AI agent for accounting firms it prepares, your accountant approves
A firm serves dozens of companies at once and every one of them delivers documents differently: a scan to a shared inbox, a photo on a chat app, a ring binder once a month, a file exported from their own software, a structured e-invoice. We reduce that to one intake, attach each document to a client and a period, read the data and run the formal checks. The entry into the books is approved by your accountant.
- 5 channelsthat is how many routes documents take into a typical firm today: email, a chat app, paper, a file from the client's own software and a structured e-invoice. We reduce them to one queue
- 14 daysfrom the end of the workshop to the first automation running, tested on your own documents from a period that is already closed
- EUR 3,500net per week of workshop, a published price. We quote the rollout only after it, once it is clear how many systems are in scope
The problem
Why is an accounting firm always short of people at month end?
The month does not spread evenly in an accounting practice. For the first few days you wait for documents that do not exist yet, and then everything has to be posted in a handful of days, because the filing dates are fixed and they do not move because a client brought in a ring binder at the last minute. Once a year the same work comes back stacked up, at the annual close and the statutory accounts.
The second problem is the daily one. Documents arrive by five routes at once, so an accountant first collects them, then sorts them, then retypes them field by field into the ledger. In between she answers a question the same client has already asked four times this year, and writes an email chasing an invoice that has been missing for three weeks.
The third thing is new and it does not depend on you. Countries across the EU are phasing in mandatory structured e-invoicing on their own timetables, and the dates differ from one market to the next. For a firm that means one more format on the intake side, plus a stretch of time in which some of your clients still work the old way and some of them already do not.
- A purchase invoice retyped by hand, field by field, one more time
- The same emails about missing documents, written client by client
- A document sits in one person's inbox and nobody else knows it arrived
- The same client question comes back to the same accountant for the hundredth time
- The close status gets collected by hand, in a spreadsheet, in the evening
Why now
What changed, so that this can be done in weeks?
The first change is the format. Some invoices arrive in an identical shape for the first time, because a structured e-invoice enforces one set of fields instead of an arbitrary PDF layout. The rest stays as it was: a scan, a phone photo, paper, a file exported from the client's own software. You still have to unify the intake, there are simply fewer variants now and some of them have a fixed form.
The second change is the reading. A model reads a scan or a photo of an invoice well enough that the work stops being retyping and starts being review. A rule checks the line totals, the tax rates, duplicates, the period and the VAT number against VIES and the national register, and your accountant gets a finished entry to approve with the document shown next to it. The decision stays with a person and that does not change.
The third thing is technical and we say it plainly, because it gets promised too casually. Comarch Optima, Subiekt and Symfonia are installed locally and keep their data in a Microsoft SQL Server database on your own server. They have no public address, so there is no webhook and no request from the cloud. You go in through the vendor's own layer or by reading the database, using a connector that runs inside your network. It works and we have been building it that way for years, it simply looks different from wiring up a cloud tool such as Xero, QuickBooks, sevDesk or Fakturownia, where the vendor exposes an API over the internet.
What changes
What does the team get after the first month?
One intake instead of five channels
Email, a client portal, a chat app, a scan of paper and a structured e-invoice all land in one queue. Your accountant stops hunting for documents across systems and starts the day with a list that is complete.
1 intake queue instead of 5 places somebody has to look into
The document knows whose it is and which period it belongs to
Assignment runs on the VAT number taken from the document, the sender address and the sale date. Whatever a rule cannot settle goes into a manual queue with the reason stated, rather than into another client's folder.
3 criteria for assignment instead of a file name and the memory of whoever sorts the post
Invoice data read and checked before posting
Number, dates, amounts, tax rates, counterparty and line items come out of the document with nothing retyped. The formal checks run before the entry is handed over, so your accountant looks at the document with the result of those checks already attached.
Line totals, duplicates, period and VAT number checked before anything reaches the books
Chasers for missing paperwork go out without anyone writing them
The system compares what should have arrived for a given period with what actually did, and contacts the client with the list of gaps. Your accountant steps in only when the client goes quiet.
A list of gaps per client sent on a fixed rhythm, with nobody checking by hand
Client questions answered from your own procedures
An AI agent handles the recurring questions about deadlines, required documents and the status of a filing, using your procedures and your data. A question that needs an accountant's judgement is passed on with the context attached.
An answer in 5 to 10 seconds, from your own base, not from the model's general knowledge
Close status visible without anyone compiling it
A view with every client, the stage, the gaps and the date of the last movement refreshes itself. The owner of the firm sees where the queue is forming before it turns into a night shift.
A view refreshed every day instead of a spreadsheet stitched together in the evening
Proof
Do you work this way yourselves?
Yes, and to be straight about it: our own measurement does not come from an accounting firm, it comes from our founders' agency. That is a services business that bills every client separately and where repetitive delivery work ate hours the same way it does in your practice. The AI agent has been running in production there since March 2026. Below is the number, together with what it does NOT prove.
Built for ourselves. These are not client deployments but implementations inside the founders' own companies. We say so plainly, because proof you cannot verify is not proof.
JustBottheir agentJustIdea Agency, a performance marketing agency of around 30 people
- 23
- people from the company work with the agent
- 38
- systems wired in permanently
- 112
- processes run without a human
- 475
- thousand lines of the agent’s code
An agent wired into Google Ads, Meta, the CRM, the task system, the SEO tool, Analytics, Search Console and the data warehouse. First commit 1 March 2026, today 23 people use it.
8 of 38 systems it works in
Google Ads
Meta Ads- HubSpot
- ActiveCollab
- Ahrefs
GA4
Search Console
BigQuery
- Before
- In the twelve months before launch the agency ran at a margin of minus 4.42 percent. Across a twenty-month window there was not a single run of three profitable months.
- After
- May, June and July 2026 are three profitable months in a row, the first such run in the whole window. Revenue barely moved, so the change sits on the cost of delivery, not on sales.
Margin up 5.55 percentage points year on year, with revenue up 1.5 percent
And in your firm
What if it does not work in our practice?
Four sentences we hear most often in accounting firms, and what we say back to them. If yours is a different one, say it on the consultation.
Our data is our clients' data, we are not sending it to a model.
We agree the scope in the workshop and write it down: the model sees specific sources and specific fields needed for the task, not the whole firm database. Processing runs at a provider you hold a data processing agreement with, in the region you name, under <a href="https://eur-lex.europa.eu/eli/reg/2016/679/oj" rel="noopener">GDPR</a>. Decisions stay with a person, so the model never approves an entry and never files anything with an authority. If you prefer it, part of the flow runs on your own infrastructure.
We run Comarch Optima on a server in the office, that will not wire up to the cloud.
Not the way a cloud service does, and we are not promising you a webhook. Optima and Subiekt sit locally and keep their data in a Microsoft SQL Server database, so writing goes through the vendor's layer and reading goes against the database for reports. A connector running inside your network does the work and calls out for tasks itself, with no port opened from the internet. The first pass runs on a copy of the database.
We are not agreeing to let AI post entries to the books.
Neither are we, and that is how we build it. The automation takes over what a person does with a keyboard today: it collects the document, attaches it to a client and a period, reads the fields and runs the formal checks. The entry into the books is approved by your accountant, with the document, the extracted data and the list of checks side by side. Nothing reaches the ledger without that click.
Our clients will never stop sending photos on a chat app.
We do not assume they will, and we do not make the rollout conditional on it. The channel stays, what changes is what happens next: the photo lands in the same queue as a PDF from email and a structured e-invoice, and gets a client, a period and a reading. Changing client habits is a separate project you can run calmly, rather than a condition for starting.
- Accounting software on your own server
- Comarch Optima, Subiekt, Symfonia, Microsoft SQL Server, a connector inside your network
- Accounting software in the cloud
- Xero, QuickBooks, sevDesk, Fakturownia, iFirma
- Invoices and registers
- structured e-invoice formats, PEPPOL, VIES, national VAT and company registers
- Mail, scans and documents
- Google Workspace, Microsoft 365, a shared inbox, drives, office scanners
- Client communication
- Slack, Microsoft Teams, a client portal, HubSpot, Pipedrive
- Data and the automation layer
- Google Sheets, BigQuery, SQL databases, n8n, Make, custom API integrations
Where the data comes from and which route we take in
The systems we meet most often in accounting firms, together with the route in and what decides the scope of a rollout in each case. The split between software on your own server and software in the cloud matters more here than the names, because it decides whether a connector inside your network is needed at all.
| System | Route in | What to keep in mind |
|---|---|---|
| Comarch Optima | reading the Microsoft SQL Server database through a connector inside your network | the program sits on your own server with no public address, so a request from the cloud will not reach it |
| Subiekt | the Sfera layer for writing, the SQL database for reading | Sfera is a separate vendor licence, we check its scope in the workshop before quoting |
| Symfonia | the SQL database and exchange files | scope depends on the version and the modules you bought, we settle it before quoting |
| Xero, QuickBooks, sevDesk | the vendor API over the internet | works with no connector in your network, because the program runs on the vendor's side |
| Fakturownia, iFirma | the vendor API over the internet | same route as the other cloud tools, what differs is which fields the API exposes |
| Structured e-invoicing | the vendor module inside your accounting program | we do not replace sending and receiving with our own code, we deliver correct data into the document |
| Shared inbox and drive | Google Workspace or Microsoft 365 | one inbox for documents as the intake, the rest of the mail stays exactly as it is |
| VIES and national VAT registers | public services queried by VAT number | we store the result of the check next to the document, so you can go back to it later |
A first rollout usually covers one accounting program and one document intake. The rest goes in along the same track, after the first one has been measured.
How we work
Five steps from five inboxes to one queue
Process first, tools second. Every step ends with something you can judge before you decide on the next one. The first three change nothing in your books.
Process workshop
One week. We take several of your clients with different profiles and walk the path of a document from the moment the client sends it to the moment it is posted. We note how many times somebody touches the same piece of paper, where the queue forms and which client questions come back most often. The output is a document, not code.
The intake queue on a closed period
We build one intake and run it on documents from a period that is already closed, so nothing goes near the accounting program. We test assignment to a client and a period on the shapes you actually get: an invoice with no VAT number, a photo rotated the wrong way, a credit note, a document in a foreign currency.
Extraction and formal checks
We add field extraction and the checks to the queue: line totals, tax rates, duplicates, the period, the VAT number in the national register and in VIES. Your accountant judges the result next to the scan and marks what the model read wrong. Those corrections give us the threshold above which an entry goes for one-click approval and below which it goes to manual handling.
Chasers and close status
We switch on the reminders about missing documents sent to clients on a fixed rhythm, plus the close status view for the whole portfolio. Only here do we add the AI agent answering recurring questions from your procedures, because before this point there is nothing for it to answer from.
Production and measurement
We switch over with error monitoring and a queue where a rejected document sits with its reason, ready to retry. At 30 and at 90 days we look at the indicators agreed before the start: the share of documents that pass with no correction, the number of chasing emails written by hand, the time from a document arriving to it being posted.
Before and after
How the week changes for an accountant
Before
- A morning round of five places where clients dropped their documents
- A purchase invoice retyped field by field, dozens of times a day
- An evening run of emails about missing documents, client by client
- A question about a deadline or a required document interrupts posting several times a day
After
- One morning queue with documents already attached to a client and a period
- An entry handed over for approval together with the scan and the list of checks that ran
- The chasers went out on their own, she steps in only when a client stays silent
- A recurring question has an answer in 5 to 10 seconds, with nothing interrupted
The first effects show 14 days after the workshop, on one intake and one accounting program. The share of documents passing with no manual correction is judged at 30 and at 90 days, because only by then has a full monthly cycle gone through the system.
For the person signing this off
What the owner of the firm gets out of it
The first thing is capacity. Today the number of clients you can take on follows from how many hours the team can survive at the peak of the month. Once the automation takes over collecting documents, attaching them to a client and retyping the fields, the peak stops being a wall, and a conversation about a new client stops opening with the question of who is physically going to do the work.
The second thing is about risk and about people. Knowledge of who delivers what and where to look for it stops sitting in one person's head, so a holiday or a resignation no longer stops client service. A new hire walks into a written process rather than into five inboxes and an oral tradition. Professional responsibility stays exactly where it was, with the accountant who approves.
- The peak of the month stops setting the limit on how many clients you can serve
- An entry reaches the books only after a person approves it, every time
- The process is written down and visible, so an accountant's holiday does not stop a client
- The first three steps of the rollout change nothing in your books
What it costs
The workshop has a published price, the rollout is quoted after it
The process workshop costs EUR 3,500 net per week and runs from 1 to 7 weeks. In an accounting firm one or two weeks is usually enough, because you have to walk the path of a document across several clients with different profiles rather than one. We quote the rollout after the workshop, once it is clear how many intake channels remain, which accounting program you run and whether a connector inside your network is needed.
Accounting software licences, the running cost of scenarios in n8n or Make and the cost of model usage go straight to those providers at their own list prices, and a single query costs the provider roughly EUR 0.005 to 0.07. We add no margin and we do not handle those payments. There is more about the workshop itself on the process workshop page.
Process workshop
EUR 3,500 net
per week, and in an accounting firm usually one to two weeks
A self-contained piece of work. It ends with a map of the document flow and a 30 day plan you could carry out without us.
- Walking the path of a document across several clients with different profiles
- A map of the intake channels and the points where the queue forms
- The route into your accounting program, split by whether it runs locally or in the cloud
- The scope of data that leaves the firm at all, written down before any rollout
- Indicators to measure at 30 and at 90 days
- Most popular
Workshop and document flow
Quoted after the workshop
scope follows from the number of channels and from which accounting program you run
The option most firms choose. One intake, assignment to a client and a period, extraction with formal checks, first run 14 days after the workshop.
- Everything in the workshop
- One intake queue for email, chat apps, scans and structured e-invoices
- Assignment to a client and a period, with a manual queue for the disputed cases
- Field extraction from the invoice and formal checks before the entry is handed over
- Automatic reminders about missing documents
- Training for the person who runs this on your side
Document flow and an AI agent
Quoted after the workshop
for firms where the same client questions come back every day
Everything in the document flow plus BOND, the company AI agent that answers from your procedures and your data. Agent rollout takes around 2 months.
- Everything in the document flow
- An AI agent answering from your procedures in 5 to 10 seconds
- A queue of client questions, with the hard ones passed to an accountant
- A close status view across the whole client portfolio
- Access for the firm's team, with the data scope agreed and written down
The workshop ends with a document that stands up without us: a map of the document flow, the route into your accounting program and a 30 day plan. If it turns out that tidying one inbox and switching on a module you already own is enough, we say so plainly and show you how. The matter ends at the workshop then, rather than at a rollout you do not need.
Frequently asked questions
Will AI do the bookkeeping instead of our accountant?
No, and that is not how we build it. The automation collects the document, attaches it to a client and a period, reads the fields and runs the formal checks, which is the work somebody does on a keyboard today. The entry into the books is approved by your accountant, with the document, the extracted data and the list of checks in front of her. Professional responsibility stays exactly where it was.
What happens to our clients' data, given it is financial and personal?
We agree the scope in the workshop and write it down before any rollout: the model gets specific sources and specific fields needed for the task, not access to the whole firm database. Processing runs at a provider you hold a data processing agreement with, in the region you name. Part of the flow can stay on your own infrastructure if you prefer it that way.
Can this be wired up to Comarch Optima or Subiekt?
Yes, but differently from a cloud service, and we say that at the start. Both programs sit locally and keep their data in a Microsoft SQL Server database, so they have no public address and there is no webhook. Writing goes through the vendor's layer, reading goes through read-only views, and a connector inside your network handles the whole thing by calling out for tasks itself. The first pass runs on a copy of the database.
How accurate is the invoice extraction and what happens when it gets something wrong?
Accuracy depends on the quality of the document, so we do not quote one number for everybody. We measure it on your own documents in the third step of the rollout, and your accountant's corrections give us the threshold: above it an entry goes for one-click approval, below it goes to manual handling. On top of that sit the formal checks, which catch a line total that does not add up, a duplicate and a wrong period.
What about structured e-invoices?
We treat them as one intake channel alongside email, chat apps and scans of paper. Invoices in a structured format carry the same fields every time, so extraction is simpler than from a photo. We do not replace the sending and receiving with our own code, because the vendor module in your accounting program handles that. Requirements differ by country and are being phased in on different timetables, so in the workshop we check what your clients' markets demand.
How long does a rollout in an accounting firm take?
One to two weeks of workshop, then 2 to 4 weeks of rollout, with the first working automation usually standing 14 days after the workshop. We start on documents from a closed period, so nothing goes into the books during the early steps. The AI agent answering client questions comes later, and its rollout takes around 2 months.
Do we have to change our accounting program?
No. We leave the program where it is and go in by the route that suits its kind: the vendor layer and the database for programs on your own server, the API over the internet for cloud tools such as Xero, QuickBooks or sevDesk. Migrating an accounting program is a separate project measured in quarters, and the first automation is meant to exist in 14 days.
We are 3 people with a dozen or so clients, is it too early for us?
It may well be too early, and we will say so on the consultation rather than sell you a rollout. With a dozen clients the biggest value usually comes from unifying the intake and from the chasers about missing documents, which is a narrow scope for narrow money. The full flow with extraction and an AI agent pays back once the peak of the month starts setting the limit on how many clients you can take.
Let us find where the queue forms in your firm
The free consultation takes 30 minutes. You tell us how many routes documents take into your office, which program you post in and what happens in the last week of the month. We tell you what can be taken over straight away, what needs the intake tidied up first, and what we do not touch at all because it stays with your accountant. You do not have to give us access to any system at this stage.
We run at most two rollouts at a time, so that someone is genuinely sitting on each of them.
Book a free consultationPS. If in the last week of the month somebody in your office is retyping invoices field by field and then writing emails about missing documents in the evening, that is exactly this problem. A document can find its own client and period, with the data read and the checks done, and your accountant can simply approve it. You write directly to Pawel, the founder, not to a shared inbox.