Proposal: purchase order for 120 units
Review and approve →KnowOS add-on · Agents on real data
AgentsThey detect, alert and propose the action.
AI agents connected to your real ERP and eCommerce data. They watch stock, orders, margins and returns, and propose what to do. With their own token and always with approval.
Concentrated in one batch and one carrier
See detail →Off-list discounts on 4 orders
Notify finance →The challenge
Information arrives late. When it arrives, it is already an incident.
The data to anticipate exists. It is in the ERP, the store and the CRM. What is missing is someone watching it continuously.
The customer discovers the stock-out
The product runs out on a Friday. On Monday the complaints arrive.
Margin drops without anyone seeing it
Off-list discounts that only show up at month-end close.
Returns spike
A batch or a carrier with problems, detected weeks later.
Reports are built by hand
Every Monday, someone crosses three exports to know how the week is going.
A team cannot watch thousands of references, orders and customers at once. An agent can, and it alerts you when something deserves attention.
What really matters
An agent is not a chat that answers. It is a process that watches while the team works.
It reads real data with its own permissions, detects what falls outside the norm and proposes the action. The decision remains with a person.
Integrafy Agents
Agents that access your real data, not generic AI.
Agents is built on KnowOS: it uses your business ontology to understand what a customer, an order or a margin is, and the permission engine to know what it may see and do.
That is why it is contracted as a KnowOS add-on. Without that map, an agent would have no context to act on safely.
In one sentenceDetect, alert and propose.
Where Agents sits in Integrafy Platform
Agents read through KnowOS and, when an action is approved, write to the ERP through Connect and Flow.
The cycle of an agent
An agent is just another user. With its token, its limits and its log.
The ontology does not distinguish between a person and an agent: both ask permission to read and to act.
Every action goes through the same cycle, and none is executed without approval from a person with permission.
AgentSame permission engine
- 01 · DetectsReads stock and consumption with its token
- 02 · ProposesPurchase order as a draft
- 03 · ApprovesA person with permission
- 04 · ExecutesWriteback to the ERP via Connect and Flow
- 05 · LogsObserve records what, who and when
Agent types
Six agents to start. And more can be created.
They also answer natural-language questions on the same data. No technical skills needed to use alerts and dashboards.
Critical stock
Detects references below minimum before reaching zero.
Order anomalies
Unusual orders by amount, quantity or customer.
Margin drops
Off-list discounts or rising costs not passed on.
Return spikes
Concentrations by batch, family or carrier.
Replenishment suggestions
Based on sales velocity and seasonality, as drafts.
Automatic reports
The weekly summary someone builds by hand today, in their inbox every Monday.
Each agent, its token · illustrative example
Each agent sees what it needs. And nothing more.
The scope of each agent is defined like a person’s: what it reads, what it may propose and what stays out of its view.
| Agent | Reads | Can | Cannot |
|---|---|---|---|
| Replenishment agent | Stock, consumption, supplier lead times | Propose purchase orders as drafts | Margins, selling prices, customers |
| Margin agent | Sales, costs, applied discounts | Notify finance and the account rep | Change prices or terms |
| Returns agent | Returns, batches, carriers | Open an incident and propose a cause | End-customer personal data |
| Catalogue agent | Sheets, translations and images in Data | Weekly report to the owner | Publish changes to the store |
Step by step
What happens when stock drops below minimum.
ERP real data
Agent with its token
KnowOS context · permissions
Owner person
ERP writeback
1 The stock of a reference crosses the minimum
2 Queries sales velocity and seasonality
3 Returns the data its token allows it to see
4 Calculates the quantity and prepares the proposal
5 Sends the purchase order as a draft
6 Approves or adjusts; it is written to the ERP
7 Confirms and closes the alert
Observe · records who detected, who approved and what was written · the person decides, the agent prepares
Endpoint · MCP · per-token permissions
Connect your AI, whichever it is. Agents follow the same rules.
Who asks · people, models and agents
ChatGPTClaudeGeminiCopilot · TeamsOwn agentsn8n · ZapierInternal apps
One endpoint and one MCP serverREST API · GraphQL · MCP — the same contract for everyone
Switch models without touching the integrationPermission engineEvery request carries a token · the token decides the scope
Rolesales, finance, agent
Rowsonly their portfolio or area
Columnsmargin and cost hidden
Actionsread, propose or execute
OntologyThe tools the model sees have business names
find_customer()orders_of(customer)current_price(product, customer)search_documents(text)propose_discount() · actionRelational dataERP · CRM · WMS · PIM
Vector dataDocuments · email · audio
Audit · who asked, with which token, which tool they used and which data they received
Security
Rules of use that are not negotiable.
Read-only by default
Agents read. They only act if expressly authorised.
Nothing is written without review
Every action stays as a draft until a person approves it.
Same rules as a person
Scoped token: role, rows, columns and actions.
Everything is logged
What it detected, what it proposed, who approved and what was written.
Model-agnostic
The model is chosen by cost and case. MCP-compatible.
Your data stays in
In on-premise deployments, everything runs on your infrastructure.
Comparison
A threshold alert, or an agent with context.
Many systems already alert when something crosses a number. The difference is what it understands and what it proposes.
What it detects Alerts and fixed rulesA value crossing a fixed threshold. Integrafy AgentsPatterns with context: seasonality, customer, family, history.
What it does Alerts and fixed rulesAlerts. Integrafy AgentsAlerts, explains the cause and proposes the action with the data.
Context Alerts and fixed rulesA table or a report. Integrafy AgentsThe full ontology: orders, customers, price lists and documents.
Configuration Alerts and fixed rulesTechnical rules someone maintains. Integrafy AgentsDescribed in natural language and tuned with use.
Security Alerts and fixed rulesTechnical credential with broad access. Integrafy AgentsToken with its own scope, like a person.
Traceability Alerts and fixed rulesAn email in an inbox. Integrafy AgentsLog of detection, proposal, approval and execution.
Where it grows
Assisted financial management
Real-time control, cash alerts and profitability projections by family and market.
Assisted sales management
Sales analysis, cross-selling detection and predictive pipeline management.
Assisted operations
Process optimisation, intelligent stock management and logistics efficiency.
How it is contracted
Agents adds to KnowOS. Agent by agent.
First the map, then the agents. Each agent is sized by what it watches, how often and what it may propose.
Integrafy KnowOS
The ontology, per-token permissions and the context the agents work on.
from €190/monthAgents
Each agent with its scope, its token and its log. Activated or withdrawn whenever you need.
per agentDiagnosis
30-minute session to decide which agents add the most in your operation.
freePrices exclude VAT. Full details in the Integrafy Platform price list.
See the full price list →Frequently asked questions about Integrafy Agents
Do agents modify data or only read it?
They read by default. Any action (proposing a purchase order, opening an incident, notifying finance) stays as a draft until a person with permission approves it. Only then is it written to the ERP through Connect and Flow, and Observe logs who approved and what was written.
Why does Agents need KnowOS?
Because an agent without a map has no context. KnowOS provides the ontology (what a customer, an order, a margin is) and the per-token permission engine. On top of that, each agent knows what it may read, what it may propose and what stays out of its view.
Which AI model do the agents use?
They are model-agnostic. The model is chosen by cost and case (Claude, GPT, Gemini or local models) and swapped without touching the integration, because agents talk to the data through the same endpoint and MCP server as people do.
Does my data leave my company?
The endpoint and the MCP server run in your deployment. On-premise, everything runs on your infrastructure. The model only receives the data the agent’s token allows for each request, and nothing is used for training.
Do I need technical skills to use the agents?
No. Alerts arrive by email, Teams or Slack with the proposal and the link to review and approve. Agents also answer natural-language questions on the same data. Each agent’s scope is described in natural language and tuned with use.
How much does an agent cost?
Agents is priced per agent as a KnowOS add-on (from €190/month). Each agent is sized by what it watches, how often and what it may propose, after a free 30-minute diagnosis.
Session on your case
Tell us what you would like to know sooner. We show you the agent that watches it.
A stock-out, a margin drop, a customer who stops buying. In 30 minutes we see which data is needed and which agent makes sense to start with.
Book a session