HomeOperational AI Agents

Operational
AI agents

AI systems that understand operational context, call business tools, follow rules, request approvals and take controlled action. Not chatbots with a company name on them — agents wired into systems, APIs, databases and workflows.

See the architecture Agent portfolio
The distinction that matters

A chatbot answers. An operational agent changes the state of the business.

The interesting part of an agent is not the language — it is everything after the sentence: which tool it is allowed to call, which rule decides the number, who has to approve before anything is written, and what the audit trail says afterwards.

Deterministic, not generated

Quantities, stock movements, reorder points and approvals come from databases, deterministic tools and business rules — never from the model's own arithmetic.

Retrieval where it belongs

RAG carries SOPs, product information, supplier policies and operational procedures — the text a person would otherwise have to go and find.

Authority stays human

Agents prepare, check, route and escalate. Approval of spend, financial commitment and anything irreversible remains a person's decision.

01Agent architecture

How a request travels from a person to a business action

Five layers, each with a different job. The sequence below walks one real request down through all of them.

Multi-agent operating architecture
PEOPLEEmployeeraises a requestManagerasks for statusField technicianneeds contextORCHESTRATIONOperations Assistantsupervisor agent · context · routing · accountabilitySPECIALIST AGENTSInventoryProcurementMaintenanceCRMFinanceKnowledgeTOOL & CONTROLAPIsDB queriesBusiness rulesApprovalRAGAuthHuman handoffSYSTEMS OF RECORDSAP / ERPCRMInventory DBFinanceDocumentsWorkflowNothing crosses a layer without passing through the controls in the layer above it.
02Flagship agent

Operations Assistant

The orchestration layer. It holds the operational context, decides which specialist agent or tool the request belongs to, and stays responsible for the request until it is resolved or escalated to a person.

Working prototype · in development

Most operational questions do not belong to one system. "Why is this job still open?" touches the work order, the parts, the supplier and the customer. The assistant's job is to hold that question, gather what it needs across systems, and come back with an answer a person can act on — or an action already prepared for approval.

What it does
  • Query operational systems
  • Retrieve context
  • Call multiple tools
  • Identify exceptions
  • Create tasks
  • Check statuses
  • Produce summaries
  • Initiate approved workflows
  • Route to specialist agents
  • Escalate unresolved issues
  • Human handoff
Architecture

Supervisor pattern over specialist agents, with a shared tool and control layer. Context assembled per request from permissions, role and the systems in scope; orchestration concepts modelled on LangGraph-style state graphs.

PythonFastAPIPydanticTool callingMulti-agent orchestrationRAG
Inventory agent · sample run
Illustrative run against sample data

The agent did not calculate the reorder point or decide the spend. It gathered the operational facts, applied the rule where the rule lives, prepared the request and stopped at the approval boundary.

03Agent portfolio

Eight agents across the operational surface

Each one solves a specific operational problem and sits on the same tool and control layer.

Maturity disclosure

These are working prototypes in active development, built and run outside client environments. None are presented as deployed production systems, and none hold final approval authority. Demos are being prepared; where a demo is not yet available the card says so.

Inventory Agent

Prototype

Stock questions get answered by whoever can open the system and read it correctly. The agent makes the same answer available to anyone who is allowed to ask.

  • Live inventory queries
  • Stock availability
  • Low-stock detection
  • Reorder calculations
  • Approved stock movements
  • Supplier lookup
  • Purchase request preparation
  • Alerts
  • Approval workflow
  • Human handoff
Architecture

Quantities, movements and reorder rules resolved by database queries and deterministic tools. RAG reserved for SOPs, product information and supplier policy. Stock movements require approval before they are written.

PostgreSQLFastAPITool callingRAGApproval engine
View architecture Demo coming soon

Procurement Agent

Prototype

Purchase requests stall in the gap between the person who needs the part and the person allowed to commit the spend. The agent prepares the request properly and routes it to the right approver.

  • Purchase-request preparation
  • Supplier lookup
  • Procurement status
  • Approval routing
  • Purchasing information
  • Policy retrieval
  • Supplier comparison
  • Exception escalation
Boundary

Automated assistance and final approval authority are deliberately separate. The agent assembles, compares and routes; a person approves. No commitment is created without that step.

Business rulesApproval routingRAGREST APIs
View architecture Demo coming soon

Maintenance / Asset Agent

Prototype

A technician standing in front of a failed asset needs its history, its spares and the right procedure — not a login to four screens. Built directly on the SAP EAM process knowledge behind it.

  • Asset information lookup
  • Maintenance history
  • Work-order information
  • Notification support
  • Preventive-maintenance context
  • Spare-parts lookup
  • Procedure retrieval
  • Escalation
  • Technician support
Scope boundary

Read and assist, grounded in maintenance data structures and procedures. Direct write-back into SAP is not claimed — it is a controlled integration decision, not a default.

SAP EAM domainRAGTool callingEscalation
SAP process context Demo coming soon

CRM / Revenue Operations Agent

Prototype

Pipeline questions and follow-ups that normally depend on someone remembering. The agent reads the CRM, qualifies against defined criteria and creates the record and the task.

  • Lead lookup
  • Pipeline information
  • Follow-ups
  • Lead qualification
  • CRM record creation
  • Task creation
  • Sales workflow actions
Architecture

CRM APIs for read and write, qualification criteria held as explicit rules rather than prompt text, and workflow actions triggered only for states the rules recognise.

CRM integrationsWebhooksn8nTool calling
Voice intake variant Demo coming soon

Finance / Collections Agent

Prototype

Outstanding payments usually sit in a spreadsheet someone rebuilds each month. The agent assembles the account context and prepares the follow-up.

  • Invoice lookup
  • Outstanding-payment status
  • Customer account context
  • Follow-up preparation
  • Collections workflow support
  • Escalation
Boundary

No autonomous financial approval authority, no payment actions, no write-offs. Balances and ageing come from finance systems; the agent prepares and escalates only.

Finance system APIsBusiness rulesAudit trail
View architecture Demo coming soon

Knowledge / SOP Agent

Prototype

Policy and procedure answers that cite where they came from, and that respect who is allowed to see the document in the first place.

  • Policy search
  • SOP retrieval
  • Procedure questions
  • Document-grounded answers
  • Source-aware responses
  • Permission-aware retrieval
Architecture

Embeddings and vector search over controlled document sets, with permissions applied at retrieval time rather than filtered out of the answer afterwards. Every response carries its source.

EmbeddingsVector searchPermission-aware RAGCitations
Retrieval foundations Demo coming soon

Customer Support / Request-to-Resolution Agent

Prototype

Intake, classification and routing done consistently, so the request reaches the person who can actually resolve it with the context already attached.

  • Request intake
  • Classification
  • Context lookup
  • Routing
  • Response generation
  • Escalation
  • Resolution workflow
  • Ticket & status updates
Architecture

Classification against a defined taxonomy, context assembled from the customer record and job history, and escalation paths that fire on rules rather than on the model's confidence.

Workflow platformsREST APIsRAGRouting rules
View architecture Demo coming soon

Operations Assistant

Prototype · flagship

The supervisor. Holds the request, routes it to the specialist that owns it, and remains accountable for resolution or escalation.

  • Multi-tool orchestration
  • Exception identification
  • Task creation
  • Status checks
  • Summaries
  • Approved workflow initiation
  • Specialist routing
  • Human handoff
Architecture

Supervisor over specialists, shared control layer, per-request context built from role and permissions.

Full detail & sample run Demo coming soon
04Technical stack

What the agents are actually built on

AI & agent layer
LLMsTool callingRAG

Embeddings, vector search, context engineering, multi-agent orchestration, LangGraph-style state graphs where the flow warrants them.

Backend
PythonFastAPIPostgreSQL

Pydantic for typed contracts at the boundary, SQLAlchemy for the data layer, and endpoints that behave the same way every time they are called.

Integration
n8nMakeRESTWebhooks

CRM and ERP integrations, event-driven handoffs, and the unglamorous plumbing that decides whether any of this reaches the business.

Governance & control
ApprovalsPermissionsAudit

Human-in-the-loop by design, validation rules, deterministic business logic, and an audit trail that records what the agent called and what it was told.

05Why the enterprise background matters

Agents are only as good as the systems underneath them

Step 01 · Enterprise systems

SAP, ERP and CRM already hold the operational truth — equipment, work orders, stock, suppliers, customers, money.

Step 02 · Structured data & rules

Those systems produce structured data and enforce rules. That is what makes an answer checkable instead of plausible.

Step 03 · Controlled tools

A tool layer exposes exactly what an agent may read and write — no more — with permissions and validation at the boundary.

Step 04 · Agents

Specialist agents reason over that context and prepare work: a request, a summary, a task, an exception worth someone's attention.

Step 05 · Approval

Anything consequential stops at a person with the authority to say yes, and the context they need to say it quickly.

Step 06 · Business action

The action is written back into the system of record, with the trail of how it got there.

See the enterprise process it sits on GitHub Discuss a system