Most lawyers have tried some form of AI tool over the past two years. Summarising documents, explaining clauses, speeding up research. That was the first wave. What is taking shape in 2026 goes a step further: AI that does not just answer, but acts. Workflows where the system takes on a task, makes decisions, and initiates follow-on steps - inside a legal document or process.
That is called agentic AI. And it is no longer a future topic.
What agentic AI actually means
Classic AI tools work reactively: you put something in, you get something back. Agentic AI is structured differently. An AI agent receives a goal - not just a question - and works through that goal across multiple steps. It can evaluate intermediate results, build on them, and make decisions within the process itself.
In a legal context, that means: instead of "explain clause 8.3 to me", a task can read "review this contract for liability risks, flag critical passages, and suggest alternative wording". The system runs the analysis, assesses severity, identifies relevant passages, and prepares concrete edits - in a single pass.
The difference from previous tools is not gradual but structural. It shifts from assistant to executor.
Why 2026 is the turning point
According to a widely cited Harvard publication from March 2026, "the era of the AI pilot is over". Firms and legal teams that have used AI in isolated experiments now face the question of how to integrate these capabilities into real work processes. Pressure is coming from several directions at once.
Clients today expect concrete efficiency gains, not technology demonstrations. Firms working with AI are expected to show that in lower costs, faster turnaround, or better outcomes - not in slide decks. At the same time, new providers are entering the market that are built around agentic workflows from the start.
For Swiss law firms and in-house teams, another factor is relevant: data protection. Agentic workflows process sensitive contract data, sometimes at scale. Where that data is processed and who has access to it is not a trivial question under Swiss regulation or client mandate.
The most relevant agentic workflows in legal practice
Contract analysis with automatic prioritisation
One of the most practically relevant use cases is automated contract risk analysis. Traditionally: a lawyer reads the contract, marks issues, assesses them, and drafts improvements. Agentically: the system identifies risks from each party's perspective, assigns a severity to each finding (low, medium, high), and delivers ready-to-use drafting options - which can be applied directly in the document without copy-paste.
CASUS, a Swiss legal AI platform, takes exactly this approach in its Risk & Quality Review. The module identifies contract parties, analyses risks from the respective party's perspective, and presents findings in a structured format with assignment, relevance, and severity.
Checking against playbooks and standards
Many law firms and legal departments work with internal standards: template SPAs, NDA frameworks, data protection clauses benchmarked to their own requirements. Agentic workflows can check a document automatically against that standard - not just for the presence of clauses, but for their completeness and deviation.
The Benchmark module at CASUS checks whether standard clauses are present and sufficiently detailed, flags missing topic areas, and outputs a match percentage. Knowing that a contract aligns 64% with your internal standard lets you focus on the gaps directly - rather than re-reading the entire document from the top.
Parallel analysis of large document sets
Due diligence and compliance reviews do not fail for lack of expertise - they fail because of volume. Dozens or hundreds of contracts need checking for specific clauses, deadlines, or risk factors. An agentic workflow can parallelise that analysis.
The AI Data Room at CASUS allows uploading many documents simultaneously. Extraction fields are defined by prompt - which clauses, which parameters, which deviations should be captured. The output is a structured table that translates directly into Excel or a due diligence template. Anomalies like missing liability caps or unusually long notice periods are flagged and prioritised by risk.
Source-based legal research
Agentic legal research workflows connect research and document work. Instead of switching between a legal database and a word processor, research runs directly in the context of the task - and the results feed immediately into clause rationales, internal assessments, or text changes.
CASUS offers a Legal Research mode inside its AI Chat, drawing on over 660,000 cantonal and federal court decisions as well as statutory law. Relevant reasoning sections are highlighted directly in the results; a preview of decision content appears inline, without any click-through. That enables structured first assessments with traceable sources - directly usable for further document work.
What agentic workflows mean in practice for Swiss law firms
The efficiency gain is real, but it also requires a redistribution of tasks. Using an agentic workflow means more than delegating an activity - it also shifts the control function. The system executes; the lawyer reviews, decides, and bears responsibility.
That is not a weakness of the model but its nature. Agentic systems are designed to handle routine work and prepare outputs - not to replace legal judgment. The question is which tasks are genuinely delegable and which are not.
In practical terms: contract analysis, benchmarking, data point extraction, and structured research are well suited to automation. Strategic negotiation, mandate-specific risk assessment, and client communication remain human tasks.
Data protection as a Swiss-specific requirement
Agentic workflows often process complete contract texts, sometimes including personal data, trade secrets, or mandate-sensitive content. For Swiss law firms, the question of data processing is not optional.
CASUS processes data exclusively in Switzerland and the EU, transfers no data to the US, and operates with neither human review nor session-based data retention (zero data retention). That architecture is not a marketing statement - it is a prerequisite for deployment on many Swiss mandates, particularly in M&A, banking, and regulatory work.
Further detail on the security architecture is available on the CASUS security page.
Where the limits are
Agentic AI produces structured, source-based, traceable outputs. What it does not produce is certainty. No system guarantees complete results or replaces legal responsibility.
In practice: extraction results from the AI Data Room are based on defined fields and prompts, not automatic completeness. Legal research outputs are structured and source-linked, but not final legal opinions. Risk review findings are prioritised by severity, but the decision of which risk is negotiable rests with the lawyer.
That is not a disadvantage compared to the status quo - it is a transparent description of what these tools can and cannot do.
Try CASUS
Anyone who wants to test agentic legal workflows in their own practice can use CASUS directly in the browser or as a Word add-in - with no US data transfer, no human review, and no data retained after a session. A free trial is available at app.getcasus.com/signup.
FAQ
What is agentic AI in a legal context?
Agentic AI refers to AI systems that do not simply respond to inputs but execute multiple steps toward a defined goal - for example, analysing a contract, prioritising risks, and preparing improvement suggestions without requiring a separate prompt for each step.
How does agentic AI differ from conventional legal tech tools?
Conventional tools answer questions or perform individual functions. Agentic workflows connect multiple steps into a single process: analysis, assessment, decision preparation, and document editing run sequentially or in parallel.
Which legal tasks are well suited to agentic workflows?
Strong fits include contract analysis, playbook benchmarking, clause extraction across large document sets, structured legal research, and formal document review. Less suited are strategic decisions, mandate-specific risk balancing, and direct client communication.
How secure are agentic AI workflows when handling Swiss contract data?
It depends heavily on the provider. Key factors are: where data is hosted (Switzerland/EU vs. the US), whether data is stored after a session, and whether human reviewers can access content. CASUS hosts in Switzerland/EU, retains no data after a session, and has no human review.
Does agentic AI replace legal review by a lawyer?
No. Agentic systems handle structured, repetitive tasks and prepare outputs. Legal responsibility, judgment, and final decision-making remain with the lawyer.
What role does the EU AI Act play for agentic legal workflows?
The EU AI Act classifies certain AI applications as high-risk, including those that influence legal decisions. For Swiss firms handling EU mandates or working with EU-based companies, this is relevant. Providers should offer clear transparency about how their systems work and where their limits are.
What does an agentic workflow look like in practice?
An example: an NDA is uploaded. The system checks it against an internal playbook, identifies missing clauses (e.g. no deletion obligation, no carve-out for regulatory requests), assesses the severity of each finding, and suggests concrete additions - which can be inserted directly in the Word document at the right location.
What does it cost to get started with agentic legal AI workflows?
It varies by platform. CASUS offers a free trial, so the fit with your own practice can be tested without upfront investment.
