# AI Agents for Legal and Compliance Work in Mid-Sized Companies

> An AI compliance agent can pre-review contracts against your own playbook, track deadlines, draft answers to supplier and security questionnaires, prepare export-control checks and sort GDPR requests. Legal judgement on the individual case and final sign-off stay with people and, where needed, outside counsel.

Canonical URL: https://www.ruebstudio.de/en/guides/ai-agents-by-department/ai-agents-legal-compliance  
Author: Dr.-Ing. Marcus Rüb, CEO  
Published: 2026-11-16

Most mid-sized companies have no legal department. The managing director reviews contracts between other tasks, deadlines live in personal calendars, supplier questionnaires end up with quality management, and anything contentious goes to an outside law firm. That gap is where an AI agent can help: not as a substitute for legal judgement, but as a diligent first pass that sorts, compares, pre-fills and reminds. This guide maps the tasks that fit, where the legal boundary sits, and how the hand-off to counsel should look. It is orientation, not legal advice.

## What is an AI agent for legal and compliance work?

A compliance agent is a language model with access to selected documents and tools that prepares recurring, law-adjacent tasks: checking contracts against internal rules, extracting deadlines, answering questionnaires from existing evidence. It proposes and explains; a person decides and signs off.

It is easy to confuse with two other things:

- **It is not a legal advice chatbot.** An open chat where staff can ask any legal question is hard to control and carries more regulatory risk. An agent instead works on defined processes with fixed inputs.
- **It is not just a contract management system.** A contract register stores files and sends reminders. An agent also reads the content, compares it with rules and writes a review note.

For a company without in-house lawyers, the agent is essentially an assistant to whoever already carries legal topics alongside their main job, often in finance, procurement or quality.

## How does a compliance agent work?

A compliance agent has four parts: a knowledge base of your own rules, access to the documents in a given case, a small set of tightly scoped tools, and a mandatory approval step. The model itself is replaceable. What matters is the rules, the data and the sign-off.

**Your rules, written down.** The core is not a legal commentary but your own position: which liability cap you accept, which governing law and venue you insist on, which payment terms are negotiable. In-house teams call this a playbook. Someone with legal knowledge writes it once; the agent checks against it rather than against whatever the model happens to remember.

**Document access.** Draft contracts in PDF or Word, questionnaires in Excel, supplier certificates, your records of processing. The agent sees only what the case needs. Existing repositories such as SharePoint or a document management system are connected through their APIs, not copied.

**Tools with narrow permissions.** Typical tools read text from documents, create a calendar entry, save a draft to a folder or open a ticket for the responsible person. Editing signed contracts, sending anything to third parties or signing are not among them.

**Approval and audit trail.** Every output arrives as a proposal with reasoning and a pointer to the clause or page. The log records who approved what, what was changed and which source was used. That is good practice, and it is also how you later show how a decision was reached.

## Which tasks can a compliance agent take on?

Five areas suit companies without a legal team: contract pre-review, deadline tracking, questionnaires, export-control preparation and data protection routines. They share fixed inputs, recurring patterns and a person who reviews the result anyway.

### Pre-reviewing contracts against your playbook

A customer sends its purchasing terms; a supplier sends a framework agreement. The agent compares each clause with your playbook and flags deviations such as uncapped liability, a foreign venue, unusual penalty clauses or automatic renewal with a long notice period. The output is a list with clause reference, your standard position and a suggested fallback wording. A person decides what to negotiate. A dedicated guide on playbook-based contract review is in preparation.

### Extracting deadlines and obligations

Notice periods, renewal options, warranty end dates, certificate expiry, follow-ups demanded in letters from authorities: the agent extracts the date, creates a calendar entry with a lead time and links the source. The real value is in the extraction, because deadlines are often buried in annexes and amendments.

### Customer, bank and auditor questionnaires

Supplier self-assessments, information security questionnaires and sustainability or due diligence surveys repeat themselves with slightly different wording. The agent answers from a curated library of approved answers and evidence, flags questions without a matching answer, and never makes a new commitment without asking. This pairs well with a general [knowledge management agent](https://www.ruebstudio.de/en/ai-agents/knowledge-management) that maintains the same answer library.

### Preparing export-control checks

Companies shipping technical goods abroad must check whether a licence is required. In the EU, the common list is Annex I of the Dual-Use Regulation, and Germany, for example, adds a national list; Germany's export authority [BAFA](https://www.bafa.de/DE/Aussenwirtschaft/Ausfuhrkontrolle/Gueterlisten/gueterlisten_node.html) names both Annex I and Part I of the German export list as the relevant control lists, and states that its tariff-code conversion table is only an aid that does not replace the exporter's own check. An agent can compare datasheets with the technical parameters of list entries, screen recipients against sanctions lists and assemble a review file. Classification itself is signed by the responsible person in the company. Outside the EU the lists differ, but the pattern is the same.

### Data protection routines

Several GDPR duties are deadline work. Under [Article 12(3)](https://dsgvo-gesetz.de/art-12-dsgvo/), data subject requests must generally be answered within one month, extendable by two further months for complex or numerous requests. Under [Article 33](https://dsgvo-gesetz.de/art-33-dsgvo/), a personal data breach must be notified to the supervisory authority where feasible within 72 hours, unless it is unlikely to result in a risk to individuals. An agent can recognise incoming requests, set the deadline, gather where the data sits and draft a reply. It can also compare system changes with the records of processing under [Article 30](https://dsgvo-gesetz.de/art-30-dsgvo/) and report missing entries. Whether a breach is notifiable remains a human judgement, ideally with your data protection officer. For the agent's own data protection footprint, see [GDPR and AI agents](https://www.ruebstudio.de/en/guides/ai-law-and-compliance/gdpr-ai-agents-business).

## Where is the legal line? Practising law and the hand-off to counsel

Most jurisdictions restrict who may give legal advice to others, but they draw the line differently. The key question is usually whether the agent works on the company's own matters or advises third parties on theirs. An internal agent reviewing your own contracts is in a different position from a tool that answers customers' legal questions.

**Germany as an example.** The Legal Services Act (Rechtsdienstleistungsgesetz, RDG) defines a legal service in [Section 2(1)](https://www.gesetze-im-internet.de/rdg/__2.html) as any activity in a specific matter of another party that requires legal examination of the individual case. [Section 3](https://www.gesetze-im-internet.de/rdg/__3.html) prohibits providing such services independently without authorisation. A company's own contracts are not another party's matter, and Section 2(3) no. 6 expressly excludes handling legal matters within affiliated group companies. In 2021 the Federal Court of Justice held in [case I ZR 113/20](https://www.bundesgerichtshof.de/SharedDocs/Pressemitteilungen/DE/2021/2021171.html) that a question-and-answer contract generator was not a legal service, because it is built for typical situations and users do not expect a review of their specific case. That does not make every AI agent safe: one that assesses individual cases is closer to legal examination than a template is.

**The UK as a contrast.** England and Wales take a different approach. The [Legal Services Act 2007](https://www.legislation.gov.uk/ukpga/2007/29/section/12) reserves six activities, including rights of audience, conducting litigation and reserved instrument activities, rather than legal advice as a whole. Other jurisdictions draw the line in their own way. Before an agent answers anyone outside the company, ask counsel.

A dedicated guide on what an internal legal agent may do under German law is in preparation.

### What does a good hand-off to counsel look like?

Counsel is not replaced but better briefed. An agent can prepare a hand-off so that the lawyer starts at the substance:

1. The matter in two sentences: counterparty, contract type, value and urgency.
2. Deviations from the playbook with clause references.
3. The specific question to be answered.
4. Relevant correspondence attached, stripped of unnecessary personal data.

Decide escalation triggers in advance: liability above an internally agreed threshold, governing law outside your home region, exports to sensitive destinations, or any case where the agent finds no matching playbook rule.

## What do you need before you start?

Three things must exist: written rules, findable documents and a named person who approves. Without a playbook the agent checks against nothing. Without a repository it searches in the dark. Without approval nobody owns the result.

**Data.** Template contracts and your standard terms, a playbook with accepted and rejected positions, a library of approved questionnaire answers with evidence, your records of processing, and product datasheets for export screening. Much of this exists in scattered form; collecting it is often the biggest piece of work and useful on its own.

**Systems.** A central document store, a calendar or ticket system for deadlines and, where relevant, access to sanctions lists. Fewer manual hand-overs mean fewer places for errors.

**Organisation.** A person with legal responsibility who maintains the playbook and signs off, plus an agreed escalation path to outside counsel. You also need to decide where the model runs. Contracts, HR data and technical datasheets are confidential; depending on sensitivity, a model hosted in a suitable region under a data processing agreement or a [self-hosted model](https://www.ruebstudio.de/en/sovereign-ai) may fit. How to protect know-how in that setting is covered in [trade secrets and AI](https://www.ruebstudio.de/en/guides/ai-law-and-compliance/trade-secrets-ai-protection).

**Rules for the agent itself.** A company using AI in the EU is usually a deployer under the AI Act, with its own obligations such as AI literacy for staff. The overview [EU AI Act for mid-sized companies](https://www.ruebstudio.de/en/guides/ai-law-and-compliance/eu-ai-act-obligations-deployers) explains what that means in practice.

## A step-by-step approach

Start with one narrow task, measure how often a person has to correct the agent, and only then expand. A sensible order is: write the rules, collect test cases, build the agent, run it in shadow mode alongside a human, then go live with sign-off.

1. **Pick a task.** Good candidates repeat often and have clear inputs: questionnaires, supplier terms, deadlines in framework agreements.
2. **Write the playbook.** One to three pages is enough to begin: topic, acceptable position, unacceptable position, fallback wording, escalation.
3. **Collect test cases.** Ten to twenty real, closed cases where you know the right answer. The agent is measured against these.
4. **Build and restrict.** Read-only access to the folders it needs; write access only for drafts and calendar entries.
5. **Shadow mode.** For a few weeks, a person and the agent review in parallel. Disagreements lead to playbook additions, not prompt tinkering.
6. **Go live with sign-off.** The agent proposes, a person approves, and the log shows who decided what.
7. **Expand.** Add the next task only once the first is stable, ideally reusing the same knowledge base.

## Limits and risks

A compliance agent can be wrong, and convincingly so. The main risks are misread clauses, deadlines missed in annexes, an outdated playbook, and creeping trust that turns approval into a rubber stamp. Data protection and confidentiality in model hosting come on top.

- **Invented law.** A language model can make up statutes or rulings. That is why the agent checks against your playbook and cites locations in the document instead of giving legal answers from memory.
- **Missed content.** Scanned annexes, tables embedded as images or handwritten changes are easily skipped. The agent should report which parts it could not read.
- **Automated decisions about people.** [Article 22 GDPR](https://dsgvo-gesetz.de/art-22-dsgvo/) gives individuals the right, in principle, not to be subject to a decision based solely on automated processing that has legal or similarly significant effects. An agent deciding about employees or applicants does not belong in this setup.
- **Stale knowledge.** Control lists, sanctions lists and laws change. BAFA notes that Annex I of the Dual-Use Regulation is updated by delegated regulation, so sources must be queried live or refreshed regularly.
- **Rubber-stamping.** When the agent is usually right, people start clicking through. Spot checks and a duty to justify overrides keep the review real.

When does this approach not fit? For rare, highly individual matters such as acquisitions, shareholder disputes or ongoing litigation. There are no patterns for an agent to learn from, and the cost of a mistake is high.

## How do you get started?

Take the task that annoyed you most this month, perhaps the third security questionnaire or the next set of supplier terms, and write a one-page playbook for it. Collect ten closed cases as test material. That is enough to find out quickly whether an agent can reliably do the first pass.

Useful questions for a first team discussion:

- Who owns contracts, deadlines, questionnaires, export control and data protection today, and who signs off?
- Which rules are written down, and which live only in someone's head?
- Which documents may go into which environment, and where must the model run as a result?
- When is outside counsel involved, and in what format do they want the material?

For how agents work in other departments and connect to existing systems, see the overview of [AI agents](https://www.ruebstudio.de/en/ai-agents). For compliance in particular, the agent is only as good as the playbook it checks against, and only as safe as the sign-off that follows.

## Frequently asked questions

### Is an internal legal agent practising law?

It depends on the jurisdiction and on whose matter it handles. Working on the company's own contracts and obligations is generally treated differently from advising third parties, but check the rules where you operate.

### Can an AI agent replace outside counsel?

No. It prepares, sorts and flags deviations so that counsel receives a clean brief. Binding opinions, negotiation strategy and disputes remain with qualified lawyers.

### What is a good first use case?

Recurring questionnaires or first-pass review of incoming terms against a short playbook. Both have clear inputs, a checkable output and someone who signs off anyway.

### Where should a compliance agent run?

Contracts and HR data are confidential. Depending on sensitivity, use a model hosted in a suitable region under a data processing agreement, or a self-hosted model.

## Sources

- [German Legal Services Act (RDG), Section 2 – Definition of legal services](https://www.gesetze-im-internet.de/rdg/__2.html)
- [German Legal Services Act (RDG), Section 3 – Authorisation to provide out-of-court legal services](https://www.gesetze-im-internet.de/rdg/__3.html)
- [German Federal Court of Justice, press release 171/2021 on a contract document generator (I ZR 113/20)](https://www.bundesgerichtshof.de/SharedDocs/Pressemitteilungen/DE/2021/2021171.html)
- [UK Legal Services Act 2007, section 12 – Reserved legal activities](https://www.legislation.gov.uk/ukpga/2007/29/section/12)
- [GDPR Article 12 – Transparent information and time limits](https://dsgvo-gesetz.de/art-12-dsgvo/)
- [GDPR Article 22 – Automated individual decision-making](https://dsgvo-gesetz.de/art-22-dsgvo/)
- [GDPR Article 30 – Records of processing activities](https://dsgvo-gesetz.de/art-30-dsgvo/)
- [GDPR Article 33 – Notification of a personal data breach](https://dsgvo-gesetz.de/art-33-dsgvo/)
- [BAFA (German export control authority) – Control lists](https://www.bafa.de/DE/Aussenwirtschaft/Ausfuhrkontrolle/Gueterlisten/gueterlisten_node.html)
