It reflects the actual users, data, tools, actions and consequences of the deployed system.
Approach
Trust is a property to demonstrate—not a label to apply.
Operisys approaches AI assurance as a disciplined argument supported by evidence: this system, for this purpose, within these boundaries, has behaved well enough to justify this level of reliance.
Assurance does not eliminate uncertainty. It makes uncertainty, control and the basis for a decision visible to the people who remain accountable.
How the models connect
One model for the control. One cycle for the work.
The Control Model identifies what has to be governed. The Assurance Cycle is how Operisys examines it. Each pass through the cycle tests all four control dimensions.
01 · What is governed
Control Model
- Purpose
- The intended use, accountable owner and legitimate outcome.
- Authority
- What the system may access, decide and execute—and where approval is required.
- Behaviour
- How the deployed system performs under normal, difficult and failed conditions.
- Evidence
- What was designed, tested, approved, observed and changed.
02 · How it is examined
Assurance Cycle
Define
Bound
Test
Observe
The depth of work follows the consequence, whether the system is a narrow assistant or an agent able to take operational action.
- 01
Define
Frame the intended use, affected people, accountable owner, operating context and consequence of error. Map the model, data, tools, vendors, workflow and human dependencies as one system.
- 02
Bound
Specify what the system may access, recommend, decide or execute—and when a person must take control.
- 03
Test
Evaluate normal tasks, edge cases, misuse, tool failure, ambiguity, escalation and recovery under realistic conditions.
- 04
Observe
Record the deployment basis and residual uncertainty, then monitor the signals, incidents and system changes that trigger reassessment.
What counts as evidence
Useful evidence is specific enough to change a decision.
It connects an organisational requirement to a control and to observable system behaviour.
Another reviewer can understand the conditions, method, result and material limitations.
Its validity is reconsidered when the model, vendor, data, permissions or operating context changes.
Working definitions
Precise terms for practical decisions.
- AI assurance
- An evidence-based examination of whether an AI system is suitable for a defined use and remains within its intended boundaries.
- Deployment assurance
- Evaluation of the assembled system—model, data, tools, permissions, workflow and people—in its intended operating context.
- Agent assurance
- Examination of an AI agent’s tools, access, actions, limits and routes back to human control.
- Authority controls
- Enforced rules that limit what an AI system may access, decide or execute, including when human approval is required.
Operating principles
The positions that guide the work.
Authority before autonomy.
Define what a system may do before optimising how independently it can do it.
System over model.
Assess the model together with its data, tools, permissions, workflow and users.
Evidence over confidence.
Use observable tests and records instead of relying on fluency or vendor claims.
Review follows change.
Reopen the assurance position when a change invalidates the evidence beneath it.
Working with Operisys
Clear boundaries apply to the engagement too.
Scope, access, data handling, third-party services and evidence requirements should be agreed before sensitive information is introduced. Where a question or specialist need sits outside the agreed scope, that boundary should be explicit.
Operisys does not claim that governance removes AI risk. The work is designed to make decisions, limits, behaviour and residual uncertainty clear enough for accountable people to act.
Book a call