The Change Engine · Advanced · E4
Supervising AI Decisions
Agent oversight, human-in-the-loop design, managing autonomous systems
- Duration
- 2 days · 9:00–17:00 each day (13h contact, incl. breaks + lunch)
- Participant level
- Advanced: experienced practitioners
- Format
- In-person facilitated workshop
- Participants
- 12 included · 20 maximum (flexible group size)
- Prerequisite
- E3 (or equivalent)
- Ecosystem
- Your existing tools: exercises and demos on your real stack
- Price
- EUR 8,500 (EUR 450 per extra participant)
Amounts in EUR, excl. VAT.
ai-generatedWhat participants will be able to do
- Apply a risk-based framework to evaluate AI outputs
- Design human-in-the-loop checkpoints for your team's AI uses
- Build an audit trail for AI-assisted decisions
- Respond to an AI failure or incident using a structured playbook
Tools & resources
What changes after this module
Managers can tell tools from agents, set the boundaries and human-in-the-loop checkpoints an agent needs, monitor its performance, and run an incident when it goes wrong, leaving with a working supervision model for one real agent use case.
Who should attend
Middle managers. Advanced: experienced practitioners.
Programme
This agenda is indicative. Content, sequencing, and examples are adapted to your team's context, tools, and objectives.
Before you start
- Inventory of AI/agent use in your area (official or shadow)
- One decision currently delegated to a system
- Your escalation or incident procedure if one exists
- E3 or equivalent baseline
| Day 1 | 9:00–10:30 | What AI agents are: Tools assist; agents actThe autonomy spectrum from copilot to fully autonomousWhy this changes the manager’s job: you supervise outcomes and exceptions, not keystrokesTreating an agent as a headcount decision, not a software purchase (who owns it, who is accountable). Hands-onThe group classifies a set of real workflows on the autonomy spectrum and debates, for one, whether it should be a tool, a supervised agent, or off-limits, and who would own it. |
| 10:30–10:45 · ☕ Coffee break | ||
| Day 1 | 10:45–12:30 | Setting boundaries: What an agent may decide vs must escalateApproval thresholds, exception handling, and kill switchesDesigning the guardrails before you deploy. Hands-onEach manager writes the decision boundaries for one real agent use case: the approve/escalate thresholds and the kill-switch condition. |
| 12:30–13:30 · 🍽 Lunch break | ||
| Day 1 | 13:30–15:30 | Human-in-the-loop design: Where a human checkpoint genuinely reduces risk vs where it just slows things downAutomation bias: why reviewers rubber-stamp: and checkpoint designs that stay awake (a duty the EU AI Act’s Article 14 names explicitly)Escalation matricesDesigning HITL for real scenarios. Hands-onWorkshop: managers design human-in-the-loop processes for three common scenarios and build the escalation matrix for their own use case. |
| 15:30–15:45 · ☕ Coffee break | ||
| Day 1 | 15:45–17:00 | Agent governance frameworks: The federated model: your domain owns day-to-day oversight of the agent while a central team owns the platform, guardrails, and audit trailAccountability chains and documentation. Hands-onManagers map the accountability chain for their agent use case (who owns what between the team and the centre) on a one-page governance sheet. |
| Day 2 | 9:00–10:30 | Monitoring agent performance: What to watch: quality metrics, drift, alert thresholds, and costDashboards a manager will actually read. Hands-onEach manager defines the monitoring metrics and alert thresholds for their agent and sketches the dashboard. |
| 10:30–10:45 · ☕ Coffee break | ||
| Day 2 | 10:45–12:30 | Workshop: build a monitoring plan: Assembling monitoring into an operating routine for a real scenario. Hands-onManagers build a complete monitoring plan for their agent use case: metrics, alerts, and review cadence. |
| 12:30–13:30 · 🍽 Lunch break | ||
| Day 2 | 13:30–15:30 | Incident response: When agents go wrong: and they willPlaybooks, communication templates, and a blameless post-mortem processGartner estimates 40%+ of agentic projects will be cancelled by 2027; disciplined supervision is how you stay in the other 60%. Hands-onTabletop exercise: the group runs a simulated agent failure and each manager drafts the first-hour response and the stakeholder message. |
| 15:30–15:45 · ☕ Coffee break | ||
| Day 2 | 15:45–17:00 | Ongoing supervision model: Turning it into a rhythm: weekly reviews, monthly audits, quarterly governance. Hands-onEach manager assembles their ongoing supervision model (cadence + owners) and presents it in trios for challenge. |
Deliverables
- Decision-boundary spec (one agent use case)
- Human-in-the-loop + escalation design
- Agent governance / accountability map (federated model)
- Agent monitoring plan
- Incident-response playbook
- Ongoing supervision model
Interested in running this module for your team? Get in touch and we'll tailor the format, dates, and delivery to your context.
Request this trainingYour trainer
Senior expert
Has designed human-in-the-loop controls and oversight for autonomous systems
Discovery Session
This module requires a full day to deliver a meaningful learning experience. We only offer it in its complete format.
Frequent questions
Can modules be taken individually?
Yes. Every module stands alone at a fixed price, and every module counts toward a programme if you continue.
Where does training happen?
At your premises or remote, on your dates, for private cohorts. Open sessions run on a fixed monthly calendar.
Which AI tools do you train on?
Yours. Every module ships in four ecosystem editions and runs its exercises on your real stack.
Who delivers?
Inforca's senior consultants and trainers. Flagged modules and the executive track are delivered at senior-expert level.
Discuss this module in a First Call: fit, dates, and the path around it.
No commitment · our team responds within one business day