Agents with a narrow job and a hard boundary
Nine business agent blueprints. Each one has an explicit tool allow-list, approved knowledge, escalation rules and a complete execution history. High-impact actions have no automated path at all.
Never available to an agent
- Moving money or issuing refunds
- Deleting records
- Changing payroll
- Signing contracts
- Making employment decisions
- Shell, database or unrestricted internet access
These capabilities do not exist in the runtime. A prompt cannot grant them.
Pick a role, then restrict it further
Blueprints are starting points with sensible defaults. You can narrow the tool list, tighten the prompt and add approval gates — never widen beyond the registry.
AI Receptionist
receptionistCover inbound enquiries 24/7, capture lead details and book appointments without hiring a front desk.
- Making commitments on pricing discounts
- Booking outside published availability
Escalation: Urgent or complaint language is escalated to a human within the same conversation.
AI Sales Agent
salesRespond to inbound interest in minutes, qualify against your criteria and hand a briefed lead to sales.
- Negotiating discounts or contract terms
- Issuing quotes not approved by sales
Escalation: Deals above your approval threshold are routed to a human rep.
AI Customer Support Agent
supportDeflect repetitive tickets, categorise issues and draft replies your team can approve.
- Issuing refunds
- Changing billing details
Escalation: Unresolved, angry or account-security issues are escalated to a human immediately.
AI Administrative Assistant
adminRemove inbox and paperwork overhead from operations teams.
- Deleting records
- Sending email without approval
Escalation: Anything legally sensitive or involving personnel decisions goes to a human.
AI Finance Assistant
financeCut invoice processing time while keeping every payment human-approved.
- Initiating payments or transfers
- Issuing refunds
Escalation: Every payment action requires explicit human approval; discrepancies are flagged.
AI HR Assistant
hrAutomate HR paperwork while keeping employment decisions human.
- Ranking or rejecting candidates
- Employment decisions
Escalation: Any hiring, discipline or termination question is routed to a human with no AI recommendation.
AI Marketing Operations Assistant
marketingAutomate marketing operations and reporting, not social media management.
- Sending campaigns without consent checks
- Purchasing ad spend
Escalation: Any audience without documented consent is escalated before any draft is produced.
AI Document Processing Agent
documentsTurn unstructured documents into validated structured records.
- Approving invoices
- Deleting source documents
Escalation: Low-confidence extractions are routed to a human with the source reference.
AI Business Intelligence Agent
biGive operators a weekly read on where time and money leak.
- Modifying source data
- Presenting projections as measured results
Escalation: Data quality gaps are reported rather than filled with estimates.
The complete tool surface
Anything not on this list cannot be invoked by any agent, regardless of what a model asks for.
| Tool | What it does | Scope | Impact |
|---|---|---|---|
| knowledge.search | Search the agent's approved knowledge base. | agent.knowledge | Reversible |
| leads.create | Create a lead record in the workspace CRM. | crm.write | Reversible |
| leads.lookup | Find a lead by email inside this workspace. | crm.read | Reversible |
| leads.update_stage | Move a lead between pipeline stages. | crm.write | Reversible |
| tickets.create | Open a ticket and optionally escalate. | support.write | Reversible |
| tickets.lookup | Find tickets for the current customer. | support.read | Reversible |
| tickets.reply | Draft a reply for a human to send. | support.write | Reversible |
| usage.read | Read this workspace's metered usage. | billing.read | Reversible |
| approval.request | Ask a teammate to approve a high-impact action. | approval.request | Reversible |
| email.draft | Produce a draft only — never sends it. | comms.draft | Reversible |
Third-party writes (CRM, telephony, payments) are mediated by the workspace's own connection, encrypted at rest, and every call is logged with its arguments and outcome.
Prompt injection is treated as a data problem
Documents, emails and web pages are wrapped in an untrusted-data block with an explicit instruction that they are data, not instructions. Known injection patterns are flagged on the message and recorded in the execution log.
The real defence is that the dangerous capability does not exist. Even a successful jailbreak cannot call a tool that was never registered.
Approval with detail, not a rubber stamp
An approval request contains the account, amount, destination and consequence, plus the conversation that produced it. The approver can decide without reconstructing context.
Decisions record who approved, when, and any comment — and the run resumes from exactly that point.
See an agent work on your process
Start a workspace, install an agent blueprint and run it in simulation mode. Configure a provider key when you want real model output.