AI-drafted court filings: why review now decides the cost

Key facts (as of 15 October 2026)

• 30 September to 3 October 2026: the presidents of Germany’s higher administrative courts and of the Federal Administrative Court asked the legislator for a legal basis to set binding rules on the length and structure of written submissions.
• 1 September 2026 (reported 5 October): Austria’s Federal Administrative Court (BVwG) fined a party €200 for flooding it with unlabelled AI-generated submissions and referred to Article 50 of the EU AI Act.
• August 2026: Hertie School researcher Chris Schmitz described the pattern as “agentic flooding”. Arbitration bodies, including the Schlichtungsstelle Reise & Verkehr, report record case numbers.
• 8 October 2026: Austria’s Bundesrat cleared § 18a AVG, which allows fully automated first-instance decisions, using fixed rules or AI, under documented conditions.

In the first week of October 2026, courts in Germany and Austria responded to AI-drafted filings from two directions. In Germany, the presidents of the higher administrative courts of the federal states and the president of the Federal Administrative Court (BVerwG) asked the legislator for a legal basis for binding rules on the length and structure of written submissions. In Austria, the Federal Administrative Court (Bundesverwaltungsgericht, BVwG) fined a party for flooding it with AI-generated submissions that had not been labelled as such. Neither court banned AI from legal work. Both said, in effect, that producing text has become cheap and that the cost of reading it is starting to fall on whoever files it.

On 10 October 2026, the Swiss daily NZZ put it in one sentence: „Das Schreiben kostet inzwischen aber fast nichts mehr – nur das Lesen ist noch immer aufwendig.“ (“Writing now costs almost nothing, but reading is still laborious.” Our translation.)

This article covers what the courts said, why the Austrian decision needs careful reading, what the research on “agentic flooding” adds, and where the review step belongs when a law firm or in-house legal department drafts with AI. It describes good practice and is not legal advice.

What the German administrative courts asked for

The presidents of the German Oberverwaltungsgerichte and Verwaltungsgerichtshöfe, together with the president of the BVerwG, met for their 65th annual conference in Lüneburg from 30 September to 3 October 2026. According to a press release from the Higher Administrative Court of Lower Saxony and a beck-aktuell report of 5 October 2026, they adopted a declaration asking the legislator to add a sufficiently specific basis to the Code of Administrative Court Procedure (VwGO). Courts could then issue binding requirements on the scope and structure of written submissions.

The presidents are concerned that the use of generative AI to draft submissions is increasingly confronting the administrative courts with considerable challenges.

The declaration contains no ban, no disclosure form and no technical filter. The judges want the power to require structure: where the application has to appear, how arguments are organised and how long a submission may run. If they get it, structure stops being a matter of house style and becomes a procedural requirement. The Federal Ministry of Justice is separately working on a broader procedural reform, but this request concerns the courts’ own ability to set the format.

Christian Wagner, a social-law specialist at the German Bar Association (DAV), told the NZZ: „Eine Klage, die 200 Seiten hat, ist dem Gericht ungeheuer lästig.“ (“A 200-page claim is an enormous nuisance to the court.”) He also said prospective clients increasingly send him their matter already prepared by AI. So the volume problem has reached lawyers’ intake too, where the person who has to read the text is the lawyer.

The duty to check is older than the volume debate. In November 2025, the Kammergericht in Berlin (17 WF 144/25) told counsel that the AI-inserted citations in her submissions had evidently not been checked, and that a lawyer must check them both under the mandate and as an officer of the administration of justice. California has since written that duty into statute with SB 574 (see our article on SB 574).

The BVwG’s first fine, and how to read the Article 50 reference

In Austria, the BVwG imposed what it described as its first Mutwillensstrafe, a fine for wanton or abusive conduct in proceedings, on a party who had flooded the court with AI-generated submissions. The decision is dated 1 September 2026 (W298 2341849-1), and beck-aktuell reported it on 5 October 2026. The fine was €200. The court counted it against the party that the machine-generated submissions had not been labelled as such, and it referred to Article 50 of the EU AI Act.

At €200, the case matters for its reasoning more than its penalty. A court treated unchecked volume as a burden on proceedings and attached a cost to it.

The Article 50 reference needs care. Article 50 sets transparency obligations for providers and deployers of certain AI systems. It does not obviously create a general duty for every party to label every AI-assisted submission, and the Austrian decision should not be read as saying it does. What is new is that a court reached for the AI Act at all when it assessed a party’s conduct. For firms and legal departments the practical lesson is narrower: decide internally when and how AI involvement in outgoing documents is disclosed, before a court raises the question.

Agentic flooding: why the volume keeps growing

The court declarations did not come out of nowhere. In August 2026, Chris Schmitz, a doctoral researcher at the Hertie School in Berlin, gave the pattern a name in a SPIEGEL interview (19 August 2026). He and his co-authors call it “agentic flooding”: a surge of applications, complaints and claims to public bodies and courts, written with AI and increasingly sent by AI agents.

In a dpa report published by beck-aktuell on 18 August 2026, Schmitz was careful about causality: case numbers also rise for economic and legislative reasons, so the share caused by AI is hard to prove. His study collected 84 potential cases in eleven jurisdictions; most of the filings come from people with real claims.

The same dpa report shows the signals in Germany: the Versicherungsombudsmann received around 50 percent more complaints in the first five months of 2026 than a year earlier, and the Schlichtungsstelle Reise & Verkehr recorded a record number of applications in the first half of 2026.

Schmitz sees the main risk in how institutions react. If they respond with citizen-unfriendly measures, such as closing digital channels or allowing more cases to be rejected without reasons, access to justice suffers. His alternative is less dramatic. Many processes could run fully deterministically, with ordinary rule-based software and no human needing to look at each case. Digitise first, add identity checks, structure the interfaces and the data, and only then ask whether AI should be used in discretionary decisions.

What handling volume looks like in practice: the SRUV

The Schlichtungsstelle Reise & Verkehr (SRUV) is one of the bodies on the receiving end, and it shows what that alternative can look like. The SRUV handled more than 45,000 cases in 2025. Classifying them used to be manual: staff read each complaint and each company response, compared them, corrected inconsistencies and sometimes checked external facts such as the weather on the travel day. That took five to seven minutes per case and depended partly on who did it.

Using e! by Lexemo, the SRUV built legal bots that read the traveller’s complaint and the company’s response, identify what is undisputed and what is disputed, and assign the case to a category. Each assignment comes with a documented chain of reasoning that shows how and why the bot reached it. Classification now takes about one minute per case. The SRUV has since built more than 400 automations across its arbitration workflow, and its own team maintains them (more on how the SRUV transformed its workflow). People stay in the loop at the steps where an assessment is made. The arbitrator opens a case that is already sorted and summarised, and the reasoning behind the sorting is visible.

The SRUV is not a law firm, and its bots sort incoming cases rather than draft outgoing filings. The design is still the same one the courts are asking for: fixed structure, rules where rules are enough, AI where language has to be read, a documented reason for every result, and a person at the decisive step.

Austria’s § 18a AVG: conditions for decisions no person has seen

On 8 October 2026, the Austrian Bundesrat cleared an amendment to the General Administrative Procedure Act (AVG) and the Administrative Penal Act (VStG), after the Nationalrat had adopted it on 23 September (Parlamentskorrespondenz No. 923). The amendment covers chatbots that help people submit applications, fully automated written decisions, and “no-stop” procedures in which a benefit is granted without a prior application.

The core is the new § 18a AVG. In matters designated by regulation, first-instance written decisions can be issued using fixed rules or even AI, without a person ever having looked at them. It is notable in itself that the text names fixed rules and AI side by side. According to the adopted text, a documented test phase is a precondition. It must in principle be possible to intervene, and the system must be sufficiently traceable in technical terms. Instead of checking every single decision, the law provides for spot checks. AI faces an additional limit: it must not be self-learning. Each decision must state that it was issued fully automatically.

Anyone who receives such a decision can lodge a Vorstellung, a simple objection, within two weeks, which sets the decision aside. According to heise, this is meant to satisfy Article 22 GDPR. The authority can also repeal or amend an automated decision on its own initiative for two months afterwards.

The debate in the Bundesrat is almost as interesting. The opposition voted against. The FPÖ questioned whether automated decisions are traceable and who is responsible for them. The Greens warned of legal uncertainty because automated decisions can still be repealed on the authority’s own initiative for two months, and said it remained unclear how the AI reaches its decisions. State Secretary Alexander Pröll replied that full automation will only be allowed where it is suitable, with test phases and controls as safeguards.

§ 18a does not, of course, apply to a law firm or a legal department. The view to Vienna is still worth it. Taken together, the conditions and the criticism read like the questions that come up internally with any automation:

• Was it tested on real cases before launch, and was that documented?
• Is there a place where a person can intervene, and does everyone involved know where it is?
• Can you show an outsider how a specific result came about?
• Does the system change between testing and live use only when someone deliberately changes it?
• And who is responsible when a result has to be corrected afterwards?

Five checks before a filing goes out

Read together, the German request, the Austrian fine and the flooding research object to the same thing: text that nobody checked, with the cost of reading it passed to someone else. A firm saying it uses AI responsibly does not solve that. A drafting process does, if the review is a defined step with a named person, a fixed point in the sequence and a record.

Most firms and legal departments already review their filings, but often implicitly. “Someone always looks over it” works until that person is on holiday, or the filing goes out late on a Friday. These five checks cover most of it.

Can the court find the application quickly?

If courts may soon require a set structure, the prayer for relief and the core argument should be findable without reading forty pages first. A short, fixed opening section is a cheap habit to build now.

Has every citation been checked against the source?

Every case reference and statutory quote that came out of a model should be opened and read by a person before filing. The Kammergericht said this in 2025. It is still the most basic part of the review and the one most often skipped under time pressure.

Has the firm decided how to handle AI labelling?

Make a deliberate internal decision on whether and how AI involvement is disclosed in outgoing documents. The Article 50 question is open, but the Austrian court has shown that missing labelling can count against a party.

Can someone justify the length?

If a submission runs long, someone should be able to say why. Volume that exists only because it cost nothing to produce is exactly what the German presidents want the power to limit.

Who checks, and when?

The other four checks depend on this one. The review needs a named role and a fixed place in the sequence, before signature and before filing. Where the drafting process runs as a workflow, for example a bot in e! by Lexemo, the review can be its own approval step: a named person receives a one-time approval link, and the decision is recorded in the bot’s report. The same works with paper and a checklist, as long as the step is written into the process.

One more layer for in-house legal departments

Business units increasingly send material to legal that was prepared with AI. That is agentic flooding inside the company. The intake step then needs its own check: what came in, who prepared it, and what has to be verified before anyone builds on it. The SRUV approach applies here too. Sort and structure incoming requests by fixed rules, document why each one went where it went, and let a person decide the cases that need judgment.

Frequently Asked Questions

What did German administrative courts ask for on AI filings?

At their annual conference in Lüneburg from 30 September to 3 October 2026, the presidents of the German higher administrative courts and the president of the Federal Administrative Court asked the legislator for a legal basis in the Code of Administrative Court Procedure (VwGO) to set binding requirements on the scope and structure of written submissions. They cited concern about the growing use of generative AI to draft submissions. They did not ask for a ban on AI-assisted drafting or for a disclosure rule.

Why did Austria’s BVwG fine a party over AI submissions?

In a decision of 1 September 2026 (W298 2341849-1), the Bundesverwaltungsgericht imposed its first fine for wanton conduct (Mutwillensstrafe) on a party who had flooded it with AI-generated submissions. The fine was €200. The court treated the missing labelling of the submissions as an aggravating factor and referred to Article 50 of the EU AI Act.

Does Article 50 AI Act require lawyers to label AI-assisted filings?

That should not be assumed. Article 50 sets transparency obligations for providers and deployers of certain AI systems, and its reach into individual court submissions is not settled. The Austrian decision shows that a court may take missing labelling into account. It does not establish a general labelling duty. Firms and legal departments should make a deliberate, documented decision on disclosure and review it as courts and regulators develop the point. This is good practice, not legal advice.

What is agentic flooding?

Agentic flooding is a term used by Chris Schmitz (Hertie School) and co-authors for the rising volume of applications, complaints and claims that reach public bodies, arbitration services and courts because people draft them with AI or send them through AI agents. Most of these filings come from people with legitimate concerns. The difficulty is the reading and checking load they create for whoever receives them.

What does Austria’s new § 18a AVG allow?

Cleared by the Bundesrat on 8 October 2026, § 18a AVG allows Austrian authorities to issue first-instance written decisions in matters designated by regulation fully automatically, using fixed rules or AI, without a person looking at them. Preconditions are a documented test phase, the ability in principle to intervene and technical traceability; instead of checking every decision, the law provides for spot checks. AI used for this must not be self-learning. Recipients can object within two weeks, and authorities can repeal such decisions on their own initiative within two months.

How can law firms and legal departments prepare now?

Make the review an explicit step in the drafting process. Five checks cover most of the risk: the application is easy to find, every citation has been checked against its source, AI labelling has been decided, the length can be justified, and a named person reviews the document at a fixed point before signature and filing. Where drafting runs through a workflow, the review can be an approval step whose outcome is recorded, so it is clear afterwards who checked what and when.

This article gives a general overview as of 15 October 2026 and is not a substitute for advice on a specific case.