AI Agents for Lawyers

Practice Areas

AI for Competition and Antitrust Lawyers: Merger Filings, Dawn Raids and Data

AI competition lawyers use review agents for merger-filing documents, dawn-raid triage and economic data. A filing-support workflow, the recall numbers and the sanction risk.

By Dr. Niklas SchmidtUpdated 13 November 20268 min read
On this page
  1. Where the hours go in a merger filing
  2. What the recall numbers say about generative review
  3. A filing-support workflow
  4. Dawn raids: triage under time pressure
  5. Economic data and expert reports
  6. Advising on pricing algorithms
  7. The sanction risk in submissions to authorities
  8. Setting this up in a competition team

AI earns its place in competition work wherever documents outnumber lawyers: internal-document review for merger notifications, dawn-raid triage, and the summarising of economic evidence for submissions. Used with a written review protocol, it compresses a document sweep from weeks to days. Used without one, it puts unverified facts in front of a regulator that has the power to fine for misleading information.

Where the hours go in a merger filing

The expensive part of a notification is rarely the form. It is the internal-document exercise behind it: board papers, strategy decks, market analyses and competitor assessments that the EU Form CO and the US HSR item 4(c) and 4(d) requests demand, and that a case team will read for signs that the parties themselves see the deal as removing a competitor.

That is precisely the task that the first commercial legal agents were built for. On 6 April 2025 A&O Shearman and Harvey announced four agents, the first of them "antitrust filing analysis", sold to corporates and other firms on subscription with a revenue share for the firm. David Wakeling's line was that they "do in minutes what previously took several hours". The FT has since reported that Ashurst used AI to sift roughly 100,000 documents in 48 hours to decide what to submit to a regulator (paywalled, reported detail). Addleshaw Goddard adopted Legora after nearly a year of evaluation partly because Tabular Review analyses up to 100,000 documents in parallel; Thomson Reuters' CoCounsel Tabular Analysis runs 10,000 documents against 100 questions in one pass.

Those are throughput numbers. The question a competition partner should ask is about recall.

What the recall numbers say about generative review

The best independent data as of September 2026 comes from Redgrave LLP's case study (June–July 2026) comparing Relativity aiR with continuous-active-learning TAR on a 45,000-document set at 6% richness.

MetricRelativity aiR (generative)Active-learning TAR
Recall87–88%64%
Human effort18.5 partner hours (43 prompt iterations)1,123 contractor hours
Cost$33,451$57,405
PrecisionLower (over-includes)Higher

Two caveats travel with those figures. Redgrave's peers said they had never seen TAR results that poor in real cases, so the baseline may be unfair; and the over-inclusion meant second-pass volume roughly doubled. For a merger filing that trade-off is the right one. A regulator asks for completeness, and an omitted hot document costs more than a few hundred extra documents in the second pass. In Schulte v. LinkedIn (N.D. Cal., 30 June 2026) the court treated generative-AI review as ordinary TAR and refused "discovery on discovery" about its validation, which is helpful in litigation but no comfort with an authority that can simply ask how you reviewed. The wider mechanics of generative review are covered in AI in e-discovery.

A filing-support workflow

This is the sequence that survives contact with a case team. It assumes a governed review platform (Relativity aiR, Legora, Harvey Vault or CoCounsel), never a consumer chatbot.

  1. Scope with the deal team. Custodians, date range, and the document categories the form actually asks for. Write these down; they become the definitions in the prompt.
  2. Collect into the platform. If the documents live in the DMS, connect through a permission-bound route such as the iManage MCP server (14 May 2026), which respects ethical walls, rather than exporting to a laptop.
  3. Write the protocol as a prompt. Responsiveness definitions tied to the form's categories; a hot-issue taxonomy (market shares, named competitors, pricing intentions, closure or capacity language); privilege markers. The prompt below is a starting point.
  4. Calibrate on a sample. A senior associate codes 200–300 documents blind; compare, then iterate the prompt. Redgrave needed 43 iterations over 18 hours. Budget for that.
  5. Run at scale, quoting the evidence. Require the tool to quote the passage that supports each code. A code without a quoted passage is treated as unreviewed.
  6. Second human pass. Everything coded hot or privileged is read by a lawyer. Sample the non-responsive set to estimate elusion and record the number.
  7. Draft the narrative from the coded set only. Every factual assertion in the narrative carries a document ID. The model works only with the documents you give it, never from memory.
  8. Verify and file. A lawyer checks each cited document and each number against source; the review memo (protocol, sample results, elusion rate, who signed off) goes on the file. The citation-checking protocol applies to any authority cited in the legal analysis.
You are assisting a competition law team preparing a merger notification.
Work only from the document provided. Do not use outside knowledge about
the parties or the market.

For the document, return a JSON object with:
- "responsive": true/false against these categories: (a) analyses of the
  transaction's rationale; (b) assessments of market shares or competitive
  conditions; (c) references to named competitors; (d) plans concerning
  prices, capacity, output, product range or closures.
- "hot_issues": a list of any passages suggesting the transaction removes
  a competitor, raises prices, reduces capacity or output, or that a party
  is "dominant", "leading" or faces "no real competition". Quote each
  passage verbatim with its page or paragraph reference.
- "privilege_flags": any indication of legal advice, external counsel
  involvement or "privileged" markings, quoted verbatim.
- "confidence": high / medium / low, with one sentence explaining why.

If the document is illegible or truncated, say so. Do not guess.

Dawn raids: triage under time pressure

The raid itself is a lawyer's job: shadowing inspectors, asserting privilege on the spot, recording what is copied. AI arrives afterwards, when the client wants to know within days what the authority took and what it will find. Two uses hold up. First, privilege review across the seized set at speed, using a tool built for it (Relativity's aiR line includes a privilege product) with a lawyer making the call on every document flagged. Second, a chronology and issue map of the seized material, so the defence strategy starts from the same picture the case team will build.

The rule that never bends is that these agents are read-only. A document set assembled by an adversary, or drawn from an untrusted mailbox, is exactly the surface for prompt injection: an agent asked to summarise can be steered by text inside the documents. Zenity's "PleaseFix" research (disclosed 3 March 2026, presented at Black Hat on 5 August 2026) showed zero-click hijacks of browser agents through emails and calendar invites, and Anthropic's own view is that "no browser agent is immune to prompt injection". No send, no delete, no external tool calls during review. The mechanics are in agent security for lawyers and the prompt injection entry.

Economic data and expert reports

Models that can run code have changed what a lawyer can do with economic evidence without waiting for the economist. Gemini Notebook attaches a "secure cloud computer" to each notebook (16 July 2026), and Claude Cowork, Claude Code and ChatGPT Work all execute analysis on uploaded spreadsheets. Anthropic reports that Claude Opus 5 (24 July 2026) raised financial-modelling accuracy by nine points with a third fewer tool calls, which is a vendor claim but consistent with what lawyers see in practice.

Sensible uses: reproduce a table from the other side's expert report and check whether it foots; summarise a 200-page report against your economist's reply, section by section; sanity-check a diversion-ratio or GUPPI calculation. The line is that the model never becomes the source of a number in a submission. It drafts around the figures the economists supply, and every figure in the final text is traced back to a schedule with a page reference. The same discipline applies to due diligence and to M&A work generally.

Advising on pricing algorithms

Competition authorities have been interested in algorithmic pricing for years, and the advisory work has a document-heavy core: how the client's pricing tool works, what data feeds it takes, who the vendor's other customers are, and what the internal messages say about "monitoring" competitors. A model is good at reading several hundred pages of vendor documentation and API descriptions and flagging where competitor prices enter the system or where a shared vendor could act as a hub. That is reading, not judgment. The characterisation of what you find, and the case law you rely on, stay with you, and research on that case law goes through a grounded tool with citation checking rather than a general chatbot.

The sanction risk in submissions to authorities

Court sanctions for fabricated citations dominate the headlines: Damien Charlotin's database held 2,041 decisions on 14 September 2026, 811 of them involving lawyers, and the cases are collected in AI hallucination sanctions. Filings to competition authorities do not appear there, because they are not court decisions. The exposure is different in kind, not smaller.

Under Article 14(1) of the EU Merger Regulation, supplying incorrect or misleading information in a notification can be fined at up to 1% of aggregate turnover, and a clearance based on incorrect information can be revoked. National regimes have equivalents. Two 2025–26 episodes show how this goes wrong in practice without any court being involved. Deloitte Australia partially refunded an AU$440,000 government report in October 2025 after a made-up court decision and non-existent academic papers were found in it. Sullivan & Cromwell apologised to a bankruptcy judge in April 2026 for AI-generated citations that the firm's review "did not identify". Neither organisation lacked a policy. Both lacked a verification step that someone actually performed.

Setting this up in a competition team

A practical sequence for the next quarter:

  • Pick the review platform the team already litigates on and confirm with the vendor, in writing, that data is not used for training and where it is processed. The vendor due diligence questions cover the rest.
  • Write two protocols as prompts: filing review and post-raid triage. Store them with the precedents, version them, and record who changed what.
  • Agree the second-pass rule (all hot and privileged codes read by a lawyer; elusion sample recorded) and put it in the engagement plan so the client sees it.
  • Train the team on the difference between a workflow and an agent, and on why review agents stay read-only. Module 4 of the AI Lab for Lawyers course builds a document-review agent live with exactly these approval gates, using browser-based tools only.
  • Log the training: Article 4 of the EU AI Act asks deployers to support AI literacy, and national authorities have supervised it since 2 August 2026 (see AI literacy for law firms).

The teams that get value from this are not the ones with the most tools. They are the ones whose review memo, six months after the filing, still shows exactly what was reviewed, how, and by whom.

Frequently asked questions

Can AI review internal documents for a merger notification?

Yes, and this is where competition teams get the most value. Generative review tools code documents for responsiveness, hot issues and privilege at higher recall than older TAR methods, but they over-include, so plan a human second pass. Run the review inside a governed platform, write the protocol as a prompt tied to the form's categories, and calibrate on a sample first.

Is it safe to use ChatGPT during a dawn raid?

Not with seized or client documents. Consumer chatbots may retain and train on inputs, and an agent reading hostile documents can be hijacked by instructions hidden in them. Use the firm's governed review platform, keep any agent read-only with no ability to send or delete, and keep privilege determinations with a lawyer.

What happens if an AI-drafted merger filing contains a false statement?

Authorities treat the notifying party as responsible regardless of tools. Under Article 14 of the EU Merger Regulation, incorrect or misleading information can be fined at up to 1% of turnover, and the authority can revoke a clearance obtained on false information. Verify every fact and figure against its source document before submission and keep a record of the review.

Can AI help with algorithmic collusion advice?

It helps with the reading, not the judgment. A model can go through hundreds of pages of a client's pricing-tool documentation, vendor contracts and internal messages and flag where the algorithm ingests competitor prices or shares data. Whether that crosses the line is your call, and you should not rely on a model for the case law.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion