AI Agents Overview
Turn on AI agents across finance, HR, vendors, nonprofits and compliance, then see everything they found on the results board, by site and grant, with who owns each one and the dollars at stake.
Muin includes a growing library of AI agents across 8 modules: Document Intelligence, Finance, HR, Vendors, CRM, Nonprofit, Compliance, and General. Agents run your workflows across those modules. Unlike a workflow that follows a fixed branch, an agent can read a document, weigh what it finds, and propose the step you configured. Every action it may take is one you granted it, and any change it proposes to your records waits for a person to approve it. See Approving Agent Changes.
What Are Agents?
Agents are intelligent automation that:
- Run on their own - Read and check your records on a schedule or when something happens. Any change they propose waits for your approval
- Make decisions - Analyze context and choose actions
- Handle exceptions - Know when to escalate to humans
- Learn patterns - Improve over time
- Report results - Full transparency on what they did
Think of agents as skilled virtual employees who never sleep and never forget.
Agents vs. Workflows
Both agents and workflows automate work, but they’re suited for different tasks:
| Aspect | Workflows | Agents |
|---|---|---|
| Logic | Rule-based (if/then) | AI-driven decisions |
| Complexity | Single path, predictable | Multi-step, adaptive |
| Exceptions | Follow defined paths | Handle novel situations |
| Oversight | Runs exactly as configured | Reports and escalates |
| Best for | Routing, notifications, simple actions | Processing, analysis, complex tasks |
When to Use Workflows
- Route documents for approval
- Send notifications on events
- Update status based on conditions
- Simple, predictable processes
When to Use Agents
- Process and categorize documents
- Match invoices to purchase orders
- Verify document completeness
- Monitor compliance and deadlines
- Handle tasks with many variables
How Agents Work
Agents follow a consistent pattern:
Triggered → Analyze → Plan → Execute → Report
│ │ │ │ │
Started Understand Decide Action Results
1. Triggered
Agents start when:
- Triggered by a workflow
- Scheduled to run
- Manually started
- API call
2. Analyze
The agent examines:
- Documents in scope
- Extracted data
- Historical context
- Related records
3. Plan
Based on analysis, the agent:
- Determines what needs to happen
- Identifies the best approach
- Considers exceptions and edge cases
4. Execute
The agent takes action:
- Updates document fields
- Creates or completes tasks
- Routes for human review
- Triggers other workflows
5. Report
Every agent run produces:
- What was processed
- Decisions made
- Actions taken
- Exceptions flagged
- Performance metrics
Agent Library
Muin includes a growing library of pre-built agents across 8 modules, ready to deploy (15 are live today). Here are some of the most-used agents.
Invoice Processing Agent
Purpose: Process incoming invoices end-to-end
What it does:
- Verifies invoice completeness
- Extracts and validates data
- Matches to purchase orders (three-way match)
- Detects duplicates
- Routes for appropriate approval. A payment it schedules always waits for a person
- Flags anomalies for review
Configuration:
- PO matching tolerance (% variance)
- Approval bands, read from your rules in Settings > Approvals
- Duplicate detection rules
- Approval routing rules
Document Verification Agent
Purpose: Check documents for completeness and accuracy
What it does:
- Verifies required fields present
- Validates data formats
- Checks for inconsistencies
- Flags incomplete documents
- Requests missing information
Configuration:
- Required fields by document type
- Validation rules
- Action on verification failure
Compliance Monitoring Agent
Purpose: Monitor certifications, licenses, and compliance documents
What it does:
- Tracks expiration dates
- Sends renewal reminders
- Flags expired credentials
- Monitors compliance status
- Generates compliance reports
Configuration:
- Alert thresholds (days before expiration)
- Notification recipients
- Escalation rules
- Report schedule
Reconciliation Agent
Purpose: Match transactions across documents and systems
What it does:
- Matches invoices to payments
- Compares POs to receiving records
- Identifies discrepancies
- Reconciles across periods
- Reports unmatched items
Configuration:
- Matching rules
- Tolerance thresholds
- Period to reconcile
- Exception handling
Data Extraction Agent
Purpose: Extract and structure data from documents
What it does:
- Processes batches of similar documents
- Extracts specified fields
- Validates extractions
- Exports to structured format
- Handles variations in document format
Configuration:
- Target document type
- Fields to extract
- Output format
- Validation rules
Grant agents
For nonprofits that seek or make grants. See Finding Grants for how each one shows up in the app.
- Grant Finder: scans federal listings, IRS grant records, program pages for your area and the pages you watch, and scores each grant against your Company Profile.
- Funder Research: reads a funder’s IRS grant records (giving per year, typical and largest cash grant, past grantees like you) and finds foundations that paid organizations like yours.
- Organization Profile Builder: drafts your Company Profile from your website and your public tax filings. Nothing is saved until you choose Save.
- Proposal Drafter: writes a grant proposal from your interview answers, the tracked grant and what is known about the funder. Every figure names its source.
- Gap Collector: turns each fact a proposal draft is missing into an inbox task for the person who has it, and closes the task when the fact is in the draft.
- Grant Recommender: for grantmakers, checks an applicant’s public filings, peer benchmark, federal awards and IRS standing, then scores the application for your review.
Activating an Agent
From the Agent Library
- Go to Sidebar → Agents → Agent library
- Browse the library
- Click Set up on the agent you want
- Work through its setup steps and click Turn on
Recommended for your organization
At the top of Agents → Agent library, Muin shows a short row of starter agents picked for your kind of organization, with the heading “Recommended for your mosque” (or church, school, food bank, or organization):
| Kind of organization | Starter agents |
|---|---|
| Mosque | Zakat Eligibility Check, Thank-you drafts |
| School | Tuition Arrears Follow-up |
| Church | Duplicate payables |
| Food bank | Duplicate payables |
| Any other organization | Duplicate payables |
A starter appears only when you could turn it on from the library yourself: its module is on for your organization, and a paid agent appears only once you have bought it. An agent you already run shows Already on with a link to it, so you never get a second copy by accident.
A paid agent runs only while you hold it. When its purchase ends, its next run is refused and the agent is paused; Run now and Turn on say which item to buy to use it again.
Setting up and changing agents is limited to owner and admin roles.
Configuration Steps
Setup writes nothing until the last step, so you can leave at any point without turning anything on. Which steps appear depends on the agent:
- Settings - the values this agent exposes, when it has any
- Alerts - whether findings raise inbox tasks
- Schedule - shown when the agent ships with one. You change it later on the agent’s own Schedule tab
- Test with these settings - a real run against a sample, where nothing is saved
- Review and turn on - name the copy, check the values, turn it on
Full walkthrough: Turning On an Agent and Reading Its Runs.
Building Your Own
When nothing in the library fits, Create Custom Agent builds one from a sentence: describe what it should do, check the review screen, try it on your last 30 days of records, then turn it on. The seven-step wizard is still there under Advanced: Building a Custom Agent.
Example: Activating Invoice Processing Agent
Name: AP Invoice Processor
Scope: Finance Module - Invoices
Settings:
- PO matching: Required for >$500
- Match tolerance: 2%
- Approval bands: from Settings > Approvals
- Duplicate window: 30 days
Schedule: Continuous (process as documents arrive)
Escalation: Create task for AP Manager
Results board
The Agents page opens with what your agents found and what it is worth. The parts below appear in this order.
What your agents did for you
Four tiles. Issues caught and Issues resolved count the last 30 days. Hours saved and Money saved count this month. The two findings tiles open your Inbox. The two savings tiles open Value. The row is hidden until there is something to count.
Under it are the run metrics for today: Active Agents, Tasks Today, Documents Processed and Tokens Used. Before any agent has run, the page says so instead.
What your agents found
One row per agent, with its Open findings, High priority findings, findings Closed (30 days) and its Last run. The agent’s name opens that agent. Closing a finding in the Inbox updates its row.
Value this month
Value found and Value realized, in dollars and in hours. Found is at stake: identified and cited, not yet acted on. Realized is confirmed. The two are never added together. Both tiles open the monthly value report.
When a rental booking lacks the insurance its space requires, the uncovered amount appears on its own Coverage at risk line: coverage a renter is required to carry but has not shown. It is not money saved, and it is never added to Value found or Value realized. The line is hidden when nothing is at risk.
Findings by site, grant and severity
Every finding that is open now or was raised in the last 30 days, one row each:
| Column | What it shows |
|---|---|
| Finding | The finding’s title. It opens the record the finding is about. Show your work opens the facts, sources and checks behind it. |
| Severity | How serious the agent rated it |
| State | Open or Closed, read from the finding’s Inbox task |
| Owner | The person the task is assigned to, the role it went to, or Unassigned |
| Due | The task’s due date, or No due date |
| Site | The site it is about, or Organization-wide |
| Grant | The grant it is about, if any |
| Rule cited | The rule the finding relies on, linked to its official source, or No source cited |
| Verification | Verified, Not yet verified or Statement removed |
| Raised | When the agent raised it |
Narrow the list with the Site, Grant, Severity and Open or closed filters. The heading shows how many findings match out of the total. A finding you close in the Inbox moves from Open to Closed here.
The board reads the newest 500 findings. When there are more, it says Showing the newest 500 of N findings, and the filters search only those 500.
Dollar figures on a finding
A finding carries a dollar figure only when your own records support it, and it names the basis. For example, a facilities finding states the property tax your exemption is worth, computed from the assessed value and levy rate you recorded. When a record is missing, the finding says the dollar figure cannot be computed and what to add, and it shows no number. See Rent Out a Room or Hall for the facilities figures.
What your agents produced
The latest output each agent keeps, with the date it was produced: reports (such as the facilities audit), calendars, and the spend plan for each grant award. A spend plan opens the award’s Compliance tab. When no agent has produced anything yet, the shelf says No reports yet.
The next 12 months
Dates ahead of you, from the records that already hold them: permits and licenses, inspections, statutory filings, registration renewals, grant application windows and report due dates, and the days left under a property-tax exemption’s use cap. Each row shows the date, what is due, the days left and what is at stake. A row with no dollar basis reads Not modelled. When there are more rows than fit, See all dates opens the monthly value report.
How much of this was checked
Of the findings raised this month, how many statements were Computed from your records, Verified by a second model, Not yet verified, or removed (Unsupported statement removed). It also shows how many questions the agents asked back, and the monthly accuracy test: how many of a fixed set of known mistakes the checker caught and the month of the run it reports. None of those test claims reach your records. A check that did not pass says what fell short. A month with no check says so, never a perfect score. Value this month and this panel show only to people who can view intelligence reports.
Export the findings
Export CSV and Export PDF download the findings that match your current filters, including state, owner and due date. When the board holds only the newest 500, the export ends with a line saying it was read from the newest 500 of N findings and that older ones are not in the file.
Who gets findings
Open an agent you have turned on and choose its Alerts tab. Findings go to names one person, and every finding that agent raises becomes an Inbox task assigned to them. Choose Nobody named (use the routing below) to go back to the default: a run you start goes to you, and a scheduled run goes to the role in Scheduled runs go to. Which findings sets the lowest severity that raises a task.
Monitoring Agents
Individual Agent Status
Each agent shows:
- Current status (Active, Paused, Error)
- Last run time
- Items processed
- Success/failure count
- Pending items
Run History
Open an agent and use its Runs tab. Click a run to read it.
A run is in one of six states: pending, running, completed, failed, cancelled or timeout. A failed run can be retried; one still going can be cancelled.
Run Details
Each run record includes:
- Start and finish time, and duration
- The trigger and the triggering event
- The definition version that ran
- Findings - what it concluded, and the inbox tasks it raised
- Extracted Fields - the values it read out of a document
- Execution Steps and AI Decisions - the sequence it worked through, each marked Completed, Failed, Skipped or Running
- Quality Flags - anything it was unsure about
- Confidence, the model used, and input / output / total tokens. An agent that worked from your records rather than a document reports “Not used. Worked from your records.” instead
- The error code, message and detail, when it failed
Runs record token counts, not a dollar figure.
Agent Notifications
When Agents Notify You
Agents send notifications for:
- Exceptions - Items requiring human review
- Errors - Processing failures
- Completions - Batch processing finished
- Milestones - Significant thresholds reached
Configuring Notifications
Open the agent and choose its Alerts tab:
- Raise inbox tasks for findings turns findings into Inbox tasks, each with a due date, until someone closes it.
- Which findings: Medium and above, High and above or Critical only.
- Findings go to and Scheduled runs go to decide whose Inbox each task lands in. See Who gets findings.
Exception Handling
When an agent encounters something it can’t handle:
- Agent logs the exception
- Creates a task for human review
- Sends notification
- Continues processing other items
- Human resolves exception
- Agent can re-process if needed
Agent Performance
Agent health
Sidebar → Agents → Agent health lists every agent you turned on, with the ones that need your attention first. For each agent it shows the credits it used this month and how often its findings were rejected compared with last week.
Each agent’s own Overview tab shows what it did for you: hours saved, items caught, runs that finished, how often you approved and rejected its findings, tool errors, credits per run, and credits this month against its monthly limit. See Turning On an Agent and Reading Its Runs.
Understanding Metrics
| Metric | Description | Good Range |
|---|---|---|
| Throughput | Items processed per hour | Varies by agent |
| Success Rate | % completed without exception | Above 95% |
| Exception Rate | % requiring human intervention | Under 5% |
| Processing Time | Average time per item | Depends on complexity |
Improving Performance
Reduce exceptions:
- Review common exception types
- Adjust agent settings to handle
- Improve document quality
Increase throughput:
- Process in batches where possible
- Optimize configuration
- Ensure adequate resources
Benchmarking
Compare your agents to baselines:
- Industry averages (when available)
- Your historical performance
- Before/after configuration changes
Using Agents and Workflows Together
Agents and workflows complement each other.
Workflow Triggers Agent
A workflow can start an agent:
- Workflow triggered by document upload
- Initial routing and tagging by workflow
- Workflow action: “Run Invoice Processing Agent”
- Agent handles complex processing
- Agent completion triggers next workflow step
Agent Creates Workflow Tasks
Agents can trigger workflows:
- Agent processes documents
- Agent proposes a task or a status change, and a person approves it
- Status change triggers workflow
- Workflow handles routing and notifications
Example: Combined Automation
Invoice received
↓
Workflow: Tag by vendor, check for PO number
↓
Agent: Process and match invoice
↓
Agent outcome: Matched / Not matched / Exception (a recommendation, never an approval)
↓
Workflow: Route based on outcome
- Matched, within your approval band → Approver's inbox, marked "within your approval band"
- Matched over $500 → Manager approval
- Not matched → AP review
- Exception → Senior AP review
Best Practices
Deployment
- Start with one agent - Learn before deploying many
- Use conservative settings - Widen scope as confidence grows
- Monitor closely initially - Watch early runs carefully
- Iterate based on exceptions - Adjust to reduce manual intervention
Configuration
- Match to business rules - Align with existing policies
- Set appropriate thresholds - Balance automation vs. oversight
- Define clear escalation - Know who handles what
- Document custom settings - Future reference
Ongoing Management
- Review exceptions weekly - Look for patterns
- Adjust settings monthly - Optimize based on data
- Audit quarterly - Ensure still aligned with needs
- Track ROI - Measure time saved
Troubleshooting
Agent Not Running
- Check agent status (Active/Paused). A paid agent is paused when its purchase ends
- Verify trigger conditions
- Check schedule settings
- Review error logs
High Exception Rate
- Review exception types
- Identify common causes
- Adjust settings or validation
- Improve data quality at source
Performance Degradation
- Check for volume spikes
- Review recent configuration changes
- Verify system resources
- Contact support if persistent
Next Steps
Now that you understand agents:
- Turning On an Agent and Reading Its Runs - Setup, schedules and run detail
- Building a Custom Agent - Build one of your own
- Building Workflows - When the job is a fixed sequence rather than a judgement
- Task Management - Where an agent’s findings land