CCCasestack Cloud
Guides › Legal AI

What Can Agentic Legal AI Do Without a Lawyer Watching? Less Than the Demos Suggest

Agents can build a chronology or a first-pass diligence table with little supervision. Most other multi-step legal work still needs checkpoints, and your contract should say who checks it and how.

Casestack Cloud editorial team · November 2026 · 9 min read

Which Steps Can Run Without a Lawyer in the Loop

Every vendor pitching agentic AI for law firms shows roughly the same demo. You give a goal, the system plans the steps, opens the documents, pulls the facts and returns finished work product. The buyer's question isn't whether the demo is real. It's which of those steps you'd let run on a live matter with nobody checking until the end.

For most firms and in-house teams, the honest answer is a short list. Legal AI agents are dependable where the inputs are fixed, the output can be checked against a source, and nothing leaves the building. Once a task involves legal judgment, a document going to an outside party, or a write to a system of record, autonomy has to give way to checkpoints.

Advertisement

Below, multi-step legal AI tasks are sorted into those two groups, with an explanation of why demos blur the line and a list of the audit-log, permission and supervision controls to get into the contract before you sign.

Chronologies, Diligence Tables and Comparisons: Where Agents Earn Their Keep

Three task types hold up well with light supervision. They share a structure: a closed set of documents goes in, structured output comes out, and every cell or entry points back to a page.

Chronologies from a defined document set

Give an agent thousands of emails, a set of board minutes and a deposition transcript, and it can pull dated events, tag the people involved, link each entry to its source and flag conflicting dates. The output is a draft fact base, not a finding. A paralegal can spot-check entries against the citations far faster than building the chronology from scratch.

First-pass diligence tables

Pulling change-of-control, assignment, exclusivity, termination and governing-law clauses from a data room into a table is routine work for an agent. It works because the schema is fixed in advance and every answer must cite a clause. Reliability drops with poor scans, handwritten amendments, and side letters that change a main agreement without saying so. The table is still useful, but a reviewer should sign off on the exceptions column.

Document set comparisons

Comparing a counterparty's draft against your playbook, checking a batch of NDAs for deviations from your standard, or matching a final agreement to the approved version are comparison jobs. The agent isn't deciding what the law requires. It's reporting differences between texts, which is the work these systems do most consistently.

Rule of thumb: if you can audit the output by clicking citations instead of redoing the analysis, the task is a good candidate for light supervision.

Tasks That Still Need Checkpoints, and Where to Put Them

The risk changes once an agent moves from extracting facts to reaching conclusions, or from reading to acting. Use the table below to set autonomy levels in a pilot.

TaskSuggested autonomyRequired checkpoint
Chronology from a closed document setHighSample-check citations before relying on it
First-pass diligence extractionHighReviewer signs off on exceptions and low-confidence fields
Playbook comparison and deviation reportHighLawyer reviews flagged deviations
Research memo with conclusionsLowLawyer checks every authority cited and every conclusion drawn
Redlines or emails to a counterpartyNoneHuman approval before anything is sent
Privilege and responsiveness callsLowLawyer reviews samples and every withheld document
Writes to DMS, billing, matter or e-filing systemsNoneHuman approval for each action

The research memo catches buyers out. An agent that searches, reads, synthesizes and writes is chaining four steps, and an early error, such as a misread holding or a superseded statute, carries through everything after it. Grounded research tools have reduced invented citations, but they haven't removed the judgment call on whether an authority actually supports the point.

Outbound actions get the hardest line for a simple reason. A bad chronology entry costs review time. A redline sent to the other side, or a filing made in the wrong court, can't be taken back.

Why the Demo Looks Better Than Your Matter Will

Demos aren't dishonest. They run under favorable conditions: clean native files, a modest document count, a task the vendor has tuned for, and one good run picked from several. Your matters arrive as scanned PDFs, foreign-language exhibits, email threads with 30 replies, and folders nobody has cleaned up.

Small errors also compound. Here's a hypothetical: an agent runs a ten-step workflow, and each step is right 95% of the time. If errors are independent, the chance that all ten come out right is about 60%. Each step looks fine alone, yet the end-to-end result isn't something you'd send to a client unchecked. That's why checkpoints belong between steps, not just at the end.

Ask any vendor three questions:

  • Can we run the pilot on our own documents, including the messy ones, under our confidentiality terms?
  • What does the agent do when it isn't sure: stop and ask, flag it, or carry on?
  • Can we replay a run and see every tool call, document opened and intermediate output?

If the answer to the third is no, you can't supervise the tool, however good the output looks.

Audit-Log, Permission and Supervision Terms to Put in the Contract

Human-in-the-loop legal AI only works if the product enforces it. Policy documents don't. Require these controls in the order form or a schedule to the master agreement, not in marketing material that can change without notice.

Audit logs

  • A step-by-step record of every agent run: prompt, plan, documents accessed, tools called, intermediate outputs, final output, user and timestamp.
  • Logs you can export in a standard format, with a retention period that matches your file-retention policy. Many firms need several years or longer, depending on jurisdiction and matter type.
  • Logs that can't be edited, including by vendor staff.

Permissions

  • Agents that inherit the user's existing matter-level access and ethical walls from your DMS, rather than running under a broader service account.
  • Tool access that's off by default. Sending email, writing to the DMS, or calling external systems should each need an administrator to switch them on.
  • Approval gates you can configure per workflow, so the firm can require sign-off before any outbound or write action.

Supervision and data terms

  • No training on your data or prompts without a separate written opt-in.
  • Named sub-processors, data-residency options, and notice before either changes.
  • Notice before model or workflow changes that could alter outputs, so you can re-validate the workflows you depend on.
  • Breach-notification timelines that satisfy your strictest applicable regime.

Red flag: a vendor that offers logs only "on request through support," or can't confirm that agents respect ethical walls, isn't ready for unsupervised work on client matters.

Supervision Duties Differ by Country, but the Principle Is the Same

Most legal systems expect lawyers to supervise work that goes out in their name, whether a junior, a contractor or software produced it. The wording varies. In the United States, bar guidance builds on the duties of competence and supervision of nonlawyer assistance. In England and Wales, the SRA's standards place responsibility for competent service and supervision on the firm and the individual solicitor. Regulators and professional bodies in India, Singapore, the Gulf states, Brazil, South Africa and elsewhere apply their own versions, and some have issued or are developing AI-specific guidance.

Data rules matter too. The EU GDPR, the UK GDPR, India's Digital Personal Data Protection Act, Singapore's PDPA, Brazil's LGPD, South Africa's POPIA, and the frameworks in the DIFC and ADGM all affect where client data can be processed and which sub-processors are acceptable. The EU AI Act adds obligations for certain uses, and how they apply depends on the use case and your role under the Act.

This is a general description, not advice for your firm. Have your general counsel, risk partner or outside regulatory counsel map your obligations in each jurisdiction where you practice or hold client data, then check the vendor's terms against that map.

Cost Drivers and a Decision Rule for Your Pilot

Pricing for AI workflow automation for lawyers is still settling. Typical 2026 market ranges, which vary widely by contract, seat count and region: per-seat assistant plans with some agentic features often run roughly $80 to $300 per user per month, and enterprise platforms with configurable agents, integrations and admin controls commonly start in the low six figures a year for mid-size firms. Some vendors charge by usage, such as documents processed, runs or compute, which can push costs up quickly on large diligence or disclosure matters. Local pricing and bundling differ, so confirm with vendors directly.

The main cost drivers are DMS and matter-system integrations, data-residency requirements, log retention, single-tenant deployment, and the internal reviewer time each workflow still needs. Business cases often leave that last one out. They shouldn't.

Say a 60-lawyer firm pilots three workflows over 90 days: chronologies on two litigation matters, NDA playbook review, and a diligence table on one live deal. It measures reviewer time per output and the citation error rate, and it records every case where the agent should have stopped and didn't.

Decision rule: give a workflow more autonomy only after its outputs pass sampled review on your own documents across at least two matters, and only if the contract guarantees the logs and approval gates you'd need to show a regulator or client how the work was supervised. If either condition fails, keep the checkpoint.

Common questions

Can agentic legal AI replace associates on diligence?

It can take over most first-pass extraction and table-building, which used to fill a lot of junior hours. It can't replace the judgment needed to assess risk, read side letters in context or advise the client. Firms that adopt it tend to shift associate time toward reviewing exceptions and analysis rather than dropping the review step.

What does human in the loop legal AI mean in practice?

The workflow has defined points where a qualified person must review or approve before the agent continues. Typical checkpoints are signing off on low-confidence extractions, checking cited authorities, and approving any outbound or write action. The software should enforce them, not just a policy document.

Which audit-log features matter most when comparing legal AI agents?

Look for a complete step-by-step record of each run, covering documents accessed, tool calls and intermediate outputs, plus user identity and timestamps. Logs should be exportable, protected against editing, and kept as long as your file-retention rules require. Make sure the contract commits to these features rather than leaving them on a product roadmap.

Should agents be allowed to send emails or file documents?

For most firms, not without approval for each action. Outbound and filing actions can't be undone and carry professional-responsibility risk, so they should be off by default and enabled only behind an approval gate. Reading and drafting are far lower-risk places to start.

How do data-protection laws affect agentic AI deployments?

Agents often touch more documents than chat tools do, so where data is processed and who the sub-processors are matters more. Regimes such as the GDPR, UK GDPR, India's DPDP Act, Singapore's PDPA, Brazil's LGPD and South Africa's POPIA each have their own transfer and processing rules. Have qualified counsel review the vendor's data terms for every jurisdiction you work in.

How long should a pilot run before expanding autonomy?

Two to three months on live but lower-risk matters is a common range. That's long enough to meet messy documents, edge cases and likely at least one model update. Expand autonomy one workflow at a time, based on measured error rates and reviewer time rather than overall impressions.

Written by the Casestack Cloud editorial team. Prices, features and rules change often - confirm details with the vendor or a licensed professional before you decide.