Agents are AI teammates you add to a matter. Unlike a chatbot that waits to be asked, a Lupl Agent has context: it knows the matter, the team, and the work in progress. It takes action, not just gives answers.
Agents are proactive by default. They respond to events – emails, status changes, task assignments – not just prompts. Because they live inside the same workspace as your tasks, documents, and workflows, they start with real matter context rather than requiring you to re-explain background every time.
| Access: Agents are currently rolling out in private preview to selected design partner firms. Contact your Lupl representative to request access. |
Prerequisites
To create and deploy agents, you need:
- Lupl Enterprise customer account
- Knowledge Admin permissions – the same permission level required to create and manage matter templates
- At least one active matter in Lupl where the agent will be deployed
- Access to the private preview release (contact your account manager to enable)
Agents are accessible from the Me menu in the top right corner of Lupl. Once created, agents are visible to all Knowledge Admins across your firm.
How Agents Work
Every agent is built from four components:
01 – Identity: Who the agent is
A name, headline, and role description. The role description is the most important field: it is a system-level prompt that tells the agent how to behave, what to focus on, and what rules to follow. Experiment with it to shape how the agent responds. Be specific about which team members handle which types of work, what the agent should prioritise, and any rules it must follow.
02 – Knowledge: What the agent knows
Documents you give the agent to work from: rate cards, practice notes, checklists, team directories. Think of it as briefing a new team member before they start. The more relevant context you provide, the more useful the agent becomes. Knowledge documents are available to the agent every time it works on a matter.
03 – Tools: What the agent can do
Knowledge Admins can enable or disable individual tools for each agent. The available tool categories are:
- Lupl tooling – create, read, and update tasks, workstreams, and steps
- Document automation – draft and assemble firm documents using your templates
- Document search – search the matter document space in Lupl
- Web search – look up live data, exchange rates, regulations, and more
04 – Triggers: How the agent is instructed
Agents do not run continuously. They are activated by specific events. There are five ways to put an agent to work – see Ways to Instruct Your Agent below.
Matter context and isolation
When triggered, an agent has access to:
- The matter structure, phases, and workstreams
- Tasks, deadlines, and status
- Documents and document comments
- Prior decisions and commentary within the matter
There is no need to paste background into a prompt or re-explain the matter. When an agent is added to a matter, a unique instance is created for that matter. This means each agent instance is completely isolated: it can only access its assigned matter and cannot read data from or leak context to any other matter.
Human supervision
Every agent has a designated human supervisor. Agents can propose and perform work, but the supervisor can review outputs, approve actions, or override decisions at any point. All agent activity is logged, just like human activity, providing a complete audit trail.
| Note: During the private preview, supervisors are not automatically notified of agent actions. You can configure automations or agent instructions to trigger escalations at defined points. Automated supervisor notifications will be introduced in a later phase. |
Creating Your First Agent
It takes less than two minutes to create your first agent. Start simple: you can always refine the role description and add knowledge documents later.
-
Open the Agents menu. Click the Me menu in the top right corner of Lupl and select Agents. You will see any agents already created by your firm’s admins.
-
Click "Create Agent".
You have two options: start from scratch, or start from a prompt.
Starting from a prompt is recommended. Describe what you want the agent to do in plain language – for example: "A virtual legal project manager who helps with pricing, scoping, and keeping work on track."
-
Review and refine. Lupl generates a name, headline, and role description. All are editable. The role description is your most powerful lever: be specific about how you want the agent to behave, what it should prioritise, and any rules it should follow.
-
Add knowledge (optional). Upload documents the agent should reference: rate cards, matter checklists, team directories. These sit in the Knowledge section of the agent and are available every time it works on a matter.
-
Add the agent to a matter. Agents only work inside matters. Click Add to Matter, select the matter, assign a supervisor, and confirm. The agent will appear in the Agents tab in the top right of the matter view and is available to anyone on that matter.
| Tip: Agents are visible to all Knowledge Admins across your firm. Coordinate with your team so you build a shared library of well-defined agents rather than duplicates. Think of them like matter templates: define them once, use them everywhere. |
Ways to Instruct Your Agent
Once an agent is on a matter, there are five ways to put it to work. Three are reactive: you initiate the interaction directly. Two are proactive: the agent is triggered by events and acts without you asking.
Reactive
-
(A) – Chat Click the Agents tab inside a matter, select your agent, and type your request. Ask it to draft emails, update tasks, look up information, create workstream items, or answer questions. The agent has full visibility of the matter and can act on anything within its tools.
-
(B) – Assign a task Assign any workstream task directly to an agent, just as you would to a team member. The agent reads the task, uses its tools, and completes the work. Well-suited for document drafting, research, and structured updates.
-
(C) – @Mention in a workstream comment Mention an agent in any workstream comment using @. The agent picks up the instruction in context and responds or takes action directly within that item. You can also @mention an agent in a document comment. Useful for quick, targeted requests without leaving the workstream view.
Proactive
-
(D) – Assign via automation Use Lupl’s Automations to trigger an agent based on workstream events, conditions, or scheduled time: when an item is created, when a status changes, or at a set time each day. The agent acts on each trigger automatically, with no manual input required.
-
(E) – Forward emails to matters (Driverless LPM) Every matter has its own email address. Set up an automation: when an email arrives, the agent reads it and updates the relevant workstream items. Notes, statuses, and assignees stay current from real-world communications, with no one needing to log in and update manually. Agents can also be assigned emails directly from the Lupl Outlook add-in and will receive the email content, including attachments, as context.
Admin Controls
Knowledge Admins have full control over agents. They can:
- Create agents – build and configure new agents with instructions, knowledge, and tools
- Configure tool access – enable or disable individual tools for each agent based on its intended role
- Assign agents to matters – add agents to specific matters, creating isolated instances
- Assign supervisors – designate a human supervisor for each agent-matter deployment
Agents cannot be added to matters or take any action without administrator involvement. Agents are not activated globally – they only become active when invited into a specific matter.
| Governance tip: Before rolling out agents, consider defining firm policies around: which matter types or practice areas will use agents; which Knowledge Admins can create and deploy agents; how supervisors will review agent outputs; and which tools are appropriate for each agent type. |
Security and Data
Permission model
Agents inherit Lupl’s existing permission framework. An agent can only access matters it has been explicitly added to. Within a matter, the agent is subject to the same data access controls as human users. If a user cannot see a document or workstream, the agent cannot see it either.
Matter isolation
When an agent is added to a matter, a unique instance is created for that matter only. Agent instances cannot access or share context with any other matter. There is no risk of cross-matter data leakage.
Audit and logging
All agent actions are logged with timestamps, action type, and context. Audit logs are accessible to administrators and include both successful actions and errors. This provides the same audit trail as human activity in Lupl.
Data handling
- All data is encrypted in transit (HTTPS/TLS 1.2+) and at rest (AES-256)
- Agent data is never used to train AI models
- Agents operate within Lupl’s existing data residency and sovereignty controls
- Hosting: Agents run on Microsoft Azure AI Foundry. Microsoft contractually commits to never using customer prompts or completions to train models. Microsoft may temporarily retain data for up to 30 days for abuse monitoring and service reliability.
| Note for firms with strict data governance requirements: Some firms require a zero data retention configuration where no prompt or completion data is retained at any point. If this applies to your firm, contact your Lupl account manager to discuss your specific requirements before enabling agents. |
AI model
Agents currently run on GPT 5.4, the frontier model from OpenAI, hosted within Microsoft Azure. The architecture is model-agnostic, allowing Lupl to update or configure models based on performance and capability requirements. Private deployment of your own Azure OpenAI service is planned but not available during the private preview.
Tips and Common Use Cases
Matter pricing and fee quotes
Give the agent your firm’s rate card as a knowledge document. Ask it to review the scope of work in your workstreams and draft a fee quote or proposal email. It can convert currencies using live exchange rates via web search.
Intake and triage
Connect a client intake form to an automation. When a new request comes in, the agent categorises it, writes a summary, and routes it to the right team member based on practice area, jurisdiction, or any criteria you define in the role description.
Daily progress reporting
Set up a scheduled automation: each day, assign the agent to review matter progress and produce a draft status email flagging risks, overdue tasks, and items needing attention. Send it automatically to the matter team or supervising partner.
Scope creep detection
When new tasks are added to a matter, trigger the agent to compare them against the original engagement scope. If work has been added outside what was agreed, the agent sends an alert before anyone realises at billing time.
Smart task assignment
Give the agent a team directory: who handles what, their specialisms, and jurisdictions. The agent reviews new workstream items and assigns the right person automatically, saving coordinators the back-and-forth of manual routing.
Document automation
Assign document tasks to your agent. Using your firm’s templates, it drafts and assembles documents such as engagement letters, proposals, and NDAs, then saves them directly to the matter document space. It can also flag the draft for human review by assigning a task to the supervisor.
GDPR and data subject requests
When a data subject request arrives, the agent classifies it (right of access, right of erasure, etc.), checks whether the required ID has been provided, calculates the response deadline, and routes it to the right person.
Driverless LPM: updating matters from email
Forward or cc the matter email address and the agent processes the email automatically. It can create, edit, or close workstream items; mark tasks as done; update statuses; and add notes. No one needs to log in and update manually. This is particularly useful on matters where the team is in back-to-back meetings or working across time zones.
| Start small: Agents work best when you treat them like a new team member: give them clear instructions, good materials, and a real task to do. Begin with one agent in one test matter. See what it can do. Build from there. |
FAQs
Are agents fully autonomous?
No. Agents operate within defined permissions and always have a human supervisor. They can propose and perform work, but accountability stays with people. Supervisors can review, approve, or override at any point.
Can agents access sensitive matter data?
Only if explicitly permitted. Agents inherit Lupl’s permission model and cannot access anything outside their assigned matter. Administrators define which tools each agent can use.
Can an agent leak data between matters?
No. Each agent instance is isolated to its matter. It cannot access or share context with other matters.
Who can create agents?
Knowledge Admins – the same permission level required to create and manage matter templates.
Are agent actions auditable?
Yes. All agent activity is logged with timestamps, action details, and context. Logs are accessible to administrators.
Is data used for model training?
No. Agent data is not used to train AI models.
How are Lupl Agents different from tools like Harvey or Legora?
Tools like Harvey and Legora focus on discrete tasks such as summarisation and document review. Lupl Agents integrate those tasks into a complete matter process, combining agents with human teammates and a project management framework so that everyone has visibility, work stays on track, and humans remain in the loop. Lupl is also working with selected customers to explore integration with tools like Harvey and Legora via API and MCP, so agents can call those tools directly.
Does it work with Outlook?
Yes. You can assign emails to an agent directly from the Lupl Outlook add-in. The agent receives the full email context, including attachments, and combines it with the matter context to take action.
Can agent behaviour be changed over time?
Yes. Agents are configured in natural language and can be updated at any time. They can also be adjusted as a specific matter evolves.
What does it cost?
Agents are a separate SKU from the Lupl Enterprise subscription. During the private preview, there is no additional charge, subject to fair use limits. Pricing will be determined following the preview period. Contact your Lupl account manager for more information.
Can we connect our own AI model?
Private deployment of your own Azure OpenAI service for Agents is planned but not available during the private preview. Contact your account manager for the current roadmap.
Private Preview Limitations
The following limitations apply during the private preview period. They will be addressed in subsequent phases based on customer feedback.
- Supervisor notifications: Automatic notifications to human supervisors are not yet implemented. Supervisors can be notified via automations or agent-initiated escalation prompts.
- External integrations: API and MCP integrations with external systems such as document management systems (e.g., iManage, NetDocuments) are on the roadmap but may not be available during the preview.
- Matter channel: Agents cannot currently be triggered via the matter channel.
- Private model deployment: Connecting your own Azure OpenAI service is not available during the preview.
Related Resources
- Introducing Agents in Lupl – Blog Post
- Agents feature page
- Questions and access requests: success@lupl.com