blueoceanfinance DE 45 minutes on your numbers
Measured 7 September 2026, in live operation

What the AI bookkeeping delivers — and what it costs.

BlueOcean Finance is an AI for tax firms. It collects your clients’ receipts, captures them, imports the bank statement, matches payment to receipt and prepares the month for closing. Whatever it cannot decide with confidence, it puts in front of you.

No pilot, no demo data: 33 clients and three firms work with it. What has come of that so far is below — with the query behind it, in case you want to check the arithmetic.

18,600receipts read and processed
135,006bank statement lines read
99.2 %of allocations made by the machine
€140.14total AI cost for all of it

476 of the allocations are not 1:1 — and the machine solved 472 of them. Precisely the cases that cost manual work: one transfer against three invoices, five prepayments against one monthly invoice, interwoven chains.

Two qualifications we would rather add ourselves: the 99.2 % refer to the allocations, not to all bank lines — 13,794 bank transactions face 4,342 allocations, and what needs no receipt (fees, salaries, owner draws) is not separated out in that figure. And they do not mean nothing is left for your people: in August 86 clarification cases stood open that nobody worked on. What the machine cannot do stays put until a person decides.

Inside the firm portal

What your bookkeeper sees in the morning.

Today
Today — Exactly three things: what blocks a close, what to check, what sits with the client. Technical failures are listed separately — that is our work, not yours.
Reconciliation
Reconciliation — Where money sits without a receipt, across month boundaries. One transfer can settle three invoices, five prepayments one monthly invoice — the system shows the chain, not just the pair.
Ready to close
Ready to close — Every client on one board, per month. Beside each one: what is missing, and the button that resolves it.
Billing
Billing — What the AI cost per client, with the receipt count beside it that produces the figure.
One real month

August 2026, across all 33 clients.

Not the total since launch, but a single completed month — so you can see what actually stands there at month end.

672allocated payments worth €627,155.50 across 20 clients
10 minfrom queued to receipt read, median across 16,384 jobs
26 of 33clients stood at "ready" or "almost" — 1 ready, 22 almost, 3 already closed

Whose move it was next

the firm12
the system9
the client0
nobody — nothing open12

Exactly one of the three is responsible, never "somebody". That nothing sat with the client means the system had already done the chasing.

What a close was waiting on

receipts under review10
payment without receipt10
receipts awaiting release9
incoming without invoice9
firm's judgement pending6
file unreadable / extraction failed4

Number of clients affected. None of these is a surprise on the 10th — each one sits on the board from day one.

How many payments stood without a receipt at month end is in the section below — separated by what genuinely needs a receipt and what never gets one. Until 9 September that split was only computed at the moment of the call; since then a nightly run records it. That is why the section below shows a series that grows by itself, not a snapshot.

As of August 2026, measured on 9 September

This is how it looks here. Unvarnished.

Every expense of the month sits in exactly one of three buckets — 1,804 expenses across all 33 clients. Not an average of the good months, not a target.

Small client265 expenses in August
99
59
107
Mid-size client562 expenses in August
188
106
268
Large client977 expenses in August
399
190
388
All together1,804 expenses, 33 clients
686
355
763
receipt present and matched never needs a receipt — fees, salaries, tax, rent, internal transfers receipt missing

About half is still missing. 763 out of 1,449 expenses that need a receipt are still waiting for one in August. That is the starting point, not the result — anyone promising you 90 per cent has either measured something else or nothing at all.

And the month before was not worse: July 879 out of 1,751 (50.2 %), August 52.7 %. We can forecast where this is heading — but this is today’s state. And when a number looks better, we check first whether something improved — or only the denominator did.

So there is no doubt about what you are seeing: these are our own books — the first reference project, and it is at the beginning. The improvement comes over the next days, weeks and months, cycle by cycle. And it does not come from the model alone: the system gets the most out of it when a bookkeeper teaches it what she knows about the client. We are not dressing anything up here.

The bucket had the wrong name

A client thought the figure was too high. Re-measured across July and August: all 1,661 lines in the “receipt missing” bucket carried the same reason — “no rule found”. Not one said “a receipt belongs here”.

Held against the document store: for 1,170 of the 1,661 lines (70.4 %) a supplier document is already in the system — it is simply not matched. Only 491 have nothing at all.

So there is no misunderstanding: that is 70 % of the lines, not 70 % of finished bookings — several lines often reach for the same piece of paper. And no receipt turned up. The bucket shrinks because the classification is right, not because paper was found.

The learning is no longer a claim

The system applied a rule it had learned from the bookkeeper 401 times — and not one of them had to be taken back by a person. For 23 of 33 clients this already happens on its own.

Plus 35 learned supplier names, applied 105 times. 19 of the 35 rules now fire without asking.

This knowledge does not come from the model but from the people who know the client. That is why it gets better month by month — and why a human stays the authority on the professional questions.

Not every payment needs a receipt — but every payment needs an explanation.

A salary payment has no invoice — it has a payroll statement. Rent has a lease. Until now our system did recognise such payments — and then asked for nothing at all. An auditor sees a payment with no paper whatsoever. Since 9 September it says which document is needed instead, and requests it from the client.

  • Payroll statementmonthly — covers every employee
  • Lease agreementonce — explains every month
  • Contribution statementper fund and month
  • Tax assessmentper payment
  • Loan agreementonce
  • Shareholder resolutiononce — on distribution

Bundled, otherwise it would be hundreds of requests: one payroll statement covers every employee of a month, one lease every month with the same landlord. Across the estate that became 111 requests at 22 clients, together explaining 206 bank lines. Bank fees, cashback, internal transfers and our own fee deliberately owe nothing — a reason that is not on this list produces no request.

And what we deliberately did not build

The obvious rule “exchange rate = currency swap” would have been an expensive mistake. The phrase appears on 637 bank lines — but only 27 really are a swap. The other 610 are card payments with an invoice. Had the phrase alone made the rule, 610 expenses needing a receipt would have vanished silently.

That is the most dangerous kind of error, because it hides a defect instead of reporting it. Six of those lines are now a counter-check in the test suite — it fails the moment anyone puts the convenient rule back in.

What it costs

€3.41 per client and month.

€2.50base price per clientfixed, regardless of volume
€0.91usage per client, avg.by actual document, statement and matching volume

€3.41 per client per month, combined — measured from 33 clients, July and August 2026, extrapolated to 400 clients with a factor of 12.1 (€1,363/month). No second, invented per-client fee: an earlier version of this page cited €17.25 per client, which never existed as a measured figure.

What stands behind the price: our own AI cost runs €0.32 for an average client, €0.85 for a large one (300 receipts, 8 statements, 500 bank lines) — formula: 0.14 cents per receipt, 1.61 cents per bank statement, 0.06 cents per bank line, measured over 44 client-months in July and August 2026. An earlier version said €0.11 to €0.77 here; it counted only the receipt reading and left out statement and reconciliation.

Profile Receipts Statements Bank lines AI cost/month
Average — measured 59 4.4 110 €0.32
Large — worked example 300 8 500 €0.85
Very large — worked example 500 16 900 €1.50
Extremely large — worked example 900 25 1,400 €2.50

Only the first row is measured. The other three are the same formula, extrapolated further, with round input numbers — not a case we have actually seen with you. A really complex client (several accounts, many statements) can end up above the average usage share in the price; the base price (€2.50) stays the same either way.

So you can trust the rest

What the system does not do.

It decides no technical questions.VAT rate, use, classification stay with the firm. The system proposes and learns from your decision, per client.
It learns nothing across client boundaries.A supplier name you confirm for one client applies there only. A dedicated check verifies this before every release.
It closes no month quietly.If something is missing it says what is missing and who must supply it: the system, the bookkeeper or the client. Exactly one of the three — never "somebody".
It invents no payment.Where an exchange rate or an amount cannot be evidenced, it asks instead of estimating.

And one figure you will not find here: hours saved. We have no measurement for that, so we do not claim it. What we have measured are receipts, cost and the share of machine-made allocations.

The next step

Run one month alongside.

We set up your folders, re-run a month you have already closed, and show you what the system would have found on its own. You see the result before you decide.

Book a call — 45 minutes on your numbers

€150 for 45 minutes. Credited in full once it turns into a project or a feasibility study.

Prefer to write? info@blueoceanfinance.ai