AI agents are turning up in more places across the law firm. A practice group may build one in Microsoft Copilot Studio. Another may use agents inside a specialist legal AI platform. An innovation team may connect an internal assistant to the document management system, while a lawyer experiments with a tool the firm has never approved.
Each agent might solve a real problem. Together, they create a management problem: Who owns them? Which client and matter information can they see? What tools can they call? How much do they cost to run? And which version of the firm's knowledge do they treat as authoritative?
That is AI agent sprawl. As agents multiply, law firms need to govern both their actions and the information they use. Cost is one consequence, but so are duplicated work, inconsistent answers, oversharing and agents that remain active long after their purpose has ended.
The scale of the issue is becoming clearer across industries. Gartner predicts that the average global Fortune 500 enterprise will have more than 150,000 agents in use by 2028, compared with fewer than 15 in 2025. It also reports that only 13% of organizations believe they have the right AI agent governance in place. Those are cross-industry figures, not a forecast for law firms, but they show why firms should put controls in place before their agent estate becomes difficult to understand.
AI agent sprawl happens when agents are created or adopted across teams and platforms faster than the firm can discover, approve, manage and retire them. A simple chatbot may answer a question; an agent may plan steps, search repositories, use tools and pass results to another system or agent. That wider reach makes its identity, permissions, connected sources and ongoing behavior matter.
Legal work raises the stakes. A transaction agent may need an approved precedent but must not see another client's restricted matter files. A research agent may need the latest jurisdiction-specific guidance rather than a superseded note. A diligence agent may use substantial compute while lawyers still need to review its findings.
An inventory alone will not solve all of this. A useful framework covers five connected controls:
Start by finding agents wherever they are built or used: approved platforms, specialist legal products, internal apps, shared workflows and unsanctioned experiments. Bring together technology, innovation, KM, information security and practice teams; no single group is likely to see the whole picture.
For each agent, record its owner, business purpose, users, platform, connected repositories, tools and any Model Context Protocol (MCP) servers it can access. Note whether it can trigger actions, share information with another agent or reach client and matter data. Flag duplicates and assign a risk level based on its permissions, data sensitivity and ability to act.
The point is to make deliberate choices. Two teams may have built similar agents because neither knew the other existed. Some experiments may deserve a supported route into production; others may need to be restricted or retired. Gartner includes discovery of both sanctioned and shadow agents in its agent-sprawl guidance.
Quick check: Can your firm account for every agent that can reach its legal knowledge or client information? For a broader starting point, see ClearPeople's AI governance overview.
An agent needs a clear identity and an accountable owner. Define who can create, approve, change and retire it. Give it only the access needed for its approved task, and review that access when its purpose, connected tools or user group changes.
For a law firm, that means testing how agent permissions interact with client confidentiality, matter restrictions and ethical walls. Access inherited from a user or connector can have consequences beyond the original use case. Where the platform supports it, use distinct identities, least-privilege or time-limited access, approval states and periodic access reviews. Record the next review date and a way to disable the agent when a project ends or its owner leaves.
Apply the same discipline to tools and agent-to-agent handoffs. An agent that can read a document has a different risk profile from one that can publish an answer, update a record or send information to another service.
Agents can perform multi-step work, revisit sources and carry growing context between steps. Their consumption can therefore vary widely, even when two workflows start with a similar request. In research on coding agents, Bai and colleagues found high and variable input-token use, while additional spending did not consistently improve accuracy. Those results come from coding tasks and should not be treated as law-firm cost ratios. The mechanism is relevant: repeated retrieval and lengthy context can increase the resources an agent uses.
Pricing is also changing. Gartner forecasts that consumption-based pricing will represent more than 35% of net new corporate legal technology spend with major vendors by 2028. That prediction concerns corporate legal departments, not law-firm procurement. In the law-firm market, Legora announced consumption-based pricing for Agent Pro in June 2026, with project or matter attribution and spending controls in its product.
Ask vendors what counts as consumption, what is included in a subscription, how overages work and whether usage can be attributed to an agent, workflow, practice or matter. Where supported, set thresholds and alerts, forecast high-volume uses and reserve expensive reasoning for work that warrants it. If matter-level allocation is unavailable, track the closest reliable unit and disclose that limitation in the business case.
The aim is not the cheapest answer. It is a useful legal result at a cost the firm can explain. For the broader productivity and measurement case, read Legal AI ROI: how law firms can prove value. To explore one component of model usage, use ClearPeople's Azure OpenAI cost calculator; its estimates are not a full model of agent workflows or vendor contracts.
An agent's instructions and permissions determine part of its behavior. The quality of the information it retrieves determines much of what it can say with confidence. Gartner's agent-sprawl guidance therefore includes AI information governance: keep information current, manage access to prevent oversharing and archive it when obsolete.
For legal knowledge, this means recording more than a document's title and location. An agent needs signals about authority, provenance, permissions, jurisdiction, currency, lifecycle status and suitability for the task. A precedent may be approved for one practice but unsuitable for another. A matter document may contain a useful clause while remaining restricted to the matter team. A research note may be accurate when written but outdated today.
Classification and retention rules need to follow information as it moves through connected systems. Firms should also decide what an agent may pass to another agent or tool, and prevent redundant, expired or restricted content from being treated as an approved source. The practical goal is for each workflow to retrieve the smallest useful set of authoritative, relevant and permission-appropriate knowledge.
Imagine a lawyer asking for the firm's preferred limitation-of-liability wording for a particular transaction. An agent searches the DMS, SharePoint, Teams and precedent collections. It finds an approved template, an earlier version, a clause negotiated for a different client and a practice-group note. Without clear authority and jurisdiction signals, it may retrieve more material and spend more steps working out which source matters. A lawyer then has to verify its conclusion.
Now imagine 25 agents serving different teams, each connected to iManage, Microsoft 365, CRM and practice resources. Each may reconstruct the same institutional position from scratch. That multiplies retrieval effort and the opportunities for inconsistent answers. A governed, reusable knowledge layer can give those agents a more consistent basis for finding and checking the firm's approved knowledge. Whether that reduces a particular vendor bill must be tested against the workflow and pricing model; the value of clearer provenance and permissions also extends beyond cost.
Approval is the beginning of an agent's lifecycle. Monitor what it does in production: which sources and tools it calls, how much it consumes, where it sends information, whether its output meets quality requirements and whether its behavior changes after an update.
Assign someone to investigate unusual access, unexpected spending, repeated failures and answers grounded in outdated or restricted content. Keep records suited to the firm's confidentiality requirements and the capabilities of each platform. Establish a way to pause, correct or retire an agent when it exceeds its intended scope. Review both agent configuration and the underlying knowledge source: sometimes a poor answer reflects a broken connector or obsolete precedent rather than the model itself.
An agent can have an owner, a unique identity, a budget and a complete audit trail, yet still choose the wrong precedent. Those controls govern what it can do. Knowledge governance helps determine what it can access, which sources are approved and what it should trust.
This is particularly important when firms use several AI platforms. Rebuilding the same rules about authority, ownership, review dates and access separately for every agent is hard to sustain. A shared approach to curating knowledge lets different tools draw on more consistent, permission-aware context while the firm retains the agent-specific controls needed for each workflow.
If several answers are unclear, start with one high-value workflow. Map its agent, access, source knowledge, costs and human review. That gives the firm a practical pattern to apply to the next one.
AtlasFuse by ClearPeople helps firms structure and govern knowledge across Microsoft 365 and connected sources so people and AI can work with more reliable context. Knowledge Collections, metadata, permissions and lifecycle controls can help teams identify and reuse approved content, with controls for AI use and visibility into consumption within the product's supported experiences.
Its strongest role in this framework is AI information governance: improving what an agent can find, access and understand. The firm still needs agent discovery, identity management and runtime monitoring across its chosen platforms. For the deeper explanation of how retrieval and context affect expenditure, see the original AtlasFuse article on legal AI cost.
For a fixed-fee matter, an agent that cuts lawyer hours might improve the margin, but only if its consumption and the time spent checking or correcting its work stay within the expected range. A cheaper run may be worse value if its answer needs extensive rework. Compare consumption by workflow, available matter or use-case allocation, review and correction time, and whether the output meets the firm's quality threshold. The more meaningful measure is cost per trusted legal outcome, not cost per token.
The management sequence is practical: discover agents, assign ownership and permissions, control consumption, govern the information they use, monitor their behavior and assess the resulting legal work. Before scaling another agent, talk to us about the knowledge and governance foundations it will need.
AI agent sprawl is the growth of AI agents across teams and platforms without a complete view of their owners, permissions, connected systems, costs and lifecycle. It can include approved agents and unsanctioned experiments. For a law firm, the concern is not simply the number of agents but whether each has a defined purpose and operates within the right client, matter and knowledge boundaries.
Why is agent sprawl risky for law firms?Agents may reach confidential matter information, invoke external tools, produce inconsistent advice or keep running after a project ends. Multiple agents can also duplicate work and repeat costly retrieval across the same repositories. The actual risk depends on an agent's access, ability to act, information sources and human oversight, so firms should assess each use case rather than treat every agent as equally risky.
How should a law firm inventory AI agents?Combine platform administration records, vendor information and interviews with practice, innovation, KM and IT teams. For every agent, record its owner, purpose, users, identity, connected data and tools, ability to act, approval status and review date. Include pilots and shadow use. Prioritize agents that can access client or matter information or perform actions in connected systems.
What is AI information governance?AI information governance is the set of rules and processes that controls which information AI can access, how that information is classified and maintained, and when it should no longer be used. In a law firm, it includes matter permissions, ethical walls, source authority, jurisdiction, version control, retention and provenance. It helps agents find material that is appropriate for a particular task and user.
How can law firms control AI agent costs?Measure consumption by agent and workflow where the platform allows it. Check what a vendor includes in its subscription, how it charges for overages and which tools or models drive spending. Set approval thresholds and alerts for expensive workflows, and compare consumption with lawyer review time and output quality. Some vendors support matter-level allocation; others require an estimate or a use-case-level proxy.
How does governed knowledge reduce context sprawl?Governed knowledge gives agents clearer signals about which sources are current, approved, relevant and accessible. That can reduce the need for separate agents to search many similar documents and reconstruct the firm's position each time. It may improve consistency and make results easier to verify. A reduction in AI charges is possible in some workflows, but depends on retrieval design, usage and vendor pricing.
What does legal AI cost include?Legal AI cost can include subscriptions, usage-based agent or model charges, integrations, external information, knowledge infrastructure and the time lawyers spend reviewing or correcting outputs. An isolated token count does not show whether the work was valuable. Firms should compare the full cost of a workflow with the quality and speed of the legal outcome it helps deliver.