Human oversight of AI
The H.A.D.A. methodology helps you decide when AI may act on its own, when a human decision is needed and how to document that decision. Operations are assigned to one of four zones, from A to D.
- Intent
- Input control (applies)
- Authorization (applies)approves the operation
- AI
- Result validation
- Human
- Evidence (applies)
Highlighted: the points of the flow that this description refers to. Orange marker: a human decision at this point.
What problem the methodology addresses
The principle “a human oversees AI” does not by itself say who should check what, and when. Without dividing operations into classes, oversight turns into a formality or misses the situations where an error costs the most.
Too many approvals
When every response requires approval, approving becomes a reflex rather than a decision.
Too little control
Without thresholds, sensitive data or actions with serious consequences go through without a human decision.
No trace of the decision
When a decision is not recorded, it is hard to establish later who made it and on what basis.
What the procedure looks like
The four H.A.D.A. steps lead from setting the human decision threshold to recording who made the decision and on what basis. The letters of the acronym come from the English names of the steps.
Step 1 · H (Human Threshold)
Human decision threshold
Decide in which situations an operation requires a human decision, for example because of the type of data, the recipient of the result or the consequences of the action.
Step 2 · A (Autonomy Spectrum)
Scope of autonomy
For each process, define whether AI only makes suggestions, prepares a result for approval, or performs designated actions on its own within agreed boundaries.
Step 3 · D (Decision Class)
Operation classes in zones A to D
Assign operations to one of four zones, from A to D, in order of increasing data sensitivity and impact. For each zone, set the handling: sending unchanged, restricting the content, a human decision or holding.
Step 4 · A (Accountability)
Accountability
Name who makes the decisions in each zone and what is recorded: the zone, the decision, the person and the time. The organization sets the evidence retention period.
Human decision before a query is sent to the model
- Step 1: Task contentThe user prepares a query or a task for AI.
- Step 2: Zone and checkThe content is checked for designated data categories and assigned to an H.A.D.A. zone.
- Step 3: User decisionThe user sees what will be restricted and decides: cancel, edit, check again or send the restricted version.
- Step 4: RecordAn evidence trail for operations in the controlled path; the scope of the record depends on the configuration.
H.A.D.A. zones in AiTrust
How AiTrust classifies content before it is sent to the model: from A, with no sensitive data detected, to D, with critical data. The handling in each zone follows from the organization’s policy.
| Zone | Example content | Handling (standard profile) | Who decides |
|---|---|---|---|
| A (zone A) | Example contentContent with no sensitive patterns detected. | Handling (standard profile)Sent unchanged. | Who decidesThe user sends. |
| B (zone B) | Example contentFinancial and business data, e.g. revenue, budget, forecasts, report exports. | Handling (standard profile)The version with masked data is sent. | Who decidesThe user approves the masked version. |
| C (zone C) | Example contentPersonal and contact data, e.g. e-mail address, postal address, customer data, salaries. | Handling (standard profile)Approval request. | Who decidesA person with approval permission, other than the requester. |
| D (zone D) | Example contentCritical data, e.g. passwords, tokens and API keys, PESEL (Polish national ID number), NIP (Polish tax ID), account and phone numbers, medical data, prompt injection attempts. | Handling (standard profile)Sending is held. | Who decidesNo one approves sending: the content must be corrected or canceled. |
- The zone is assigned by the organization’s rules; when several rules match, the most restrictive zone applies. The content examples are built-in patterns, used when the organization has not defined its own rules.
- The built-in patterns do not yet recognize all medical terms in the same way. ICD diagnosis codes and English medical terms go to zone D, while the Polish terms „dane pacjenta” (patient data) and „dane medyczne” (medical data) go to zone C. An organization that processes health data can assign zone D to them in its own rules.
- The organization sets the handling for each zone in its policy; some deployment preparation variants have stricter settings, e.g. approval as early as zone B. In the example below (zone D, Redact decision), restricting the content was configured instead of holding.
Which steps of the method AiTrust supports
A methodology describes how an organization sets the rules for working with AI; AiTrust provides technical support for selected steps, to an extent that depends on the deployment configuration.
Check before sending
- By default in AiTrust
- Before sending, the user sees the H.A.D.A. zone, the decision and the list of masked data. The user then cancels, edits the content, checks it again or sends the restricted version.
- Your organization decides
- The organization sets the rules that assign zones and the handling in each zone.
Related steps
Human decision threshold (H), Operation classes in zones A to D (D)
Operation approval
- By default in AiTrust
- Operations of designated classes require a human decision before execution, and in the standard profile this applies to zone C.
- Your organization decides
- The organization sets the scope of AI autonomy separately for each process.
Related steps
Human decision threshold (H), Scope of autonomy (A)
Zones in the panel and in audit records
- By default in AiTrust
- Zones A to D are visible in the application’s control panel, and audit records can be filtered by zone.
- Your organization decides
- The organization sets the evidence retention period.
Related steps
Operation classes in zones A to D (D), Accountability (A)
Zone in the agent passport
- By default in AiTrust
- The baseline policy of the Agents area has an initial configuration of the agent passport with an approval threshold, operation and cost limits and grants to tools.
- Your organization decides
- For each AI digital worker, a maximum H.A.D.A. zone is set, along with the threshold above which approval is needed.
Related steps
Scope of autonomy (A), Operation classes in zones A to D (D)
How this step looks in the AiTrust application
The organization sets its H.A.D.A. zone policies in HADA Policy Studio: it creates rules, checks them in a preview and publishes successive versions that can be rolled back.
What the available example shows
This example shows the human decision step before content is sent to the model.
What this example shows
- The task “Prepare a reply to a customer complaint” assigned to zone D.
- The version sent to AI, in which the e-mail address, phone number and account number were replaced with the [EMAIL], [PHONE] and [IBAN] tags.
- The list of three redactions and the “Redact” decision, that is, restricting the content.
- The “Sending blocked” message, because zone D requires sensitive data to be removed before sending.
- The user’s choice: “Cancel”, “Edit prompt”, “Apply to prompt and check again”, “Check again” or “Send restricted version”.
What this example does not show
The screenshot shows the interface in one example, on synthetic data and in one configuration. It does not prove that every data category will be recognized, and it does not show other zones or cases.
Related answers on the platform and compliance
Answers on the platform page and on the AI Governance and compliance page
- Human decision in the process
Where in the process a human decision is required and what it covers.
- Masking data in queries
What data masking provides and what it does not ensure on its own.
- Reviewing the result before a decision
What requires human review before an AI result is used.
- AI Act
Information on the AI Act in the security and compliance area.
Other methodologies
- AI operating boundariesB.O.R.D.E.R.
- Readiness assessment methodologyI.A.G.R.A.
- Protection and resilienceS.H.I.E.L.D.
See the check before sending in a demo
We will show the check in a demonstration configuration and discuss which operations in your processes should require a human decision.