This is the sixth agent in the Season 2 run of my AI Agent Series, where I break down every agent I actually run inside Purposeful Profits and across the brands I work with. This one is the least glamorous and the fastest to pay for itself, because it deals with money that already belongs to you.
Picture a Monday where you already know your cash position. Not roughly. Exactly. You know which invoices landed over the weekend, which client is now nine days late, which payout came in three hundred pounds lighter than the order value said it would, and which subscription quietly stopped billing in June. You know all of that before you open your laptop, because something read every transaction while you slept and wrote it down in four lines.
What this used to cost, and why founders keep putting it off
Reconciliation is the task everyone agrees is important and nobody wants to own. Most founders running a brand under a few million do it themselves, and the honest number is eight to twelve hours a month on a simple setup, climbing to fifteen or twenty once you have multiple currencies, a couple of processors and a subscription product. A brand doing five thousand orders a month is dealing with thousands of line items across Shopify, Stripe, PayPal, refunds and bank feeds. Finance teams at larger brands routinely lose six to twelve days a month just getting to a state clean enough to begin the close.
So it slips. It becomes a quarterly job, then an annual panic before the accountant's deadline. And the cost of that delay is not admin time, it is money. The average small business is owed around twenty two thousand pounds in overdue invoices at any moment, UK small businesses carry roughly twenty six billion pounds of unpaid invoices between them, and late payment is pushing an estimated fourteen thousand businesses under every year. None of that gets chased if nobody has looked.
Reconciliation is not accounting hygiene. It is the only process that tells you whether the money you earned actually arrived. Skip it for a quarter and you are running the business on a number you have never verified.
What the agent actually does
It has read access to the bank accounts and the payment processors, and it has the list of invoices and orders that have been issued. That is the whole input. Here is what it runs, every day, unattended.
01
Pull every transaction across every account
Multiple accounts, multiple currencies. In my own case that is a GBP account and a CAD account plus the processors, because clients pay from different countries and the money does not politely land in one place. The agent pulls the full transaction list for the window, normalises the currencies at the rate on the day the money landed rather than today's rate, and stores the raw records untouched. Nothing is interpreted at this stage. The job is completeness.
02
Match payments against issued invoices
The naive version matches on amount and reference. The real version has to handle the mess: a reference that dropped the invoice prefix, four invoices settled in a single transfer with the numbers jammed together, an international payment that arrives eleven pounds light after intermediary fees, and a client whose bank sends nothing but their own company name. The matcher works through amount, reference fragments, payer identity and date proximity, and it scores its own confidence rather than pretending every match is certain.
03
Classify what did not match
Everything unmatched gets sorted rather than dumped in a pile. Invoice issued and nothing received. Partial payment received against a larger invoice. Money received with no invoice attached, which is usually a new client or a refund reversal. Outgoing transactions that look like duplicated payments. The classification is the useful part, because each category has a completely different next action and mixing them is why most reconciliation spreadsheets get abandoned.
04
Age the receivables and rank them
Every unpaid invoice gets an age in days from the due date, not from the issue date, and the list is ranked by amount at risk rather than alphabetically or by date. Anything past thirty days gets escalated in the report. Anything past sixty gets called out specifically, because that is the point where the probability of collection starts falling off a cliff and a polite reminder stops being enough.
05
Reconcile the payout, not the deposit
This is the ecommerce-specific piece. A Shopify, Amazon or TikTok Shop deposit is a net figure with processing fees, refunds, chargebacks, shipping labels and ad deductions already taken out. Booked as revenue, it is wrong in about nine directions simultaneously. The agent breaks each payout back into its component parts and reconciles gross sales against the settlement, which is where the two to four percent of quietly disappearing transaction volume shows up.
06
Write the summary a human will actually read
The output is not a dashboard. It is a short written report: closing balance per account, what came in and from whom, what is outstanding and how late, what did not reconcile and why, and a running month-to-date income figure. Then a numbered action list. If nothing needs a decision, the report says so in one line and you move on with your morning.
The context layer is what stops it guessing
Any script can compare two numbers. What makes this behave like a finance person rather than a spreadsheet formula is the memory it carries in. It knows the client list and which entity name each brand pays under, which is rarely the name on the contract. It knows that one client always pays in two instalments, so a half payment is expected behaviour and not an exception. It knows the retainer amounts, so a transfer that is three hundred pounds short is flagged as an FX shortfall rather than a partial payment dispute. It knows which invoices were issued in CAD and which in GBP, and it knows the bank details that belong to us so an outgoing transfer between our own accounts never gets counted as revenue.
It also carries the history of how each relationship behaves. A client who has paid on day twenty nine for eleven months running is not late on day twenty five, and flagging them wastes attention and damages the relationship if it triggers a chase email. A client who has never paid before the second reminder is a different risk profile and should be surfaced earlier. That memory is the difference between an agent that reduces your workload and one that generates a new queue of false alarms for you to triage.
What the output actually looks like
Four blocks, arriving as a message rather than a login. Balances first, one line per account. Then movement: what arrived since the last run, matched to a name and an invoice number. Then exceptions, which is the only section that ever needs me. Then the month-to-date income figure against the same point last month.
A real exceptions block from a recent run read something like this. One payment of five figures received with a reference listing four invoice numbers stripped of their prefixes, which the matcher resolved by amount and flagged at medium confidence for a human to confirm. Two invoices sitting unpaid past thirty days on a client who has historically paid on time, which is a signal worth a phone call rather than an email. One invoice marked as sent that had never actually left drafts, which is the single most common and most expensive error in any small business finance process.
That last one is the point. The money was earned, the work was delivered, and the invoice was sitting in a drafts folder because a human got interrupted. No amount of chasing finds that. Only a system that compares what should exist against what does.
Where it still needs a human
An agent described as flawless is an agent nobody should trust, so here are the limits. It has read access only and it stays that way. It does not move money, it does not send payment reminders on its own, and it does not mark an invoice as settled without a human confirming a medium-confidence match. That is a deliberate constraint rather than a technical one. The blast radius of a reconciliation error that also fires an email at a client is much larger than the time saved by automating the send.
It also breaks in predictable ways. Bank API tokens expire and the run fails silently unless you build an alert for the failure itself, which took me one embarrassing week to learn. Genuinely ambiguous references still need a person, because guessing wrong and telling a client they underpaid is worse than asking. And none of this is accounting. The agent produces a matched ledger and an exceptions list. Your accountant still owns the VAT treatment, the accruals and the filings, and they do a better job of it when they receive clean data instead of a bank export.
Inside the system
How we build this for brands
For a brand rather than an agency, the same agent points at Shopify payouts, processor settlements and marketplace deposits instead of client invoices, and it feeds the profit and cash-flow dashboards we build from live Shopify and ad data. Once the money side is verified rather than estimated, the reporting agent that surfaces leakage or risk weekly is finally working from a real contribution margin instead of a gross figure with the fees still hiding inside it.
That accuracy then travels. The lifecycle flows deployed in Klaviyo can trigger on a genuinely failed payment rather than a guess, so the dunning sequence stops emailing people whose card actually went through. And the ad account gets budget decisions made against verified net revenue rather than a top-line number that was two to four percent optimistic all quarter. Part of this runs live for portfolio brands today; the full system is what we deploy when we take a brand on.
Build Your Agent Team
Get your own AI agent team working on your brand
Caner can help you build your own AI agent teams and unlock growth for your brand. This is not a template or a SaaS tool. It is a custom system designed around your data, your workflows, and your growth targets.
Book A DemoFrequently asked questions
What is a revenue reconciliation agent and what does it do?
It is an automated system that reads every transaction across your bank accounts and payment processors, matches incoming money against the invoices and orders you have issued, and reports what is paid, part paid, unpaid and unexplained. It runs on a schedule rather than on demand, so the picture is current every morning instead of once a month when someone finally opens the spreadsheet. The output is a short written summary and an exceptions list, not a dashboard you have to interpret.
Can I build a reconciliation agent myself?
Partly. Pulling transactions from a bank API and comparing amounts to an invoice list is a weekend of work if you can code. The hard part is the matching logic. Real payments arrive with the invoice prefix stripped, four invoices settled in one transfer, an FX conversion that lands three pounds short, or a reference that is just the client's surname. Encoding those cases, and knowing which mismatches to escalate rather than guess at, is what takes months rather than a weekend. The other half is the memory layer: which client pays in instalments, which entity name each brand pays under, and which transfers are just your own money moving between your own accounts.
How long does it take to set up for an ecommerce brand?
About a week of real work. Two days to get read access to the bank and processor accounts and confirm the token refresh works unattended, two days to map your invoice or order data into a shape the agent can match against, and a day or two of supervised runs where a human checks every match before the agent is trusted to run alone. After that it runs daily and only asks for attention when something does not reconcile.
Does this replace my accountant or bookkeeper?
No, and it is not trying to. The agent handles the mechanical part: reading transactions, matching them, and flagging what does not fit. Your accountant still owns the treatment of those numbers, the VAT position, the accruals and the filings. What changes is that they receive a clean, already matched ledger with a short exceptions list instead of a bank export and a shrug, which usually means they spend their hours on advice rather than data entry.
Why do bank deposits never match reported sales in ecommerce?
Because a payout is a net figure, not a sale. Every deposit from Shopify, Amazon or TikTok Shop already has processing fees, refunds, chargebacks, shipping labels, advertising deductions and tax withholding netted out of it. If you book the deposit as revenue, your books are wrong in several directions at once and the reconciliation can never close cleanly. Undetected settlement discrepancies commonly run at two to four percent of transaction volume when nobody is checking, which on a brand doing two million a year is a full-time salary disappearing quietly.
How much money do brands typically find when they reconcile properly?
It varies, but it is rarely zero. The average small business is owed around twenty two thousand pounds in overdue invoices at any one time, and UK small businesses collectively carry roughly twenty six billion pounds in unpaid invoices. The first proper reconciliation almost always surfaces something: a payment received but never marked off, an invoice that never left drafts, a refund processed twice, or a subscription that stopped billing months ago and nobody noticed because revenue was growing.
About the author
Caner Veli built Liquiproof to global distribution across 3,000+ retailers, then exited. He now runs Purposeful Profits using a combination of operator strategy and AI-powered systems he has built and uses daily, having 10x'd monthly revenue in his own business in the last 90 days.