What is an AI assistant for accountants, and how is it different from an agent?
An AI assistantAI assistant. An AI assistant is a helper built on a neural network that, at a person’s request, answers questions, prepares texts and does simple tasks in work programs. answers an accountant’s question in a chat window and returns text: a letter, a summary or a table of mismatches. It changes nothing in the accounting software by itself, and that is the difference from an agent, which carries out operations without a person.
An agent is connected to the accounting system as a separate project with its own access rights. Our guide on how to build an agent for the books covers that work. This page is about a ready-made assistant that you write to directly.
Work with an assistant always has four parts: the question, the attached document, the answer and a human check. The last part is not optional. An assistant can sound sure even when it is wrong.
European statistics show how common this kind of work is. According to Eurostat (11 December 2025), analyzing written text was the most common AI use among EU businesses in 2025 [4]. That year, 11.8% of companies with 10 or more employees used AI this way [4]. This is exactly the work you can hand to an assistant.
I spent twenty years leading software development for Walmart, Disney, Apple and Microsoft. I work with AI every day, but I am not an accountant, so I tie every rule below to a named regulator or law. I described the change in my book How Not to Go Bankrupt!? [7]. A few decades ago, the idea that a computer could analyze thousands of pages of legal or financial documents in seconds sounded like pure fantasy.
What can you hand an AI assistant every day?
Every day you can hand an assistant drafts, summaries and matching: work whose result you see in full and check in minutes. The table lists eight such tasks, with what the assistant does, what the person checks and how high the risk is.
| Task | What the assistant does | What the accountant checks | Risk |
|---|---|---|---|
| Letter to a counterparty about a balance mismatch | Drafts it from anonymized amounts and dates | Amounts, dates and the reconciliation statement number | Low |
| Reply to a tax authority’s letter asking for explanations | Drafts the text without figures | Adds the figures from the accounting records and checks them against the tax provision named in the request | Medium |
| New guidance letter from the tax authority or finance ministry | Summarizes it in plain words with a reference to the paragraph | Finds that paragraph in the letter itself | Low |
| Spreadsheet formula | Writes the formula and explains it | Tests it on five rows with a known answer | Low |
| Matching two exports | Finds rows where amounts or dates differ | Recalculates the rows found and the totals | Medium |
| Note to management on growing accounts payable | Drafts it from the trial balance figures | Figures and reasons against the documents | Medium |
| Sorting incoming documents by type | Suggests the document type and folder | Every item the assistant is unsure about | Medium |
| Details from a supplier’s incoming invoice | Copies the tax ID, bank details and amount into a table | Every field against the invoice itself | Medium |
All examples use labels such as “Employee 1” and “Counterparty A” instead of names, and drop account numbers. The assistant does not invent figures for a tax explanation. The accountant adds them from the accounting system once the text is ready.
I treat an assistant like a new employee who gets part of the work. In the same book I compared it to the way great leaders once delegated work to their staff [7]. Now we can delegate a certain part of our tasks to artificial intelligence. The key word is part: you delegate what you can check afterwards.
What should an AI assistant never do?
An assistant cannot sign, approve a document for the accounting records or file tax and financial reporting. Each of these acts carries a person’s name and a person’s liability. Regulators draw the same line when they describe what AI tools are for.
The FRC guidance for audit firms of 30 March 2026 works through two examples [6]. One summarizes board minutes, the other reviews contracts while testing revenue recognition [6]. Both are drafting and checking tasks, and in both a person reaches the conclusion.
Fake paperwork is the hardest line. An assistant can produce a convincing invoice or certificate for a deal that never happened if someone asks for one. So the ban has to rest with a person, not with the program.
Disputes over the books end the same way. Someone with authority decides, puts the decision in writing and answers for the result. A chat window has no place in that chain.
Tax calculation also stays with the accountant: it rests on the firm’s own accounting records, rates and reliefs. The assistant can still help: ask it to explain the method and to check the arithmetic on figures you took from the accounting system. Then you check the result against the tax law’s rules for that tax.
What data should you never paste into an AI assistant?

Personal data of staff and individual clients, and your firm’s confidential information, should not go into an outside service without a lawful basis. For data about people, that means a basis under Article 6 of the GDPR [11] or the UK GDPR [12].
ICAEW, the Institute of Chartered Accountants in England and Wales, wrote on 9 November 2026 about client data in AI tools [16]. It warns against entering confidential information into open source AIopen source AI. Open source AI is an approach in which a model is published so that anyone can freely use, study, change and share it: with the weights, the code and information about the training data. tools [16]. Under R114.1 of its Code of Ethics, accountants must comply with confidentiality, one of five fundamental principles [16]. The UK data regulator ICO wrote its guidance of 11 November 2026 for organizations receiving AI-drafted requests for personal data [5].
The reason is where the data goes next. ICAEW’s Sarah Brook says entering information into a tool outside a secure environment can be like posting it publicly [16]. Once a client’s salary sits on someone else’s servers, you cannot be sure it is gone.
Confidential business information needs the same care. The terms you agreed with partners, your margins or a pending deal belong to your firm and often to your clients and counterparties as well. A leak of them hurts both sides, even when nobody meant any harm.
Before you send anything, remove or replace every field that identifies a person or a deal:
- full names of employees and individual clients;
- passport and social security numbers;
- individual tax IDs;
- the salary of a specific person;
- bank account and card numbers;
- deal terms your firm treats as confidential.
Upload a bank statement only in anonymized form: it holds payee names, account details and amounts paid to specific people. The practical route is to export it to a spreadsheet and delete the name and account columns. Keep dates and amounts, and delete payment descriptions that name a person. For the full statement you need a service whose agreement with your firm covers where that data is stored.
To reconcile the statement with the accounting records, put the bank account’s daily movements from the general ledger next to it. Ask the assistant to find the days where the statement and the ledger totals differ. The accountant checks every day it finds in the accounting software.
In the US, a tax preparer who knowingly or recklessly discloses return information commits a misdemeanor under 26 U.S.C. §7216 [13]. The section and Treasury regulations list the exceptions [13]. The Federal Trade Commission’s Safeguards Rule (16 CFR Part 314) covers tax preparation firms [14]. They must oversee service providers and require those safeguards in their agreements with them [14].
In a 2026 recording I warned that the data we leave on other people’s servers is used to train AI [10]. The world feels safe now, I said, but in the future that data can be turned against you. That happens when the provider has not ruled out learning from your data in its terms or in a setting: the third question below. For the legal side, read our analysis of AI and personal data.
Can you use an AI chatbot hosted in another country if you remove the names?
Yes, if the text no longer contains personal data or confidential business information. If data that identifies a person stays in it, you are sending it across a border. A labeltarget variable. A target variable is the value a machine learning model learns to predict: whether a customer will leave, how much product will sell, how much a flat costs. is not enough if a salary or date still points to one person [11]. Under GDPR Recital 26, such text is still personal data [11].
Chapter V of the GDPR and the UK GDPR then applies [11] [12]. The data may leave only for an adequate country or under safeguards such as standard contractual clauses, with narrow exceptions in Article 49 [11].
The same ICO guidance notes that AI systems can retain and reuse what they receive for training [5]. So the transfer does not end when the chat closes.
In the EU, the AI ActAI Act. The EU AI Act is Regulation (EU) 2024/1689 on artificial intelligence. It sorts AI systems by level of risk and sets requirements for each level. adds a duty for the firm itself. The Digital Omnibus on AI of 8 July 2026, Regulation (EU) 2026/1744, kept it with providers and deployers [15]. Under the amended Article 4(1), they shall take measures to support the development of AI literacyAI literacy. AI literacy is the ability to use artificial intelligence with benefit and without harm: to understand what it can do, where it makes mistakes, which information should not be trusted to it and how to check an answer. of their staff [1]. Those measures also cover other persons who deal with the operation and use of AI systems on the firm’s behalf [1].
The duty does not require them to guarantee that any individual reaches a specific level [1]. Under Article 4(2), the Commission and EU countries support these efforts, in particular those of small and medium-sized enterprises [1]. That is still the consolidated text on the Commission’s AI Act Service Desk, version of 27 July 2026 [1]. A firm whose accountants use an assistant is a deployer in the sense of the Act.
The practical conclusion is simple: anonymizedata anonymization. Data anonymization is the processing of information about people so that nobody can tell whom it concerns without extra information. the text before it leaves your desk, rather than trusting the service’s settings. Legal reviews take days, and you open the assistant dozens of times on an ordinary working day.
How do you check an AI assistant’s answer?
You check an assistant’s answer against the source, by recalculating and with a second answer, never by how confident it sounds. In its guidance of 11 November 2026 on subject access requests drafted with AI, the ICO lists their typical errors [5]. They include references to other countries’ laws and case law that does not exist, which the ICO notes are sometimes called hallucinationsAI hallucination. An AI hallucination is an answer from a neural network that sounds confident but contains invented or wrong facts. [5]. I see the same risk in any AI answer an accountant gets.
The check takes four steps, and each next step matters only if the previous one passed:
- Open the law article or letter the assistant cited and find the paragraph it named: no paragraph, no answer.
- Recalculate the amounts and check the dates in the accounting system, not in the assistant’s answer.
- Ask the same question a second time, or ask another model, and compare the answers.
- If the answers differ, the accountant decides from the document, not by a majority of answers.
I do the third step myself every day. In a 2026 talk on how AI tools collect data, I described my method [9]. I ask one question to several products, and several times to the same one. Sometimes the answers match, and then I understand that the truth is most likely somewhere there: I call it cross-checking.
Matching answers still prove nothing. Two models can make the same mistake. For me, agreement is a reason to open the document, not a reason to skip it.
Who is responsible if the assistant makes a mistake?
ICAEW wrote on 27 January 2025 that its Code of Ethics leaves the judgment on technology output with the accountant [17]. Under R220.8, whoever uses such output to prepare or present information sets the appropriate steps by professional judgment [17]. R320.11 adds a step before use [17]. An accountant who intends to apply the output in professional work first determines whether that is appropriate for its purpose [17].
In my experience, people answer for an error, not the program. That means whoever signed the document and whoever runs the accounting. The UK audit regulator FRC issued guidance for audit firms on generative and agentic AIagentic AI. Agentic AI is an approach where a neural network receives a goal and carries the work to a result on its own. It plans the steps, calls software tools and checks what came out. on 30 March 2026 [6].
Its press release says that regulatory responsibility for deploying AI tools and for the quality of audit work stays the same [6]. As auditing standards set out, the human auditor is always accountable, and the guidance does not change that [6]. The FRC wrote this for auditors; I apply the same principle to bookkeeping.
The same logic holds for a draft. If the assistant made an error in a letter and you signed it, the error became yours.
The FRC describes this as the first guidance on generative and agentic AI from an audit regulator anywhere in the world [6]. It is also the FRC’s second guidance on AI in audit [6]. So I read the position as settled rather than new.
In a 2025 conversation I retold a Harvard professor’s idea, not my own; I did not name him then [8]. I agree with it. If I use artificial intelligence as a tool, and I personally made the decision to use it, then I am responsible. An accountant who chose to open the chat answers for what he or she carried from it into the accounting records.
How many companies already use AI, and what do the figures say about finance?

Across the EU, 20.0% of enterprises with 10 or more employees used AI in 2025, according to Eurostat (11 December 2025) [4]. A year earlier the share was 13.5% [4].
Adoption also varies widely across Europe. In 2025, Denmark (42.0%), Finland (37.8%) and Sweden (35.0%) had the highest shares of firms using AI, according to Eurostat [4]. Romania (5.2%), Poland (8.4%) and Bulgaria (8.5%) had the lowest [4].
In Russia, text processing is used by 84.2% of financial-sector organizations that already use AI, according to HSE ISSEK (9 October 2025) [3]. ISSEK is the statistics and research institute of HSE University in Moscow. It reports that this is the highest share of any sector, and 66.2% of these organizations also use AI decision support [3].
Rosstat, Russia’s federal statistics service, surveyed over 15,000 large and medium organizations for 2024 [3]. They had over 242,000 workers with AI skills, 17% of them in the financial sector; 203,000 use AI at work [3]. HSE ISSEK puts all 242,000 at under 1% of the surveyed organizations’ staff [3]. In Russia, then, the assistant is not yet a mass tool.
Both sets of figures cover all kinds of AI, and neither source gives a separate share for accountants. If someone quotes a share of accountants using assistants, ask where it comes from: these surveys do not have it.
Does an AI assistant pay off if you still check every answer?
It pays off on tasks where checking the result is faster than doing the work, and not where the check equals the work. The FRC (30 March 2026) tells auditors that professional judgment sets how much checking an AI tool’s output needs [6]. That depends on the tool and the task, the FRC adds [6].
The main objection goes like this: checking eats all the savings. For tax calculations and journal entries that is true, since the accountant repeats the whole job. For letters, summaries and finding mismatches it is false: checking a finished text is faster than writing it.
My working rule, not a measurement: if checking a task takes over half the time of doing it by hand, do it yourself. The formula is simple: divide the minutes of checking by the minutes of manual work. A result above 0.5 keeps the task with a person.
An example with round figures. A letter to a counterparty takes 30 minutes by hand. The assistant’s draft plus your check takes 8 minutes: 8 / 30 = 0.27, so hand it over. An explanation of a tax calculation takes 40 minutes by hand and 30 minutes to check: 30 / 40 = 0.75, so there is no gain.
The 30 minutes come from my book [7]. There I wrote that a message to a business partner once took 30 minutes, and AI now writes it in seconds. That is my example, not a measurement, so time your own 10 tasks over a week.
How do you choose an AI assistant for bookkeeping?
Choose an assistant by three questions, not features: where is the data stored, is there a contract, and can you block training on your chats? On the first, ICAEW warns that entering information into a tool outside a secure environment can be like posting it publicly [16]. For data with names you need a service that commits in writing to how it stores and deletes your data. The other route is a local modellocal LLM. A local LLM is a neural network that runs on your own computer or server, not in a vendor’s cloud, so your data goes nowhere. on your own servers.
The second question is the contract: a service processing client data for your firm needs a processor agreement under GDPR Article 28 [11]. The third is machine learningmachine learning. Machine learning (ML) is a field of AI in which a program finds rules in data by itself instead of getting them ready-made from a developer. on your chats: text that a model learns from stays with the provider after the chat window closes.
I compare services by these features rather than by price plans. Features last longer than plans. A comparison by task, language and access is in our review of AI assistants.
An assistant built into the accounting system itself is judged by the same three questions. 1C is accounting software common in Russia. What a built-in tool can be trusted with there is covered in our analysis of AI in 1C.
Where should an accountant start in the first week?
Start with three low-risk tasks from the table above and one written rule on what data never goes into the assistant. Agree the rule with your manager or the partner in charge. The firm answers for how its staff use the tool. A rule nobody signed off is a rule nobody follows.
- Pick three tasks: a letter to a counterparty, a summary of new guidance and matching two exports.
- Write down the anonymization rule from the data section and agree it with your manager.
- Keep a log: the task, minutes by hand, and minutes with the assistant including the check.
- After a week, keep only the tasks where the check took less than half the manual time.
I wrote about this kind of time tracking in my book [7]. Intelligent assistants and specialized platforms help you analyze where your time goes and sort tasks by priority automatically. That improves overall efficiency.
A week of logging gives you figures for your own tasks, and with them any talk about wider use rests on facts. The assistant’s value shows in minutes, not in impressions.



