Agents under human control
How can you prototype a useful agent without delegating sensitive decisions to it?
3500 € per session, excl. VAT
In short
Over two days, comprising 14 hours excluding breaks, you design an agent prototype capable of preparing an action, requesting human approval and logging its simulated execution. You define its permissions, build its checkpoints and measure its quality against documented scenarios before preparing a limited pilot.
Why this training?
An agent that links reading, decision-making and action can propagate an error beyond a simple incorrect response. Oversight must therefore intervene before sensitive commitments are made, with limited permissions. Human approval is only useful if the person responsible understands the proposal, its sources and its uncertainties. The programme translates this requirement into executable checks and test evidence. Running case study: Avelcore Équipements, a fictitious distributor, receives product return requests. The agent consults a fictitious policy, prepares a recommendation and seeks a responsible person's approval before any return is authorised. Incomplete, contradictory or suspicious cases must be escalated, without contacting any real customer. Transfer to the workplace: Adapt the prototype to an authorised process, have the permissions and data approved, then propose a bounded pilot with a designated person responsible, acceptance thresholds and a stop procedure.
Objectives of this training
- Define the boundaries of autonomous, supervised and prohibited actions
- Model approval and escalation states
- Build a prototype with minimal permissions
- Test errors, refusals and bypass attempts
- Measure quality and formalise a limited pilot
Who is it for and what are the prerequisites?
Roles and tasks concerned
- Automation managers and AI project managers
- Subject matter leads and designers of supervised processes
- Developers or integrators responsible for a prototype
Prior skills
- Be able to structure and evaluate generative AI instructions
- Understand a business process and its decision-making risks
- Be able to configure an automated workflow or read a simple script
What you will work on
Set boundaries for delegation
Objective: Define the agent’s mandate. Agent, workflow, autonomy and sensitive commitments.
Map the risks
Objective: Assign responsibility for each control. Minimal permissions, separation of roles, escalation.
Model the states
Objective: Make the process flow explicit. Waiting, approval, rejection, expiry and resumption.
Structure the inputs
Objective: Constrain the information used. Authorised sources, mandatory fields, untrusted data.
Assemble the prototype
Objective: Produce a traceable recommendation. Instructions, document consultation and simulated actions.
Set up approval
Objective: Block sensitive actions. Sources, rationale, uncertainties and the supervisor’s choices.
Secure resumption
Objective: Prevent improper execution. Versioned approval, expiry, duplicate prevention.
Prepare the tests
Objective: Establish the expected outcomes. Standard, ambiguous, incomplete and prohibited cases.
Put the safeguards to the test
Objective: Detect bypasses. Document-based injection, prohibited access, absent supervisor.
Measure and correct
Objective: Compare results with expectations. Accuracy, relevant escalations, blocked actions, human corrections.
Organise pilot operations
Objective: Prepare sustainable oversight. Logging, retention, alerts and manual shutdown.
Decide on the pilot
Objective: Justify a limited experiment. Acceptance thresholds, responsibilities and remaining limitations.
What you take away
- Permissions and responsibilities matrix
- Supervised prototype and documented configuration
- Human approval dossier and execution log
- Test report with quality measurements
- Pilot plan with stopping conditions
Method and assessment
- Short teaching inputs followed by immediate practical work on the case study used throughout.
- Individual work followed by work in pairs, with peer review of the outputs.
- Guided group discussion: what works, what needs correcting and what remains a human decision.
- Preparation: Check access to a prototyping environment authorised by the organisation Prepare a procedure and fictitious or anonymised files, without sensitive data
Observable criteria: Every sensitive action remains blocked without explicit approval ; Each human decision is linked to a versioned proposal ; Tests cover refusal, expiry, injection and resumption ; Accuracy and escalation rates are calculated on the tested dataset ; The pilot specifies responsible parties, scope and stopping thresholds.
Frequently asked questions
Does the prototype act on real systems?
No, actions are simulated in an isolated environment.
Is programming required?
Not necessarily; the configuration depends on the authorised tools and their verified capabilities.
Who approves the proposals?
A designated person with access to the sources and the authority to refuse.
Can you move to full-scale production immediately afterwards?
No; security, compliance and integrations require additional validation.
Last updated: