A category definition. Why autonomous agents under human supervision represent a structural shift in how legal work gets delivered, and what this means for the enterprise legal leaders deciding whether to adopt.
There is a structural problem at the centre of every large in-house legal team, and it is not a technology problem. It is a staffing model problem. The majority of legal work in an enterprise does not require the expertise of the people doing it.
This is not a controversial observation. Any GC will tell you, usually with some exasperation, that their team spends most of its time on routine contracting: NDAs, vendor agreements, procurement reviews, first-pass redlines, intake triage. This is work that needs to be done correctly but does not require the judgment of someone who spent seven years in law school and practice. The mismatch is simple: inexpensive work is being done by expensive resources.
The problem has been understood for decades. The CLOC survey data shows outside counsel spend dropping from 58% to 37% of legal budgets over the past decade. But the work did not disappear. It moved in-house, where it now competes for the time of lawyers who were hired for strategic roles but spend their days processing contracts. The capacity squeeze that every GC describes is not a temporary condition. It is the predictable result of bringing high-volume work in-house without changing the delivery model.
The solutions attempted so far have each addressed a symptom rather than the structure. ALSPs outsource the overflow to cheaper labour, but the economics still scale linearly with volume. CLMs organise the queue, but every request still lands on a lawyer's desk. Copilots make individual lawyers faster, but every contract still requires a human to pick it up, process it, and send it back. None of these changes the fundamental model: legal capacity is bounded by the number of humans doing the work.
The copilot model is the dominant paradigm in legal AI as of 2026. A lawyer opens a tool, uploads a document, asks a question or requests a task, reviews the output, and then does the work themselves. The AI makes the lawyer faster. The lawyer is still the operator.
The evidence for copilot effectiveness is genuine. A widely cited Harvard/BCG study found that AI assistants deliver roughly 25% speed gains and 40% quality improvements on structured tasks. These are meaningful numbers. But they describe a ceiling, not a trajectory. If a lawyer reviews 10 contracts a day and a copilot makes them 25% faster, they now review 12.5 contracts a day. The backlog shrinks temporarily, then grows back as the business adds volume.
The more telling data point is the gap between expectation and reality. An ACC/Everlaw survey found that 64% of in-house teams expect to reduce outside counsel reliance through AI. But only 7% report actually seeing a reduction in total matter cost. That 57-point gap is not a failure of any particular product. It is the structural limit of a model that accelerates lawyers without removing them from the loop on work that does not require them.
The signals from outside legal suggest the copilot phase is transitional. Dario Amodei of Anthropic tracks Claude usage at 60% augmentation and 40% automation, with the automation share growing. McKinsey now counts 20,000 AI agents as part of its 60,000-strong workforce, targeting parity by the end of 2026. Sequoia observed in March 2026 that for every dollar spent on software, six are spent on services, and that the next phase of AI will compete for the services budget, not the technology budget. These are not legal-specific signals, but legal is not exempt from them.
If copilots represent the first phase of AI in legal, what does the second phase look like? The answer, we believe, is agentic services: providers that deliver completed legal work using autonomous agents under human supervision. Not faster lawyers, but a different staffing model entirely.
The term "agentic" is used loosely in 2026, and this looseness is a problem. If every chatbot is "agentic," the word means nothing. For the category to be useful to buyers and analysts, it needs a precise definition.
An agentic legal service is a provider, whether software, service, or hybrid, that delivers completed legal work using autonomous agents operating under human supervision. The key phrase is "completed legal work." The output is not a suggestion, a draft for the lawyer to review and rework, or a summarisation. The output is the finished product: the redlined contract, the drafted NDA, the triaged and routed request, the negotiation response. A human set the rules, the agent applied them, and the work is done.
This definition can be tested empirically. We propose four tests that distinguish agentic services from adjacent categories.
The system completes work from intake to resolution without requiring human approval at every step. The human defines the guardrails: which terms are acceptable, when to escalate, what risk tolerance applies. The agent operates within those guardrails. Some outputs require human review (exceptions, high-risk items). Most do not.
The system operates on the customer's specific templates, preferred terms, fallback positions, and escalation rules. This is not generic legal AI trained on a corpus of contracts. It is an agent that has been configured to work the way this particular legal team works, with this particular playbook, for this particular business.
A single agent performs multiple functions within a workflow. It does not just triage. It triages, selects the right template, drafts the document, applies jurisdiction-specific clauses, and routes the output. A point solution for each step is a toolkit, not an agent.
Supervision is not an afterthought or the customer's problem to solve. It is built into the service. Tenured legal practitioners review agent output, refine the underlying rules as patterns emerge, and close gaps between what the agent does and what the team's standards require. This is quality control delivered as a product feature.
These four tests are not arbitrary. They trace back to the structural problem: if you want to remove routine work from the human queue, you need a system that can execute autonomously (test 1), on your specific logic (test 2), across a complete workflow (test 3), with professional oversight (test 4). Fail any one test and you have something valuable but structurally different from an agentic service.
The agentic services category is not monolithic. Three structural forms are emerging, and the choice between them matters more than any feature comparison.
Agents are deployed inside the customer's environment. The customer's own playbooks, templates, and escalation rules govern the agents' behaviour. Supervision is handled by the customer's lawyers, a partner firm the customer selects, or the vendor's supervision team, depending on the customer's preference. The capability belongs to the customer. If they change vendors, the playbooks and institutional knowledge stay.
The platform model requires more from the customer upfront: playbook definition, supervision capacity, configuration time. In return, it builds internal capability that compounds over time. The customer's agents get better as their playbooks improve, and that improvement is an asset they own.
The vendor employs the agents and, typically, human lawyers. The customer sends work in and receives completed output. From the customer's perspective, this looks like hiring a very fast, very consistent ALSP. The vendor handles everything: the AI, the supervision, the quality control. The customer's team does not interact with the agents directly.
The service model requires the least change from the customer. No deployment, no configuration, no supervision capacity. The trade-off is dependency: the operational knowledge lives inside the vendor, and the customer builds less internal capability. Switching costs are higher because the vendor's AI learns the customer's patterns over time, and that learning is not portable.
A platform deployed inside the customer's environment, with the vendor's supervision team handling quality. The customer retains ownership of the playbooks and configuration. The vendor provides the professional oversight. This model combines the capability ownership of the platform model with the reduced supervision burden of the service model.
In practice, the boundaries between these models are blurring. Service-model providers are building client-facing platforms. Platform-model providers are offering supervision services. The market will likely converge on hybrid models that give customers a choice along the spectrum, rather than forcing a binary decision.
Agentic services were not possible two years ago and may not have been adopted even if they were. Three factors are converging in 2026 that make the category viable now.
Foundation models can now handle the legal reasoning required for routine contracting with acceptable accuracy. This does not mean they are perfect. It means they are good enough for the class of work where the guardrails are clear and the risk tolerance is defined. Two years ago, the error rate on complex clause analysis was too high for production use. Today, with structured prompting, validation chains, and multi-model verification, the accuracy profile supports deployment at scale.
SOC 2 Type II certifications, single-tenant deployment architectures, zero-data-retention agreements with LLM providers, EU data residency options, and comprehensive security questionnaire responses. The infrastructure for enterprise trust now exists. Two years ago, a GC who wanted to run contract data through an LLM had legitimate security objections that no vendor could satisfactorily address. Today, the security posture of a well-built agentic service is at least as strong as the security posture of an offshore ALSP.
A generation of legal operations leaders has spent a decade building processes, templates, standard terms, escalation matrices, and institutional playbooks. This work was motivated by process efficiency, not AI readiness. But it produced exactly the inputs that agents need: structured rules, documented preferences, explicit decision criteria. The organisations that invested in legal ops over the past decade are, perhaps inadvertently, the best-prepared to adopt agentic services. Their operational maturity is the foundation the agents build on.
The models are capable enough. The infrastructure is trustworthy enough. The organisations are structured enough. The question for enterprise legal leaders is no longer "can this work?" It is "does this fit our operating model, and what does the transition look like?"
If you are a GC or Head of Legal Ops at a large enterprise, the emergence of agentic services changes several things about how you plan, budget, and staff.
The staffing model conversation changes. The question is no longer "how many lawyers do we need?" It is "which work requires a lawyer, and which work requires a supervisor?" The distinction is important. Routine contracting, first-pass reviews, NDA processing, intake triage: these are candidates for agentic execution. Strategic advice, complex negotiations, novel regulatory questions, high-stakes disputes: these remain human work. The legal team does not shrink. Its composition shifts.
The budget conversation changes. If your organisation already outsources routine work to ALSPs or law firms, agentic services compete for the same budget line. This is not a new software purchase requiring a new business case. It is a reallocation of existing services spend to a model with better economics. If your organisation handles everything in-house, the budget conversation is about capacity expansion without proportional headcount growth, which is harder to quantify but increasingly necessary as legal workloads scale faster than hiring.
The vendor landscape is dissolving its traditional categories. The distinction between buying a tool and hiring a firm is less clear than it was a year ago. AI-native law firms use proprietary platforms. Platform companies offer supervision services. ALSPs layer AI on existing headcount. The buyer who evaluates agentic services needs to assess the delivery model, not just the technology, and the questions about control, dependency, and knowledge ownership matter as much as the questions about accuracy and speed.
It would be irresponsible to close without noting the limits. Agents do not replace judgment. They do not handle genuinely novel situations. They are not suitable for work where the stakes are existential and the playbook does not exist. They do not eliminate the need for experienced lawyers. They change what experienced lawyers spend their time on.
The opportunity is not the 20% of legal work that is genuinely complex. It is the 80% that is not. For a category whose promise is to restructure how routine legal work gets done, that is a large enough ambition.