What an Accounting AI Agent Costs — and the Volume Below Which You Shouldn't Buy One

Ask a vendor what an accounting AI agent costs and you get a calendar link. I understand the reflex — the honest answer really does depend on volume. But the effect is that nobody can compare anything, and offices end up choosing on demo quality instead of arithmetic.

So here is my price. Then the more useful part: the sum that tells you whether to buy it at all. There is a number in it that costs me sales. I would rather publish it than argue about it after the invoice has gone out.

The price, plainly

A pilot starts at €1,900. That buys one document flow automated end to end — either purchase invoices arriving from KSeF and email, or bank-statement reconciliation. It runs as drafts inside your system, and a person approves every entry. The integration is included (KSeF, enova365, Optima, Symfonia, wFirma or your bank), along with the exception queue, a full audit log and a data-processing agreement. Two to three weeks to go live.

After go-live there is a monthly subscription: hosting, maintenance, model costs. I quote it together with the pilot because it tracks volume. For everything below I will assume €149 a month, which is what the AI ordering assistant on the B2B portal costs. Close enough for planning, and stated openly so you can drop your own quote in instead.

Two more numbers for scale, from the same price list: an AI readiness audit starts at €2,500, a full automation build at €5,000. The pilot is deliberately the cheapest thing I sell. It should pay for itself before you commit to anything bigger.

The arithmetic that actually decides

Price is the easy part. The real question is whether the hours you get back are worth more than the subscription. Every month. Forever — not just in year one.

Four assumptions. All arguable, all mine:

• Three minutes of human handling per document, after KSeF. Not the typing; that is gone. This is coding the document, matching the payment, and deciding what to do when it does not match.
• 55 PLN an hour, fully loaded. Salary plus employer contributions, divided by the hours actually worked — not the gross figure on the contract.
• 70% straight-through after calibration. Seven documents in ten booked without anyone touching them. Month one looks worse. A clean, repetitive portfolio does better.
• 4.30 PLN to the euro. So the pilot is about 8,170 PLN and the subscription about 640 PLN a month.

Now run a company through it. Four hundred cost documents a month is 20 hours of handling. Automate 70% and you save 14 hours, worth 770 PLN. Take off the 640 PLN subscription and you are left with 130 PLN a month. Against a pilot of 8,170 PLN, that is a payback measured in decades.

So at 400 documents a month: do not buy this. Not from me, not from anyone.

Run the same sum upward and it turns:

Documents/month	Manual hours	Hours saved	Net monthly gain	Payback on €1,900
400	20	14	130 PLN	effectively never
510	25.5	17.9	342 PLN	~24 months
700	35	24.5	707 PLN	~12 months
2,400	120	84	3,980 PLN	~2 months

About 700 documents a month is the line. Below it, you are buying a subscription that eats most of what it saves. Above it, the maths stops being close.

Why the threshold moves for an accounting office

Look at the last row again. That is not a large company. That is an office with forty clients averaging sixty documents each.

Which is the whole argument for accounting offices as the primary buyer. One agent, one integration, one calibration — spread across a portfolio sitting on a shared ERP. A single client stays stubbornly below the line. The twelfth client puts the office past it. Same software, same invoice from me, and a payback that moves from theoretical to about two months.

It also explains something buyers find backwards. The client who is hardest to automate is often the one who justifies the project, because they add volume to a threshold that only the office as a whole has to clear.

What pushes the price up

Four things, in the order they turn up in quotes:

A second flow. Purchase invoices and bank statements are separate integrations with separate ways of breaking. Adding statement reconciliation after invoice intake costs far less than doing both at once — and it gives the first one a chance to prove itself.

An ERP without a usable API. enova365 has a documented WebAPI, so the work is predictable. Older on-premise installations sometimes mean going in at database level. That costs more and ages badly.

A non-standard chart of accounts. Every deviation from the ERP's default coding is calibration work the agent cannot infer. This is the item that most often turns a fixed price into a range.

Exception handling nobody scoped. The queue where uncertain documents wait for a human is the difference between a demo and a system. It is also where the ongoing hours live. A quote that never mentions it is not cheap — it is incomplete.

What the subscription actually buys

The recurring line is the one buyers question hardest, and fairly — the software is built, so what am I charging for every month?

Three things, and none of them is profit on idle code.

Model costs, which scale with your documents. Every document classified is inference someone pays for. This is the part of the bill that genuinely moves with volume, and it is why a flat subscription quoted before anyone has counted your documents is a guess wearing a price tag.

Keeping the integration alive. KSeF's schema is not frozen, ERP vendors ship updates on their own calendar, and a bank changes its statement format without asking anyone. Each of those breaks an integration quietly — the agent keeps running and simply stops catching things. Somebody has to notice within days rather than at the quarter's close.

Calibration drift. Your document mix moves: a new supplier, a new cost category, a client who switches to split payment. The straight-through rate slides from 70% toward 50% and nobody sees it, because fewer drafts arriving still looks like everything is fine.

That last one is the question worth putting to any vendor, including me. Ask how you will see the straight-through rate — a number, monthly, per flow. If nobody can produce it, the subscription is buying hosting and optimism, and you will find out which at the worst possible moment.

If you are under the threshold

Most single companies are. The honest answer is: do not commission a custom agent yet.

Do this instead.

Turn on what your ERP already has. Most Polish systems shipped KSeF intake features that half their users have never enabled.

Pay per check where per check is genuinely cheap. KsięgoAI verifies a contractor by NIP against the VAT white list and VIES for 39 PLN per 30 checks, first 50 free — a rounding error next to one bad payment over the 15,000 PLN joint-liability threshold.

Then audit your hours. One week, every accountant tagging where the time goes. If it turns out you are at 900 documents a month rather than 400, that changes the answer — and it is the cheapest piece of analysis in this article.

Come back when the volume is there. It will be a better project, and you will know exactly why you are buying it.

The uncomfortable summary

An accounting AI agent is not expensive. €1,900 to automate a document flow is less than a month of the salary it partly replaces. But cheap and worthwhile are two different questions, and the subscription keeps asking the second one. Below roughly 700 documents a month my recommendation is to wait — and a vendor who tells you otherwise has not run the arithmetic in front of you.

Want this table with your own numbers in it? Start with a free scoping call: thirty minutes on your document volume and your ERP, and you will know whether you are above the line. Including if the answer is that you are not.