Section: Justice & right Author: V Reading Length: ~27 min Sources and further reading: 22 items Topics: AI, law, advocacy, accounting, auditing, court translation, regulation SEO / Working Title: Can AI replace lawyers, accountants or translators?
Can AI replace lawyers, accountants or translators? The question sounds simple until we notice that we are combining three different stamps. Passing the American bar exam is not registration in the Czech Bar Association. Accounting software is not an auditor's report. The BLEU score is not an act of interpretation before a public authority. And a high-stakes system for justice is not a ban on office searches. Therefore, the article does not look for which of these professions will fall first. Looking for which onethe act is bound by the law to a person with authorization, and which concept the model may speed up without bearing a signature.
1. The stamp on the translation of the judgment is not BLEU on the test set
A seemingly identical scene takes place in three offices on the same day. The assistant enters the draft contract into the tool and wants a faster first version. The accountant lets the system sort the documents and asks if he can close the month already. A translation agency will bring a machine translation of a judgment with a good score and does not understand why the court does not accept it as an act. Next to them, a colleague shows an article about GPT-4 passing the Uniform Bar Exam, and another mentions a study that they arelegal services among industries with high exposure to language models [13][16].
At first glance, it is a single dispute: the cognitive profession against the model. In fact, each describes a different action. The assistant solves the text. Accounting entries. Agency similarity of translation to reference. Bar exam test with key. Felten, Raj, and Seamans overlay of language modelers' capabilities with descriptions of American industries [16].
Therefore, they can all be partly right and the common conclusion still wrong. The model can be useful in the design. He can do research better than a junior with no experience. It can translate an email so that everyone can understand it. However, this does not mean that he was authorized to provide legal services, perform a mandatory audit, or perform court translation.
Three licenses, three prices of error. A benchmark is not a signature.
— Jiný Kontext
2. The text task has a regime other than legal effect
One model response has at least six layers. The first is a text task: summary, template design, typo check, working email translation, document type recognition. Exposure can be high here. The study by Eloundou and co-authors measures the technical exposure of tasks to language models and tools above them. A study by Felten, Raj and Seamans ranks legal services among the US industries with high exposure to language modelers. Not one of them measures Czechlicense [16][17].
The second layer is the legal effect. A submission for a client, an auditor's statement or an act of a court translator are not just sentences. They are manifestations of will and responsibility in a regime that entrusts the right to a certain person. The third layer is the license: registration, examination, pledge, chamber or ministry list. The fourth is privacy and data mode. The fifth is control and the cost of error. The sixth is the regulation of systems, if the instrument enters the judicial decision-making of the state.
So the sentence "the model can write it" is not an answer to the sentence "it can execute it". It can shorten the first draft. It can expand the variants. It can show a problem that one would notice later. However, he is not the holder of the authorization. And if an unauthorized person offers the result as a dedicated service, the problem is not that the text created the model. It arises because the concept has become an act entrusted by the law to someone else.
This boundary is often lost in practice because all layers have the same appearance on the screen. Contract clause draft, quote check, client response and finished submission are all text. They look like lines in CMS, email and PDF. But the law does not read them according to the font. It reads them by relationship, purpose, and the person issuing them. An internal note can be working material. The same sentence in a letter to a client can be advice. The same sentence in submission can be proceduralclaim. The model does not know the boundary between these modes by itself, unless a human puts it into the process.
Therefore, it is not enough to introduce the rule "AI can only help". Help with what? With the first outline, with the duplication check, with the translation of the appendix for orientation, or with the final conclusion? An organization needs an action map. For each line, it should be clear whether the output remains inside the office, whether it goes to the client, whether it enters the financial statements, whether it is attached to an official act and who approves it before sending. Without such a map, the model does not become unsafe becausewrites. It becomes dangerous because no one knows when the text stopped being just a concept.
3. The best tool is the one whose hallucination is carried by the check, not the signature
For these professions, it is not enough to ask how often the model answers correctly. Equally important is what his mistake looks like and who will see it. A failed sentence in an internal memo is cheap. A fictitious citation in submission is expensive. A wrongly assigned liability in accounting can be transferred to a management decision. The confusion of participants in the translation of the judgment is not a stylistic defect.
| A kind of failure | What does he look like? | How to test | What will limit the damage |
|---|---|---|---|
| A hallucinatory quote | Flowing paragraph, non-existent file or shifted conclusion | Check each decisive citation against a primary source | A person signs a submission; autonomous citation is not a working rule |
| Unauthorized service | Chat output is offered as legal advice, audit or court translation | Separate concept, control and dedicated action | Advocacy Act, CAČR and list of the Ministry [1][3][7] |
| File leak | Client material ends up in an inappropriate account or mode | Find out about the contractual regime, retention, anonymization and access | Confidentiality, practice texts and a guided instrument [1][22] |
| BLEU instead of effect | The translation looks similar to the reference, but does not perform the task | Human control of meaning, names and legal consequences | Registered interpreter or translator for the act [6][14][15] |
| Confusion of accountant and auditor | The AI conclusion is read as an audit opinion | Distinguish between bookkeeping and mandatory auditing | Unit responsibility and auditor's opinion [2][3] |
Therefore, the best system is not the one that speaks most confidently in the demo. It is a system whose failure has a place in the process. Who controls the spring. Who decides on publication. Who bears the signature. And where does the work stop if the decisive data cannot be verified.
The check must also have a price on the calendar. If a lawyer saves ten minutes drafting a paragraph and then spends twenty minutes verifying an obscure citation, the system can still be useful as a brainstorming tool. However, it is not cheaper for a decisive output. If the accountant gets the documents quickly sorted, but has to manually go through all the items again that should have been solved by automation, the saving is moved to another column. If the translator gets a fluent versionjudgment, but must verify participants and negation in every other sentence, the tool helped with rough material, not responsibility.
This is not an argument against using models. It's an argument against buying by sample. A pilot should measure the cost of an accepted result, not the cost of a single response. An accepted result means that it has been verified, that it fits within the permission, and that someone can stand for it. For regulated professions, this last condition is stricter than for marketing text. This is precisely why a model's ability must not be mistaken for a signature.
4. A lawyer is not someone who writes a text about law
The Advocacy Act does not speak of pretty legal text. He is talking about providing legal services. § 2 paragraph 3 of Act No. 85/1996 Coll. stipulates that only persons specified by law may provide legal services on the territory of the Czech Republic. Section 16 requires the lawyer to protect and promote the rights and legitimate interests of the client. Section 21 deals with confidentiality, including that the obligation continues after delisting and also applies to other persons under the statutory regime [1].
This does not mean that a lawyer may not use the model. It means that the model is not a lawyer. A concept from a model can be a work tool like a lookup, a pattern, a comment, or an internal template. Only the result provided to the client in a certain relationship and under a certain responsibility becomes a legal service. When provided by an attorney, the attorney is responsible. When it is offered as a legal service by someone without authority, it is not a debate about the model's capabilities. This is a ban debateunauthorized provision of legal services.
NSS in decision 6 Ads 30/2012-47 works with confidentiality as a protection of the client and the advocacy institute, not as a technical choice of tool [22]. That is essential. AI can change office workflow. However, it does not change the addressee of the status responsibility.
5. 13,319 registrations at ČAK are license status, not demand for concepts
With reference to the Chamber's statistics, the lawyer's journal reports a total of 13,319 lawyers by the end of 2025, of which 528 were newly registered that year. It lists 2,601 persons and 122 newly registered legal assistants. For lawyers under the age of 30, 545 persons are reported in 2025 against 763 in 2017, i.e. a decrease of 28.6 percent [18].
Those numbers are useful, but they have a narrow meaning. It measures the status and age structure of enrollment at ČAK. It does not measure demand for legal concepts. It does not measure how many searches the model has speeded up. Nor does it measure whether clients have stopped seeing lawyers. If we see in them aging and changing entry into the profession, we read the status denominator. If we turn them into evidence that AI is replacing lawyers, we are adding a cause that the source does not provide.
This is where the article differs from the general occupation map. He doesn't ask if being a lawyer is a safe job. He asks what license means. The status of 13,319 attorneys is not an argument that the work is forever the same. It is a reminder that the legal effect of service is not the same as the ability to write a paragraph about law.
6. Passing the Uniform Bar Exam is not an entry into the Czech Bar Association
Katz, Bommarito, Gao, and Arredondo report that the GPT-4 in their work scored approximately 297 on the Uniform Bar Exam in zero-shot mode. For the written parts of the MEE and MPT, the authors report an average of 4.2 out of 6.0. This is the work published for the preliminary version of GPT-4 and the US test context [13].
This is a significant signal of ability. The model was able to solve a standardized legal exam better than many would have expected until recently. However, it does not follow that he can represent the client, take over the file, maintain confidentiality, evaluate Czech procedural consequences or pass the Czech bar exam. UBE is an exam with a key, essays and assessment methodology. The Czech practice of advocacy is a legal relationship with the client, status discipline and responsibility.
The test results are good for one thing: they shatter the comfortable notion that models are just formulation toys. They are good for no other thing: they must not be read as licenses. Therefore, the question is not whether the model can answer the legal question in the test. It says who is allowed to issue the answer as a service, who verified it, and who bears the cost if it is wrong.
7. Hallucinating a quote is not bad style
Dahl, Magesh, Suzgun, and Ho examined legal confabulations in public language models on questions about randomly selected US federal decisions. They report that GPT-4 hallucinated in 58 percent of responses and Llama 2 in 88 percent. The population included US District Courts, US Courts of Appeals, and the Supreme Court of the United States; non-compliance with metadata and case law content was measured. It is not about Czech law and it is not about a verdict on every current model veach mode [11][12].
The caveat is important. These numbers are not to be transferred 1:1 to the Czech Collection of Decisions. Models tend to be better for newer and more significant things. A corporate tool with a managed database may have a different profile than a public chat. Still, the study shows something important: a fluent legal paragraph can fail where a mistake is most costly. Not in style, but in opera.
How to test a quote
Any definitive citation must go against the primary source. For practice or anonymized text, check the name of the court, the file mark, the judgment, the legal sentence and the scope of the conclusion. If the model lists the right source, it's not done yet. One must verify that a particular passage actually supports a particular sentence. Fluency is not proof. It's just a form.
8. Confidentiality does not carry over to public chat
Attorney confidentiality is not a menu switch. It is a legal obligation of a specific person and a circle of other persons in a certain legal regime. Section 21 of the Act on Advocacy protects facts that the lawyer has learned about in connection with the provision of legal services. Povinnost trvá i po vyškrtnutí ze seznamu. The law does not bind her to whether the text was written by a person or processed by a tool [1].
Putting a client file into a public chat is therefore a different question than using a model for a practice concept. We don't just decide whether the model answers correctly. We address who is the data controller and processor, what retention applies, who has access to interactions, whether the content is used to further improve the service and how the audit is conducted. An on-premise tool or contracted enterprise mode may have a different profile. Even then, however, the model does not become the bearer of secrecy in a punitive sense.
Before inserting the file into the model
Is the account personal, public, or contractual? Is the text anonymized or is it a living document? Who approved the data flow? Can output and input be deleted according to contractual rules? And who is responsible if the decisive fact falls outside the permitted range? Only after these questions does it make sense to talk about productivity.
9. An accountant is not an auditor
Accounting is often conflated with auditing in common debate. Legally, it is different layers. Act No. 563/1991 Coll. requires the accounting entity to keep accounts so that the financial statements give a true and fair view. § 5 also stipulates that entrusting another person with accounting does not relieve the accounting unit of responsibility for accounting [2].
Act No. 93/2009 Coll. another act deals with auditors. A mandatory audit can be performed by an auditor or an audit firm authorized by law. An auditor's report is not the same as a well-kept spreadsheet. It is a regulated conclusion of a person who meets the legal conditions and is subject to the rules of the profession [3][4].
AI can help in accounting. It can suggest posting a document, spot a recurring error, prepare a checklist, or explain the difference between an account and an item. This does not mean that the entity has lost responsibility. And it does not mean at all that an audit opinion has been issued. Software can speed up documents. Mandatory audit is a different legal act.
It is important not to skip the word "unit" here. Accounting responsibility is not just a personal label of the accountant in the office. The law binds it to the accounting unit and at the same time admits that the processing can be entrusted to another person. However, according to § 5, the authorization does not remove the responsibility for keeping accounts [2]. So when a company adds AI between the document and the statement, the obligation to know what happened does not disappear. It just adds another processing layer.
The contrast against auditing is even sharper. An audit does not mean that someone has "counted the spreadsheet a second time". The auditor expresses a statement according to the legal regime and professional rules. AI can help with analytics, sample selection, anomaly detection, or working note formulation. He cannot become a person with audit authority. If these layers are mixed in an organization, a false sense of security is created: the system has checked something, so it is an audit. It's not an audit. It iscontrol tool before or alongside the audit.
10. 1,019 auditors is a license to audit, not the number of accountants in the economy
In its materials, the Chamber of Auditors of the Czech Republic lists a total of 1,019 statutory auditors by the end of 2025. For comparison, it states 1,374 auditors in 2014 and 1,133 in 2020. For self-employed auditors, it states 327 persons at the end of 2025 and a maximum of 647 in 2003. The text of the Chamber talks about the declining number of auditors, aging and the need to promote the profession [5].
It's easy to make a bad headline here. The number of auditors is decreasing, AI is growing, so AI is replacing auditors. The source doesn't say. It shows the state and development of the estate population. It does not separate demographics, the attractiveness of the profession, the difficulty of the exam, the labor market, remuneration and technological substitution. And they don't even measure the number of accountants in the economy.
The meaning of the number is narrower and more precise. 1,019 statutory auditors is the number of persons licensed to perform a specific regulated activity on a specific date. If a company uses AI to check documents, this does not change the definition of a mandatory audit. If the auditor relies on an analytical tool, his statement does not disappear. The labor savings can be real. It does not become a license.
A simple rule follows from this for company management. The tool can only be evaluated according to the speed of the task that we actually hand over to the tool. In the case of a licensed statement, the control documentation must also be evaluated. The auditor or accounting entity needs to know why the system suggested the item, where it got the input from, and whether the person saw the exception. Otherwise, automation will not create certainty. It just creates a faster path to an unclear conclusion.
11. A market translator is not a court translator
Translation is the second area where the sentence "the model can do it" is confused with the sentence "the authority will accept it". The free translation market includes emails, marketing, technical documentation, subtitles and internal memos. There, machine translation can lower the price and speed up the work. Some outputs are only indicative. Others need revision. This is the market for quality and contract liability.
Court interpreting and translation work is a different regime. Act No. 354/2019 Coll. binds activities before public authorities and acts stipulated by law to be entered in the list of court interpreters and court translators. The Ministry of Justice lists requirements for language, professional competence and other conditions for registration. Professional care, impartiality and promise are part of the regime [6][7].
The model can prepare a working translation for understanding. It can be a prelude to a human. He cannot take a vow or bear impartiality himself. If the action is intended for a court, notary or other public authority under the law, it is not decisive whether the output looks good. The decisive factor is who performed it as an act.
As with advocacy, this is not an aversion to technology. A court translator can use dictionaries, databases, memories, corpora and machine drafting. The decisive thing is that the result is not transmitted by the machine. It is delivered by an enrolled person who understands why the confusion of "may" and "shall" is dangerous in a particular sentence, why a participant's name is not cosmetic, and why a date or a negative particle can change procedural meaning. The model can suggest fluent Czech. Impartiality ahowever, professional care cannot be read from fluency.
12. BLEU has no legal effect
BLEU was created as an automatic machine translation metric. Papineni and co-authors in 2002 described a method based on n-gram matching of a candidate translation with reference translations. It is a useful technical indicator for certain system comparisons. However, it does not measure whether the translation will stand up as a legal act [14].
Callison-Burch, Osborne, and Koehn already pointed out in 2006 that a higher BLEU is neither a necessary nor a sufficient condition for better human quality assessment, especially when comparing systems with different strategies. A translation can be close to the reference in parts and still change the decisive meaning. It can retain many words and confuse the participant. It can sound good and move the legal relationship [15].
Court translation cannot solve this risk with a number. The act is not just a textual similarity. It is meaning, impartiality, accountability and an identifiable authority holder. Therefore, the sentence "the machine has a high BLEU" is similar to the sentence "the student wrote the test well". It may be relevant to the ability. It's not a stamp.
13. High-risk justice is not a ban on the search tool in the office
Regulation (EU) 2024/1689 classifies in Annex III as high-risk, among other things, AI systems intended for a judicial body or a similar alternative dispute resolution body to assist them in the interpretation of facts and law and in the application of law to a specific factual situation. This is an important boundary. It targets the system for the decision-making body, not automatically every private lawyer chatbot [8][9][10].
It does not follow that the office search is without rules. The lawyer still bears confidentiality, the duty to protect the client and the responsibility for the output. If the instrument helps to prepare a pleading, one must not mislead the court with a fabricated citation. If it works with personal data or trade secrets, it must have a legal and contractual regime. But the AI Act is not a simple “lawyers are not allowed to use AI” sign.
Annex III point 4 on employment should be read with the same caution. It refers to the recruitment, selection, evaluation, termination or assignment of tasks according to certain characteristics in a work context. It is not a chapter about the demise of accountants. Regulating the use of the system is not a profession forecast.
14. The disciplinary order and the promise of impartiality are not transferred to the model
The three professions are united by a simple principle: law addresses responsibility to a person or a legal entity in a certain regime. A lawyer has a duty to protect the client's interests and maintain confidentiality. The auditor takes an oath and is subject to the rules of the Chamber of Auditors. A court interpreter and translator performs an act with an obligation of professional care and impartiality [1][3][6].
A model can enter a work chain but not take over its end. If the attorney signs the filing with a fictitious source, the client will not complain about the model. If the auditor uses an analytical tool and overlooks an error in the audit evidence, the opinion remains his. If the registered translator lets the machine translate the text and confirms the wrong meaning without checking, the action is still his.
This is not nostalgia for manual labor. It is a construction of responsibility. Regulated professions did not come about because a person typed a sentence slower than a machine. They arose because someone must be identifiable, competent, controllable and punishable. The model can increase performance in the middle of the process. He cannot become the addressee of disciplinary proceedings.
15. The three prices of error do not add up to one occupational hazard
Legal advice, financial statements and court translation may look similar on screen. All are text or table. All can be spawned after the prompt. They can all be fluent. But their mistake is paid differently.
In advocacy, an expensive mistake is often a support: non-existent case law, missed deadline, bad procedural strategy, breach of confidentiality. In accounting, an error may appear to be a small item that translates into a true and fair view, but the entity's responsibility does not disappear because it has commissioned the processor itself [2]. In a judicial translation, it can be a name, a date, a negative particle or a relationship between the participants. High similarity of sentences does not guarantee legal applicability.
Three license tests
Advocacy: is it an internal concept or the provision of legal services? Audit: is it data processing and control, or the auditor's statement? Translation: is it a work translation of the text or an act before a public authority? If the answer falls into the third part, we are not only dealing with the model's capability. We decide who is allowed to act.
16. Three columns on a practice text will tell more than a replacement headline
The practical test can be done without a live file and without trying to bypass the license. Take a practice legal text, a public regulation, an anonymized document or an old translation for which you have the correct solution. Divide the work into three columns. The first is the concept. The second is a check against the source. The third is a permission-bound action.
In advocacy, a model can suggest the structure of a letter, but one must verify the legal support and decide whether it is a service to the client. In accounting, the system may suggest the inclusion of an item, but the accounting unit will not be released from responsibility and a mandatory audit will not occur. For translation, a machine can deliver a working version, but a court translation requires a registered person in the appropriate mode [1][2][6].
What needs to be tested
Test anonymized input, public source, and permission boundary. See where the model saves time and where it creates new inspection work. If the saving in concept is lost in citation verification, it is not a failure of the entire technology. It is a more accurate result price. And if the third column can't be passed to the model, the occupational replacement headline says more about the headline than the law.
17. An amendment without an authorization holder would break the thesis. He's not here yet
Every good thesis has a limit. This would be weakened by an amendment that would allow defined legal services, mandatory audit or court translation to be carried out by the system without the holder of the authorization. It would also be weakened by the Czech empirical series showing that models used in offices manage legal citations without expert control at a level comparable to responsible human research. Or data from ČAK, KACR and the Ministry, according to which the decline in numbers after demographic control would aremuneration was based on the substitution of AI.
The file does not have such documents. That's why it shouldn't be said as a fact. A narrower thing is to be said: the text tasks are exposed, the legal effect is regulated, and the responsibility remains with the bearer. This is less effective than the sentence "AI will replace lawyers". However, it is more accurate.
Accuracy in this debate does not hinder adoption. On the contrary, it enables it. A lawyer can use the model where he knows what he is checking. An entity can automate the routine as long as it does not lose responsibility. The translator can use the machine version as long as he doesn't pretend that the metric has replaced the promise.
The thesis boundary is also practical for suppliers. If he sells the instrument to lawyers, he should not promise to reimburse the lawyer. It's supposed to show how the tool works with sources, how it limits data leakage, how it logs resources used, and how it hands control over to a human. If it sells accounting automation, it should say where the posting proposal ends and where the accounting entity's responsibility begins. If he sells a translation for an office, he must distinguish a working translation from a judicial act. There is no exact bordermarketing weaker. It is a safeguard against a good instrument being sold as a bad legal fiction.
18. The question is not which of these professions will fall. It reads which act the law binds a person to
AI is changing the way you work with text, source, and routine review. For lawyers, accountants and translators, this can be significant precisely because their work involves a lot of language. But language is not the entire legal regime. Model capability and license are not one size fits all.
A benchmark is not a signature. The Uniform Bar Exam is not an entry into the ČAK. BLEU is not a court translation. Receipt software is not an audit opinion. And a high-risk justice system is not the same as a private search engine. Once we separate these things, the simple argument of whether AI will "replace" the three occupations disappears. A more practical map will remain: which tasks can be accelerated, which outputs must be verified, and which actions can only be done by an authorized person.
A model can write a draft. Signature, confidentiality and disciplinary liability remain with the authorized person.
— Jiný Kontext
So the right question is not: can AI replace lawyers, accountants or translators? It is: for which of the three acts, legal service, mandatory audit and court translation, does the law still require an authorization holder, and where is the model only allowed to speed up the concept, which still has to be checked by a person with a stamp?
Related texts in this series
- Which occupations will AI replace… — task exposure ≠ license.
- Can AI replace programmers? — different regulation, same distinction ability vs. signature.