If GenAI was the warm-up act, AI agents are the main event — redefining the core of legal work. We are entering an era where intelligent systems do not just generate content but plan, execute, and improve workflows on their own. But the conversation needs sharper vocabulary. “Agent” is not one thing. It is four canonical tiers, each with a distinct autonomy profile, distinct risk surface, and distinct governance burden. Without that classification, every “agentic” conversation collapses into hype.
Why the distinction matters now
The 2025 Thomson Reuters Future of Professionals Report: 80%+ of legal professionals expect AI to transform their work inside five years. Only 22% of organisations have a clear AI strategy. Those with one are 3.5× more likely to see measurable return. “Clear strategy” is no longer satisfied by “we are using AI.” The question now is at which tier, in which workflow, with what controls. A function that cannot answer those three questions has the appearance of an AI strategy, not the substance.
The four canonical tiers
Tier 1 — Augmentation
The AI assists a human inside a single step. A lawyer asks for a clause summary, a redline, a research extract; the lawyer reviews and decides. Autonomy: low. The lawyer remains the system of record. Most legal-AI usage in 2025 sits here. Defensibility posture: established — the human is the control. Module governance: standard prompt logging, output review.
Tier 2 — Co-pilot
The AI proposes multi-step actions; the human approves each step before execution. Drafting suites that propose paragraph-by-paragraph edits, intake systems that suggest routing and require approval, research tools that draft a memo then ask for refinement. Autonomy: medium. The control is still per-step. Defensibility posture: manageable — every action is gated. Module governance: workflow logging, escalation triggers when the AI proposes deviation from template.
Tier 3 — Workflow operator
The AI executes multi-step workflows end-to-end within pre-defined bounds; the human reviews the result, not each step. Contract redlining inside guardrails, intake triage with auto-routing, regulatory scans that pre-classify alerts. Autonomy: high — but bounded. Defensibility posture: demanding. The function must hold an explicit delegation-authority register (who authorised this AI to act in this workflow, under what bounds, with what fallback). Risk Taxonomy 2026 surfaces action irreversibility and reduced supervisory capacity at this tier.
Tier 4 — Autonomous agent
The AI plans goals, decides what work to do, executes across multiple systems, and selects when to escalate. The human reviews exception cases and policy boundaries, not workflows. Few production deployments exist in legal in 2025. The frontier. Defensibility posture: aspirational. Adds cascade failure and confabulated execution to the Risk Taxonomy. Requires an organisational AgentOps capability that does not yet exist in most legal functions. Not a place to be first.
GenAI vs Agentic: the underlying difference
GenAI (Tier 1–2 usage) creates content in response to prompts. Agentic AI (Tier 3–4) executes multi-step processes to achieve goals. GenAI needs prompting and fact-checking; Agentic needs goals, oversight, and validation gates. The most effective teams combine: GenAI to create, Agentic AI to execute. The trap is treating both as interchangeable — most failed AI deployments confuse a Tier 1 tool with a Tier 3 use case (or vice versa).
Where to start: five use cases, mapped by tier
- Matter intake & triage — Tier 3 candidate. Auto-classify and route requests against predefined rules; escalate exceptions to a human queue.
- Contract drafting & redlining — Tier 2 today; Tier 3 for low-complexity templates with clear deviation rules.
- Regulatory monitoring — Tier 2/3. Scan, classify, surface; humans triage and decide on response.
- Knowledge management — Tier 1–2. Auto-surface relevant playbooks in context; the lawyer still authors.
- Client onboarding — Tier 2/3. Collect, verify, organise data across systems with human approval before downstream commits.
Rule: start at the lowest tier that solves the workflow. Do not jump tiers without standing up the governance and the delegation-authority register first. Tier inflation is the failure mode.
AgentOps: the discipline that gates the tier jump
AgentOps — the operational discipline of designing, governing, and improving AI agents — is what separates a function that runs Tier 3 from a function that claims to run Tier 3. It answers three questions per agent: who is accountable for autonomous decisions; how are outputs validated and audited; how is success measured. Without those three answers, the function is not running an agent. It is running an unsupervised intern with system access.
The shift: tier discipline is Defensibility
The agentic conversation will get louder through 2026. The vocabulary that survives will not be “we deployed AI agents.” It will be “we deployed Tier 2 co-pilots in matter intake with the following delegation-authority register; we are testing Tier 3 in template redlining inside the following bounds; we do not yet operate Tier 4.” That is the Defensibility Posture the board will recognise and the regulator will accept.
Share this issue