Generative AI answers a prompt. An AI agent can decide what to do next.
That distinction matters for law firms because legal work is rarely one prompt. A real workflow might involve gathering matter context, finding precedent, extracting terms, drafting an output, checking it against a policy, routing it for review, recording the result, and following up when information is missing.
When software can coordinate more of those steps, the question changes. It is no longer only “How much faster can a lawyer draft this?” It becomes “Which parts of the delivery model still require human effort, and what is the service worth?”
That is an operating-model question, not a product demo.
Assistant, automation, or agent?
The word agent is applied loosely. Firms need a simpler vocabulary.
Assistant
An assistant responds to a user. It summarizes, drafts, retrieves, or answers, but the person decides each next step.
Automation
Automation follows a predefined sequence. If a condition occurs, the system performs a known action. It is reliable when the rules and inputs are stable.
Agent
An agent receives a goal, chooses among available actions, uses tools, evaluates intermediate results, and continues until it reaches a stopping condition or asks for help.
The categories can overlap. The useful question is how much discretion the system has and what it can affect.
An assistant that drafts a paragraph carries less operational risk than an agent that reads client files, selects authorities, edits a document, sends an email, and records time. The second system needs permissions, boundaries, review, and evidence at every step.
The adoption curve is steep and uncertain
Gartner predicts that up to 40% of enterprise applications will include task-specific agents by the end of 2026, up from less than 5% in 2025. Its best-case projection has agentic AI driving roughly 30% of enterprise application software revenue by 2035.
The same research firm also expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing cost, unclear value, and inadequate risk controls.
Both predictions can be true. Agents can become common inside software while many individual projects fail.
That is a useful warning for law firms. Buying an agent feature is easy. Redesigning a workflow, proving its quality, controlling its data, and changing the pricing model are the expensive parts.
The failure mode is not that the agent does nothing. It is that the agent does enough to change the economics without the firm changing the system around it.
The pyramid problem
Traditional law-firm economics rely on leverage. Partners supervise teams of associates and staff. Junior work creates billable volume, supports training, and contributes to the margin beneath partner rates.
Agents are aimed directly at multi-step work that often sits in that leveraged layer:
- first-pass research
- document review and extraction
- diligence checklists
- chronology building
- precedent assembly
- routine drafting
- filing and status workflows
- billing compliance review
The question is not whether every task disappears. Most will still need matter context, exception handling, professional judgment, and review. The question is how much human effort remains and where it sits in the team.
If routine associate hours fall, an hourly firm loses billable volume. If the work stays priced by outcome, the firm may gain margin. If junior lawyers no longer learn by doing the first pass, the firm also needs a new training model.
The technology, pricing, staffing, and talent pipeline are one problem.
Pick workflows, not job titles
“Automate the associate” is not a usable implementation plan.
Start with a bounded workflow that has:
- a repeatable trigger
- known source systems
- a defined output
- clear permissions
- an objective review checklist
- a responsible owner
- a safe failure state
An invoice compliance check is a better first agent than an autonomous litigation strategist. The guideline is known, the invoice is structured, exceptions can be flagged, and a human can approve the result.
A contract intake workflow can also be bounded. The agent can confirm that required documents exist, extract a defined set of fields, compare them with approved playbook positions, and route exceptions. It should not silently negotiate terms or send a final response without approval.
The narrow workflow creates data the firm can use to decide whether more autonomy is justified.
Build the governance counterweight
An agent can take action across systems. Governance needs to match that ability.
Identity and permissions
Give the agent its own identity. Do not let it act through a shared administrator account. Limit tools and data to the minimum required for the workflow.
Matter boundaries
Every retrieval and action should be scoped to the authorized client and matter. The system should fail closed when context is missing or ambiguous.
Human checkpoints
Define which actions require approval. High-impact steps such as filing, sending client communications, changing a legal document, or posting a charge should not be treated like internal summarization.
Evidence
Record the source data, tool calls, outputs, decisions, approvals, and final action. A log should let a reviewer reconstruct what happened.
Cost limits
Set ceilings on model use, queries, external services, and repeated attempts. An agent can consume resources faster than a person notices.
Stop conditions
Define when the agent must stop and escalate. Missing authority, conflicting instructions, low confidence, a permission error, or an unexpected data classification should not trigger improvisation.
Evaluation
Test the workflow against representative cases, known exceptions, adversarial inputs, and failure scenarios. Accuracy on a clean demo is not enough.
ABA Formal Opinion 512 remains relevant even when the system is more autonomous. A lawyer’s duties around competence, confidentiality, communication, supervision, candor, and fees do not move to the agent.
Measure the full economics
Agent ROI is often calculated as human hours avoided. That misses important costs:
- workflow design
- integration and data cleanup
- vendor and security review
- model and tool usage
- human review
- monitoring and incident response
- exception handling
- maintenance when systems change
- training and change management
Measure cost, quality, speed, and capacity together.
A workflow that is 70% faster but requires extensive correction may not be useful. A workflow that is only 20% faster but produces consistent intake data and fewer missed requirements may be valuable. The answer depends on the service and risk.
Then connect the result to pricing. If the firm sells time, less time may reduce revenue. If it sells a defined outcome, lower delivery cost may improve margin. If the client expects AI savings to be shared, the commercial model needs to reflect that before the invoice.
Redesign training with the workflow
Junior work has never been only production. It is also how lawyers learn issue spotting, judgment, drafting, and client context.
If agents perform more first-pass work, firms need deliberate alternatives:
- require juniors to review and correct agent output
- compare agent results with independent analysis
- rotate lawyers through workflow design and evaluation
- teach source verification and failure analysis
- preserve direct client exposure
- measure judgment, not only billable volume
An agent can remove repetitive effort without removing the learning loop, but only if the firm designs a new one.
A staged adoption path
Stage 1: observe
Map one workflow and measure its current cost, time, quality, and exceptions.
Stage 2: assist
Use AI for a first pass while a person controls every step.
Stage 3: automate
Automate stable transitions and checks with deterministic rules.
Stage 4: delegate a bounded goal
Allow an agent to choose among approved actions within strict permissions and human checkpoints.
Stage 5: expand carefully
Increase autonomy only after the evidence shows reliable performance, manageable exceptions, and real business value.
The firms that benefit from agents will not be the ones that grant the most autonomy first. They will be the ones that can absorb faster delivery without losing control, client trust, or the economics of the service.
Need to design the workflow and the guardrails together?
Jinka helps law firms build matter-aware agents with bounded permissions, human review, audit trails, cost controls, and a commercial model that fits the work.
Sources
- Gartner: 40% of enterprise apps will feature task-specific AI agents by 2026
- Gartner: More than 40% of agentic AI projects will be canceled by the end of 2027
- American Bar Association: Formal Opinion 512
- Association of Corporate Counsel: GenAI transparency and readiness questions for outside counsel