Business Intelligence & Oversight Agents¶
Thinklio Built-in Agent Specification Version 0.1 | March 2026
This document covers four agents that share a common theme: they watch, measure, and surface intelligence from structured data and ongoing activity.
- Customer Intelligence Agent — builds and maintains a picture of customers and relationships
- Finance Agent — handles financial queries, expense management, and reporting
- Monitor Agent — watches systems, feeds, and metrics and alerts on conditions
- Compliance Agent — monitors content and agent outputs against policy rules (new addition)
1. Customer Intelligence Agent¶
1.1 Purpose¶
The Customer Intelligence Agent is the business development and account management layer. It builds and maintains a structured picture of customers, prospects, and relationships — drawing on Rolodex data, interaction history, Items, and external research to answer questions about who customers are, what they need, and where opportunities exist.
It is related to but distinct from Rolodex. Rolodex stores and retrieves relationship data. The Customer Intelligence Agent interprets that data: "this customer's engagement has declined," "this prospect has been dormant for 60 days," "these three accounts share a common pain point."
1.2 UI Structure¶
Chat Tab: Intelligence queries: - "Which accounts haven't had a touchpoint in 90 days?" - "Summarise Acme Corp's relationship with us" - "Who are our at-risk accounts and why?" - "Brief me on the prospect I'm meeting tomorrow"
Accounts Tab: A structured view of all customer and prospect organisations, with relationship health indicators, last contact date, open Items, and pipeline status.
Signals Tab: A feed of notable signals across the customer base — engagement drops, significant interactions, inbound enquiries, news about customer organisations.
Pipeline Tab: Sales pipeline tracking for prospect accounts: stage, estimated value, next action, and probability.
1.3 Agent Capabilities¶
| Capability | Description |
|---|---|
| Account health scoring | Composite score based on recency of contact, open items, interaction sentiment, and engagement trends |
| At-risk detection | Identify accounts where health is declining or contact has lapsed |
| Opportunity surfacing | Identify accounts with signals suggesting an expansion or re-engagement opportunity |
| Account briefing | Produce a structured briefing on a specific customer or prospect |
| Pipeline management | Track prospects through a configurable sales stage model |
| Relationship mapping | Identify connections between contacts across accounts |
| News monitoring | Surface news about customer organisations that may be relevant |
1.4 Account Health Model¶
Account health is a composite score (0–100) derived from:
| Factor | Weight | Description |
|---|---|---|
| Recency of interaction | 30% | Days since last meaningful interaction |
| Interaction sentiment | 20% | Positive/neutral/negative trend from recent interactions |
| Open items | 20% | Number and age of open Items linked to this account |
| Engagement breadth | 15% | Number of distinct contacts engaged with at this account |
| Response rate | 15% | How often our outreach gets a response |
Health scores are recalculated on a configurable schedule and updated in real time when new interactions are logged.
1.5 Inter-Agent Behaviour¶
- Rolodex is the primary data source — all contact and interaction data comes from here
- Enquiry Agent feeds in inbound contact signals as they arrive
- Research Agent is invoked for external intelligence on specific accounts
- Briefing Agent is used for full account briefing documents
- Mail Agent is invoked for outreach when the agent recommends re-engagement
1.6 Use Cases¶
UC-1: At-risk account review A weekly review shows three accounts with health scores below 40. The agent surfaces them in the Signals tab with reasons: Account A hasn't had contact in 95 days; Account B has two open unresolved Items; Account C's last three interactions had negative sentiment. For each, the agent suggests a re-engagement action.
UC-2: Pre-sales briefing A sales lead has a call with a prospect in 30 minutes. They ask for a briefing. The agent synthesises: Rolodex history (2 previous meetings, last one 3 months ago), open pipeline stage, any recent news about the company, and suggests two talking points based on their expressed interests from past interactions.
2. Finance Agent¶
2.1 Purpose¶
The Finance Agent handles financial queries, expense management, and basic financial reporting. It is not an accounting system — it is a conversational interface to financial information, policies, and processes.
Its domain covers three areas: expense management (submitting, approving, and querying expenses), financial policy (drawing on the finance section of the knowledge base), and financial reporting (producing summaries from connected financial data).
2.2 UI Structure¶
Chat Tab: Financial queries and actions: - "What's our current budget for Q2 marketing?" - "Submit an expense: $240 taxi to the airport on 15 March, client travel" - "What is the approval threshold for capital equipment?" - "Generate a monthly expense summary for the engineering team"
Expenses Tab: Expense records — submitted, pending approval, approved, and rejected. Users see their own; managers see their team's.
Reports Tab: On-demand financial summaries and standard reports. Report templates for common financial summaries.
2.3 Agent Capabilities¶
| Capability | Description |
|---|---|
| Answer policy questions | From the finance knowledge library (expense categories, approval thresholds, etc.) |
| Submit expense | Create an expense record with receipt attachment |
| Approve/reject expense | For users with approval authority |
| Query expenses | Find and summarise expense records by period, category, team, or person |
| Generate report | Produce a financial summary from connected data |
| Budget query | Report on budget vs. actual for a category or period |
| Alert on anomalies | Flag unusual expense patterns |
2.4 Governance¶
Finance is a sensitive domain. Several governance constraints apply:
- The Finance Agent never provides investment advice
- Financial figures in responses are always attributed to source data
- Expense submission creates a record but does not process payment — payment processing is always external
- Approval actions are logged immutably
- Access to team or org expense data is role-gated
2.5 Library Dependency¶
The Finance Agent draws on an org-supplied finance library for policy questions: expense categories and limits, approval hierarchies, chart of accounts, travel policy, etc. Without a seeded library, the agent can only manage expense records — it cannot answer policy questions.
2.6 Use Cases¶
UC-1: Expense submission A team member asks "I need to submit an expense — lunch with a client, $180, today, business development category." The agent creates the expense record, requests a receipt attachment, and submits it for approval to the relevant approver.
UC-2: Policy query A manager asks "what's the limit for software subscriptions I can approve without finance sign-off?" The agent retrieves the relevant policy from the finance library and gives a direct answer with the policy source cited.
UC-3: Team expense summary A finance manager asks for the engineering team's expenses for Q1. The agent queries expense records, produces a summary by category and by person, and flags two outliers — one above the team's normal pattern and one pending approval for 12 days.
3. Monitor Agent¶
3.1 Purpose¶
The Monitor Agent watches. It observes configured sources — system metrics, web endpoints, data feeds, agent outputs, or any other measurable signal — and triggers alerts or actions when defined conditions are met.
Unlike the Digest Agent (which produces periodic summaries for human consumption), the Monitor Agent acts on conditions in near-real-time. It is event-driven, not scheduled.
It is the operational oversight layer: "if X happens, do Y."
3.2 UI Structure¶
Chat Tab: Configure and query monitors: - "Alert me if the support queue goes above 20 open Items" - "Monitor our careers page for changes" - "Set up a monitor for mentions of Thinklio on news sites" - "What monitors are currently active?"
Monitors Tab: All configured monitors, their status (active/paused/error), last checked time, and recent trigger history.
Alerts Tab: Triggered alerts, with context, timestamp, and resolution status.
Log Tab: Full activity log of all checks, triggers, and actions.
3.3 Monitor Types¶
| Type | What it watches | Trigger condition examples |
|---|---|---|
| Queue monitor | Enquiry Agent Item queue | Item count > N, SLA breach rate > X%, type spike |
| Web monitor | URL or page content | Page changes, keyword appears/disappears |
| Search monitor | Web search results | New results matching a query |
| Agent output monitor | Outputs from other agents | Research Agent finds a high-relevance source, Fact Checker flags critical issues |
| Data threshold | Any numeric metric | Budget utilisation > 80%, error rate > 5% |
| Keyword monitor | Email or message streams | Specific terms appear in Mail or Messenger |
3.4 Alert Actions¶
When a monitor triggers, the Monitor Agent can:
| Action | Description |
|---|---|
| Notify user | In-platform notification |
| Send email | Via Mail Agent |
| Create task | Via Taskmaster |
| Create Item | Via Enquiry Agent |
| Run another agent | Invoke a Research Agent, Briefing Agent, or custom agent |
| Post to channel | Via Messenger Agent |
Alert actions are configured per monitor and are subject to the same governance model as the relevant agent (sending an email requires Mail Agent permissions).
3.5 Use Cases¶
UC-1: Support queue spike A monitor watches the Enquiry Agent Item queue. When open Items exceed 30, it triggers: a Taskmaster task for the support lead ("Support queue alert: 31 open Items"), a notification to the support team channel in Slack via the Messenger Agent, and a briefing on the queue composition (types and ages of the open Items).
UC-2: Competitive news monitoring A monitor runs a daily web search for a competitor's name. When a significant new result appears (high relevance, from a major publication), it creates a Taskmaster task for the marketing lead, includes the article summary, and adds the source to the competitive intelligence section of the Knowledge Base.
UC-3: Budget threshold alert A monitor watches monthly spend against budget. When spend reaches 80%, it notifies the finance manager and creates a task to review the remaining budget allocation. At 95%, it escalates to the org admin.
4. Compliance Agent (new addition)¶
4.1 Purpose¶
The Compliance Agent monitors agent outputs, communications, and content against a configured set of policy rules. It is the automated compliance layer for organisations in regulated industries or with strict internal standards.
It doesn't replace human compliance review — it dramatically reduces the volume of content that needs human attention by filtering out clear compliers and flagging only potential issues.
4.2 What It Monitors¶
| Source | What it checks |
|---|---|
| Mail Agent outbound | Outbound emails for prohibited content, required disclaimers, tone |
| Messenger Agent outbound | Messages for policy compliance |
| Writer Agent outputs | Drafts for prohibited claims, required disclosures, tone |
| Knowledge Base additions | New documents for classification and policy compatibility |
| Enquiry Agent responses | Agent-drafted responses for compliance before sending |
4.3 Rule Types¶
Compliance rules are configured by an org admin:
| Rule type | Examples |
|---|---|
| Prohibited content | "Never claim product X cures Y" (healthcare), "Never promise a specific return" (financial) |
| Required disclosure | "All financial advice must include the standard disclaimer" |
| Tone restriction | "No language that could be interpreted as threatening or coercive" |
| Classification | "Any document containing personal health information must be marked sensitive" |
| Channel restriction | "Customer data may not be included in external emails" |
4.4 Compliance Actions¶
When a compliance rule is triggered:
- Block — prevent the action (e.g. don't send the email until reviewed)
- Flag — allow the action but create a compliance record for review
- Warn — notify the user that a potential issue was detected but allow them to proceed
- Auto-correct — for simple cases (missing disclaimer), append or modify automatically
4.5 Governance Profile¶
The Compliance Agent itself has strict governance: - Rule configuration is restricted to org admins only - The compliance log is immutable and accessible to compliance officers - The agent cannot be disabled by regular users - Its own outputs (compliance records, flags, blocks) are auditable
4.6 Use Cases¶
UC-1: Financial services disclaimer A financial services firm configures a compliance rule: all outbound emails containing investment-related content must include a standard disclaimer. When the Mail Agent drafts an outbound email matching the content pattern, the Compliance Agent intercepts, checks for the disclaimer, and if absent, appends it automatically before the email is sent.
UC-2: Healthcare claims monitoring A health services organisation configures a rule prohibiting specific clinical claims in marketing content. The Content Agent produces a social media draft. The Compliance Agent reviews it, identifies a phrase that could be interpreted as a clinical claim, and flags it for human review before the content is approved for publication.
UC-3: Data leakage prevention A rule restricts customer personal data from appearing in external emails. The Mail Agent drafts a reply that accidentally includes a customer's full name and account number. The Compliance Agent detects the personal data pattern, blocks the send, and notifies the user: "This email may contain customer personal information. Please review before sending."
5. Shared Open Questions¶
- The Customer Intelligence Agent's health scoring model requires interaction sentiment analysis — this is a non-trivial NLP capability. Should sentiment be an explicit field that agents populate when logging interactions, or inferred automatically from interaction content?
- The Finance Agent's expense submission flow assumes integration with an external approval workflow or finance system. The agent creates records and routes approvals, but payment processing is out of scope. This boundary needs to be clearly communicated to users.
- The Monitor Agent's ability to trigger other agents creates the possibility of cascading automation chains. An alert triggers a Research Agent run, which triggers a Digest update, which triggers a Mail send. Governance limits on chain depth should apply, consistent with the broader delegation model.
- The Compliance Agent is the most governance-sensitive agent in the set — it operates as an authority over other agents. Its relationship to the platform's own policy engine needs careful design to avoid duplication or conflict.
Previous: Coordinator Agents | Next: Creative Agents