Three Types of Decisions: When Humans Decide, When AI Decides
Not every decision needs a human. And not every decision should be left to AI. A framework for assignment – with concrete HR examples.
The Question Every HR Department Asks
“Which decisions can AI make on its own – and which must stay with humans?”
This question comes up in every conversation about AI in HR. From the works council (Betriebsrat), from senior leadership, from compliance. And the usual answer – “AI supports, humans decide” – is too vague for productive use.
Because in practice, a single HR process consists of dozens of individual decisions. Take sick leave processing: is the document complete? Which collective agreement applies? Does the illness duration exceed the continued pay threshold? Does a return-to-work programme (BEM) need to be initiated? Who needs to be informed?
Some of these decisions need a human. Some don’t. And some should consciously NOT rest with a human – because AI demonstrably makes them better.
Three Types of Decisions
Type 1: Human Decides
Here, a human must make the final decision. The agent can prepare, suggest, compile materials – but the decision itself rests with the human.
When: Discretion required, discrimination risk, works council obligation, individual situations.
Example: Return-to-work planning after long-term illness. The agent has all the data: illness duration, return-to-work history, occupational health recommendations, available positions. But the decision about which reintegration model fits this specific person needs a human. It’s about the individual situation, about empathy, about the conversation with the person. The works council (Betriebsrat) has co-determination rights (Mitbestimmung). And if the decision is wrong, it has real consequences for a real person.
What the Decision Layer does: It enforces human-in-the-loop. Architecturally, not organisationally. The agent cannot make this decision autonomously – even if its confidence is high.
Type 2: Rule Set Decides
Deterministic logic. If condition X, then result Y. No room for interpretation.
When: Collective agreements, works council agreements, tax legislation, pay grading, deadline calculations.
Example: Pay grade classification. New employee, job profile available, collective agreement is clear. The classification follows from the criteria in the collective agreement. This is not a decision that needs interpretation – it is rule application. The Decision Layer applies the current rule version and documents the result.
What the Decision Layer does: It ensures the current rule version is applied. When the collective agreement changes, the new version takes effect from the effective date – automatically, without someone needing to inform 50 specialists at 12 locations.
Type 3: AI Decides Autonomously
And this is where it gets interesting. Because this category is usually told wrong. The standard narrative: “For simple standard cases, AI is also allowed to decide on its own.” That sounds like permission for trivial tasks.
The reality is different. There are decisions where AI is not just faster, but demonstrably better than a human. Not because AI is smarter – but because it lacks three structural weaknesses that humans have.
Advantage 1: Consistency Across Locations and People
50 specialists at 12 locations apply the same collective agreement. Each interprets edge cases slightly differently. In Hamburg, a special payment request is approved; in Munich, the identical case is rejected. This is not a training problem – it is the natural variance of human decision-making with ambiguous rules.
An AI operating on a versioned rule set decides identically. Every time. At every location. At 9 AM and at 4 PM.
Concrete example: Continued pay period calculation. Same case, same rule, same result. Not dependent on which specialist at which office handles the case.
Advantage 2: No Fatigue
A recruiter screens differently on Monday morning than on Friday afternoon. After the 50th application, attention drops. The second-to-last candidate was particularly strong – the next one seems weaker by comparison, even though they objectively meet the requirements (anchoring bias). The recruiter just received bad news – the next three assessments are harsher (affect heuristic).
These are not personal weaknesses. This is human cognition. Well-researched, extensively documented, and measurable in every repetitive decision process.
An AI evaluates application number 1 with the same rigour as application number 200. It does not have a bad day.
Concrete example: Requirement matching in recruiting. Every application is checked against the same criteria profile. Not influenced by the order of applications, not by the time of day, not by the recruiter’s emotional state.
Advantage 3: Completeness in Rule Checking
This is the advantage that gets underestimated most.
An HR specialist checks a sick leave notification against three or four criteria that come to mind: illness duration, continued pay period, perhaps the return-to-work threshold. But do they also check the waiting period rule? The special provision for part-time employees in the company-specific agreement? The reporting obligation to the occupational health authority for certain conditions? The special case for workplace accidents? The rule for fixed-term contracts?
Every time? Also on Friday at 4 PM? Also when handling five other cases in parallel?
An AI checks against all applicable rules, in the current version, completely and documented. Not because it is smarter – but because it does not forget. And because its rule set is not stored in people’s heads, but in a versioned system.
Concrete example: Sick leave processing. The agent checks every electronic sick note against all 12 relevant criteria from collective agreement, works council agreement, and law. Every time. The result: fewer errors that only surface at the next audit.
Why a Single Process Contains All Three Types
The framework becomes practically useful when you understand: a single HR process almost always contains ALL three decision types.
Take sick leave processing as a continuous example:
| Step | Decision Type | Why |
|---|---|---|
| Receive and validate sick note data | AI autonomous | Document classification, high accuracy, structured input |
| Match against employee master data | Rule set | Deterministic, no interpretation |
| Check continued pay period | AI autonomous | Checks against ALL collective agreement criteria, more consistent than any specialist |
| Assess return-to-work obligation (>6 weeks in 12 months) | Human | Discrimination risk with health data, works council co-determination |
| Inform line manager | AI autonomous | Consistent information, no omissions, no interpretation needed for WHAT is communicated (only absence and duration, no diagnosis) |
| Initiate return-to-work measures | Human | Individual situation, discretion, works council involvement |
Note the third row: “Check continued pay period” is not under “Rule set” but under “AI autonomous”. Why? Because the AI doesn’t just apply simple if-then logic here – it performs the check COMPLETELY against all applicable rule sets. Something a human in practice never fully does, because they don’t have all special provisions in their head.
That is the crucial point: “AI autonomous” is not the category for trivial tasks. It is the category for decisions where consistency, fatigue-resistance, and completeness matter more than discretion.
What This Means for the Works Council
The works council (Betriebsrat) is often sceptical about AI autonomy. Rightly so – when it’s unclear WHY AI decides autonomously.
With the three-types framework, the argument becomes transparent:
“AI decides autonomously for deadline calculations. Not because we want to cut headcount. But because we know that 50 specialists at 12 locations calculate the same deadline differently. AI always calculates correctly. And when it’s not certain, it escalates to a human. That’s traceable, documented, and visible in the auditor portal at any time.”
That is an argument the works council understands. It’s not about replacement – it’s about quality assurance.
The Decision Layer Makes the Assignment Operational
The framework remains theory if it’s not technically enforced. The Decision Layer implements the three-types assignment for every process step:
For every micro-decision, it’s defined: human, rule set, or AI. For AI decisions, it’s documented why AI is the right choice. For human decisions, human-in-the-loop is architecturally enforced. For rule-set decisions, the current rule version is stored. Every decision – regardless of type – generates a complete audit trail entry.
→ Decision Layer – Overview and Examples
Book a meeting – We’ll show you on your specific process which decisions stay with humans and which AI handles better.