Legal AI is usually described by what it produces: a summary, a draft, a table of clauses. An agentic workflow is better described by what it does between the request and the result, because that is where the risk sits. These notes are written from a lab's point of view, about corporate and legal work, and they name the steps rather than the products.
An agentic workflow takes a goal, splits it into steps, uses tools, checks the results, and stops for a person to approve. Every step is recorded.
What an agentic workflow is
A chat assistant answers a question and stops. An agentic workflow takes a goal, splits it into steps, calls tools for each step, checks what came back and decides what to do next. The unit of work is no longer a reply. It is a run: a sequence of actions with an input, an owner and an end state.
That is the line between AI-assisted and AI-native. In an AI-assisted practice a person uses a tool and the tool stays outside the process. In an AI-native practice the process itself is built around agents, and people are placed at defined points inside it. The same distinction applies to a company's legal department as much as to a law firm using AI agents.
Where a person approves
A useful workflow is drawn with its approval points before it is drawn with its agents. The test is simple: for every step that leaves the system, such as a message to a counterparty, a filing, a payment or a change to a record someone else relies on, there is a named person whose decision releases it.
An agent that approves its own output has not been reviewed. A second agent that approves the first is a check, and worth having, but it is not a sign-off. We keep the three apart in the vocabulary: check, review, approval. Only the last one belongs to a human.
What gets logged
Each run leaves a record that a person who was not there can read: the goal as given, the sources consulted, each tool call, each intermediate result, each point where a person was asked and what they answered. A log that only keeps the final text cannot answer the question a reviewer actually asks, which is how the system reached it.
The log is also what makes AI governance for law firms and for corporate legal teams a practical matter. Disclosure, supervision and retention duties are easier to meet when the record exists by default than when someone reconstructs it afterwards.
How workflows fail
The failures we plan for are ordinary ones. A source that was not read is treated as if it said nothing. A document is matched to the wrong entity because two names are close. A step that should have stopped for approval is skipped because the run was configured with a default that allowed it. A tool returns an empty result and the agent reports success.
The common remedy is to make the unknown state explicit. A run that could not verify something ends as unverified, with the reason, rather than as a confident answer. An empty screen with an explanation is a better result than a plausible one with nothing behind it.
A reading of the professional rules
In Israel, the practical questions about legal AI sit between three sources: the Israel Bar's guidance on lawyers' use of AI, the privacy and data-security rules that apply to client information, and the contract terms with the vendor that processes it. This note does not interpret them. It points to where a workflow has to be mapped against each, and leaves the legal conclusion to a lawyer.
Questions and answers
- What is an agentic workflow in legal work?
- It is a process in which an AI system takes a goal, breaks it into steps, uses tools for each step and checks the results, with a person approving the steps that leave the system. The unit of work is a recorded run, not a single reply.
- What is the difference between AI-assisted and AI-native?
- In an AI-assisted practice, people use AI tools alongside an unchanged process. In an AI-native practice the process is designed around agents from the start, with human decision points placed inside it.
- Where must a person approve?
- At every step that leaves the system or changes a record others rely on: a message to a counterparty, a filing, a payment. A person's decision releases it, and an agent's own approval does not count as one.
- What should an agentic workflow log?
- The goal as given, the sources consulted, each tool call and result, and every point where a person was asked, with their answer. The log should let someone who was not present see how the output was reached.
The same subject from a law firm's side, in the firm's own series: AI Briefings at neemanlaw.com.