balance is printed.
Reviewed and signed Human
Nothing reaches you, or any authority, without a qualified person reading it and putting their name to it. That is the rule the whole firm is built on.
The controls, since you will ask
An agent with access to your ledger is a serious thing. Here is how ours are constrained.
It cannot invent an account
Agents read your live chart of accounts and can only use what is in it. A suggestion that does not match a real account is rejected before it goes anywhere near your books.
It must balance
Postings are checked against the source document before they are written. If the lines do not agree to the total, the agent throws the transaction out rather than posting something wrong.
It knows when it is unsure
Every decision carries a confidence score against a threshold set per client. Below it, the item is held. Certain categories are escalated to a human no matter how confident the agent was.
It cannot pay anyone
Our agents record and code transactions. They do not move money. Payment stays a deliberate human act, with your bank and your authorisation.
Everything is written down
Each action lands in a timestamped audit log: what happened, which agent did it, how confident it was, and what a reviewer decided. You can read it whenever you like.
Rules are yours, not ours
Thresholds, category mappings, vendor rules and escalation triggers are configured per client in a file you can review and change. No hidden logic.
How a build runs
Watch the process as it really is
We sit with whoever does the work today and map it step by step, including the workarounds nobody documented. Most of the value shows up here, before anything is built.
Agree the rules and the limits
What the agent may decide alone, what it must escalate, what it may never touch. Written down and signed off before a line of it is built.
Build and test against real documents
Built and run against your actual invoices and transactions in a sandbox, not sample data. We show you what it got right and, more usefully, what it held back and why.
Run alongside, then hand over
It runs in parallel with the existing process until the exception rate is boring. Then you get the documentation, the audit trail and the training, and it becomes business as usual.
Straight answers
In most cases yes. We work with the major cloud accounting platforms and with a good number of the specialist and regional ones. The agents connect through the software's own interface, the same way any other approved application would. Where a system has no way in, we say so at the mapping stage rather than after you have paid for a build.
Then that is the first job and it is a fair one. Plenty of growing businesses run on spreadsheets and a bank statement for longer than they should. We set the system up properly, build the chart of accounts around how you actually operate, get the opening position right, and only then automate on top of it.
It replaces the part of their week they hate. Someone still has to own the numbers, work the exceptions, handle the judgement calls and talk to the business. What goes away is keying, matching, chasing and re-checking. In our experience the team gets smaller in workload rather than in headcount, and the work gets considerably more interesting.
It gets caught, because that is what the controls are for. The design assumes the model will be wrong sometimes. The balance check, the confidence threshold, the chart of accounts validation and the human review queue exist precisely so that being wrong results in a held item rather than a bad entry. When something does slip through, it is logged, we can see exactly which step let it past, and we fix that step.
Your documents and records are processed in systems we control, used only to do your work, and never used to train anybody's model. For a build inside your own team, it runs on infrastructure you own and the data never leaves it. We will put the specifics in writing before you share anything.
A build is priced on the process, not on a licence per seat, because the work is in the mapping and the controls rather than in the software. A single well-scoped process is a matter of weeks. We will give you a fixed number after the mapping conversation, and that conversation is free.
Tell us the process that eats your week.
We will map it with you and tell you honestly whether it is worth automating. Some are not, and we will say so.
Book a free process review