How it works · Flow F15
When an agent gets it wrong: contain, reverse, explain, fix
Detection, severity, containment with kill switches, reversal or compensation per reversibility class, notification, root cause by replay, new evals and a post-mortem.
What the AI incident flow does
No AI is perfect. One3 plans for incidents in advance, so they are contained fast and explained honestly.
An incident is detected by a guardrail event, a customer complaint or an anomaly. It is classified by severity. Containment follows: a kill switch, demotion or token revocation. Affected Actions are reversed or compensated according to their reversibility class — the UI never promises an undo that is not real. The tenant is notified, and authorities too if a personal-data breach requires it. Root cause is found by replaying the run. The fix ships with new eval cases, a post-mortem is written and the learning goes to Knowledge.
Kill switches K1/K2 and the incident workflow are planned for W1; the Agent Liability Desk as a full app is W4.
Every step on this page is an Action: a typed verb with a risk tier (R0 read-only to R4 irreversible or financial), a permission check and a reversibility class. Agents and humans call the same Actions, so the same approval rules apply whether a person clicks a button or a Digital Employee acts.
AI incident, step by step
Nine steps from detection to learning.
- 1
1. Detection
**AI Workforce · AI Governance · AI Observability.** Guardrail events, complaints or anomalies.
- 2
2. Severity classification
**AI Governance.** S4 low to S1 critical.
- 3
3. Containment
**AI Governance · Identity & Access.** Kill switch, demotion or token revocation.
- 4
4. Reverse or compensate
**Platform · Action Bus.** Blast-radius query: every object changed by the agent in a window, then reverse or compensate per class.
- 5
5. Notify
**Platform · Notifications.** The tenant is told; authorities too if a personal-data breach requires it (verify).
- 6
6. Root cause by replay
**AI Observability · forensic replay.** Deterministic replay with cached tool responses.
- 7
7. Fix and new eval cases
**AI Workforce · Evaluation.** The fix ships with eval cases that would have caught it.
- 8
8. Post-mortem
**Human.** Mandatory for S1.
- 9
9. Knowledge
**Intelligence · Knowledge.** The learning is stored.
Severity and automatic response (defaults)
| Severity | Examples | Automatic response |
|---|---|---|
| S4 low | Style violation; minor slip caught by review | Log; add eval case |
| S3 medium | Wrong external message sent (reversible); budget overrun | Demote that skill one level; notify owner |
| S2 high | Contact outside window; missing disclosure; data shown to wrong internal user | Suspend skill and external channels for this agent |
| S1 critical | Irreversible financial or legal action outside policy; PII leak; impersonation | Kill switch for this agent; freeze related Actions; tenant alert |
Reversibility classes
| Class | What the UI promises |
|---|---|
| fully_reversible | Reverse |
| compensatable | Compensate |
| externally_irreversible | Can't be reversed. Follow-up options available. |
| legally_irreversible | Legally binding. Only formal correction processes apply. |
| privacy_impact_irreversible | Disclosure can't be withdrawn. Incident process applies. |
Apps involved
Each step runs as a typed, permissioned Action inside one of these apps. Nothing happens outside the Action contract.
AI Workforce
AI Governance holds incidents and kill switches; Agent Liability Desk handles claims and remediation in W4.
Learn morePlatform
Action Bus reverses or compensates; Trust & Audit keeps the hash-chained record; AI Observability supports replay.
Learn moreLegal, Risk & Compliance
Breach procedures and regulator notification where required.
Learn moreIntelligence
Knowledge keeps post-mortems.
Learn moreQuestions about this flow
Can One3 undo everything an agent did?
No, and it will not pretend to. Each Action declares a reversibility class. Internal changes can be reversed; sent messages and payments can only be compensated.
How fast is containment?
K1 and K2 kill switches target under 5 seconds *(estimate, verified in quarterly drills)*.
How do you find the root cause?
Every run stores context references, model and prompt versions, tool inputs and outputs and policy decisions, so it can be replayed.
Will I be told?
Yes. Tenants are notified, and authorities too where a personal-data breach requires it.
What stops it happening again?
Every fix ships with new eval cases, and the agent re-enters shadow mode before regaining autonomy.
See how One3 handles failure
Ask us for the incident and kill-switch design before you rent an agent.
Apps involved
Where this happens in One3
Platform
One governed foundation for every human, AI agent, integration and channel
ExploreLegal, Risk & Compliance
Contracts, consent, controls and risk as governed work on the data you already run
ExploreIntelligence
Cited answers, one set of metrics and decisions tracked to a measured result
ExploreAI Workforce
Rent ready-made AI Digital Employees that work for your team, with a monthly ROI statement
ExploreMore in How it works
- Tenant onboarding
- Lead to cash
- Campaign loop
- Issue to resolution
- Hire to retire
- Procure to pay
- Order to delivery
- Record to report
- Idea to release
- Goal to outcome
- Agent lifecycle
- Model improvement
- Pack upgrade
- DPDP data-principal requests
- Rent a Digital Employee
- Switch to One3 in a weekend
- Monthly Agent Investment Statement
- Goal to verified statement