Agents that work inside a governed model of the business
AskUnderstandDecideAnalyseCheckActRecord
Most enterprise agents are connected to data. A SCIKIQ agent operates inside a model of how the business works: which entities exist, how they relate, which rules apply, who may approve what, and what must be recorded. That is what makes its answers explainable and its actions safe to allow.
01 Architecture
Three tiers, one owner of meaning
SCIKIQ business layerWhat makes the answer right
Business contextOntology and knowledge graphDecision modelsPolicy and accessAgent factory
The part that is specific to your enterprise. It defines the question, the facts, the rule and the boundary.
↓
Agent runtimeWhat keeps agents running
Agent sessionsMemoryTool connectionsSub-agents and workflowsDurable execution
Interchangeable: SCIKIQ’s own runtime, or an agent runtime your organisation already operates, reached as a service. The business layer above does not change when the runtime does.
↓
Enterprise systemsWhere actions land
SAPSalesforceSnowflakeERP, CRM and workflow APIs
Actions reach these systems only through the governed path, never directly from a model.
02 The governed path
From a question to an approved action, in eight steps
1
Business intent
A person asks in business language. The question is resolved to a domain, the entities it concerns and the decision it is really asking for.
The knowledge graph returns the relevant entities, their relationships, the live KPIs and the rules that apply, already filtered to what this person may see.
A deterministic decision model evaluates the rule and returns a status, a severity, the reasons and a recommended action. The verdict is the business’s rule, not a model’s guess.
Illustrative: the supplier, material and figures are invented to show the flow.
The question
“Which suppliers could impact our European Q4 production?”
Asked by a supply chain manager
What the knowledge graph returns
SupplierABC Components, GermanyLead time 45 daysOn-time delivery 82%
MaterialHydraulic valveDemand 120 a dayOn hand 3,200Safety stock 5,000
PlantGermany plant, EuropeUses the hydraulic valve
ABC Components supplies the hydraulic valve, which is used by the Germany plant.
The decision
Rule: Shortage if stock is below safety stock and days of cover are shorter than the supplier’s lead time.
Supply shortageAt riskSeverity: High
Stock of 3,200 is below the safety stock of 5,000.
About 27 days of cover against a 45-day supplier lead time.
What the specialist agents add
Demand agent
Production needs about 120 valves a day through the quarter.
Inventory agent
Current stock covers roughly 27 days.
Supplier agent
A 45-day lead time and 82% on-time delivery: this supplier needs attention.
Finance agent
Expediting adds logistics cost, to be weighed against the cost of a stopped line.
Policy
Supply chain managers and procurement directors may raise this action. It requires approval.
Action
Purchase requisition for 5,000 valves raised in SAP, status pending approval. Nothing is ordered until a person approves it.
On the record
Execution ID · Question and role · Entities used · Decision and reasons · Policy result · Action raised
04 Across industries
The same path, four more industries
The steps do not change from one industry to the next. What changes is the ontology, the rule and
the policy. Each example below ends differently: an action held for a committee, a pause that only an investigator
can turn into a decline, an action released automatically within policy, and a consent check before a customer is contacted.
Illustrative: names and figures are invented to show the flow.
Banking
“Which corporate borrowers are drifting toward a covenant breach this quarter?”
Asked by a credit risk manager
Context
Borrower: Nordal Logistics, €40m term loan
Covenant: net debt to EBITDA no higher than 3.5x
Latest ratio 3.3x, rising for three quarters
Relationship: refinancing talks due next month
Rule
Early warning if the ratio is within 10% of its covenant limit and has risen for two quarters or more.
Decision
Early warningSeverity: Medium
Agents
Credit Ratio trajectory crosses 3.5x within two quarters at the current trend.
Relationship Refinancing talks are a natural moment to discuss headroom.
The shortage call comes from a rule the business owns and can read. Language models summarise and add perspective; they do not change the verdict.
2
Context before reasoning
An agent only sees entities resolved through the ontology and permitted to the person asking. It reasons over governed facts, not over whatever a search returned.
3
Policy before any write
The governance check sits between the recommendation and the action, so a persuasive answer can never bypass an approval rule.
4
Approval is a state
Sensitive actions are created as pending, visible in the target system, and wait for a person. Automatic execution is reserved for actions inside policy.
5
Every run is on the record
The audit record links the question to the data, the rule, the policy and the action, which is what a reviewer or a regulator will ask to see.
6
The runtime is replaceable
Sessions, memory and tool execution can come from SCIKIQ or from an existing runtime. The business layer, which is the part that is hard to rebuild, stays the same.
06 Creating an agent
What a business team defines
An agent is defined by the business it serves, not by the model behind it. A team names the
domain, the context it may use, the decisions it supports, the systems it can reach and the policies that
bound it. Every question and every action then runs through the same governed path:
ontology → knowledge graph → decision → agents → policy → approval → action → audit.
An AI agent that works inside a model of the business rather than on raw data. It takes its facts from the knowledge graph, its verdict from a decision rule the business owns, and its permission to act from policy, and it records every run.
Does the language model decide what happens?
No. The decision model evaluates the business rule and returns the status and the recommended action. Agents explain and add perspective. Policy then decides whether the action may be raised and whether it needs approval.
Can we use an agent runtime we already run?
Yes. The runtime that handles sessions, memory and tool execution is a replaceable layer. SCIKIQ provides the business context, decisions, policy and audit above it, and those stay the same whichever runtime sits underneath.
What happens when an action needs approval?
It is created in the target system as pending, for example a purchase requisition awaiting approval, and nothing executes until an authorised person approves it. The approval is part of the audit record.