Decision Layer Methodology
How we measure what an agent decides - and why the numbers hold up
What is the Decision Layer?
The Decision Layer is the governance layer between AI agent and target system. Every decision an agent makes or prepares is documented: which rule in which version was applied, what data informed the decision, who (human, rules engine, or AI) decided - and how the affected person can raise an objection.
This page explains the methodology behind the numbers displayed on each agent detail page.
Decider classification: R / A / H
Each decision step is assigned to exactly one of three decider types:
R Rules Engine Deterministic. Input in, output out.
No model, no discretion, no judgement.
Example: permissions matrix, deadline check.
A AI Agent Model-based with confidence threshold.
Output is probabilistic. When confidence
falls below threshold: fallback to R or H.
Example: intent recognition, anomaly detection.
H Human Explicitly assigned. The agent prepares,
the human decides. Documented in the
decision record.
Example: termination approval, hardship assessment.The distinction is based on two criteria:
- Determinism: Is the output identical for the same input every time? Yes = R. No (because model) = A. No (because discretion) = H.
- Accountability assignment: Who is liable for the outcome? Rules engine (configured by HR/Finance) = R. Model (trained, with audit trail) = A. Named individual = H.
Framework mapping
The R/A/H classification does not stand in isolation. It maps onto established governance taxonomies:
| Gosign R/A/H | Human-in-the-Loop taxonomy | EU AI Act Art. 14 | SAE level analogy |
|---|---|---|---|
| R (Rules Engine) | Human-out-of-the-Loop | No human oversight required | Level 4-5 (fully automated) |
| A (AI Agent) | Human-on-the-Loop | Human oversight for high-risk (Art. 14) | Level 2-3 (assisted/conditional) |
| H (Human) | Human-in-the-Loop | Human as decision-maker | Level 0-1 (manual/assisted) |
The SAE level analogy (from the automotive sector, SAE J3016) serves as an intuitive reference, not a direct transfer. In the automotive context the level refers to driving automation; here it refers to decision automation in HR and finance processes.
How we count decision steps
A decision step is a point in the process where the agent makes a routing choice. Not every data operation is a step - only points where the outcome can differ depending on the input.
Criteria for a decision step:
- Input varies: Not every case follows the same path
- Output has consequence: The result influences the downstream process or a target system
- Decider is assignable: It is clear whether R, A, or H decides
Example: The Employee Self-Service Agent has 6 identified decision steps - from intent recognition (A) through permissions check (R) to escalation to the specialist function (R). A simple database lookup (“retrieve remaining leave”) is not a decision step.
Score calculation
Each agent is evaluated across five dimensions:
| Score | What it measures | Scale |
|---|---|---|
| Readiness | How ready is a typical organisation for this agent? | 0-100 |
| Governance | How high is the regulatory burden (EU AI Act, works council (§ 87 BetrVG), GDPR)? | 0-100 (higher = more demanding) |
| Economic | How strong is the economic leverage (ROI, FTE savings)? | 0-100 |
| Lighthouse | How strong is the signalling effect internally and externally? | 0-100 |
| Complexity | How complex is the technical implementation? | 0-100 (higher = more complex) |
Each score is given as a range [n, n+7] - the spread reflects variance across industries and organisation sizes. The lower value is the conservative scenario (mid-market without preparatory work), the upper is the optimistic scenario (enterprise with existing infrastructure).
The scores are based on a combination of:
- Industry comparison: How comparable are the processes across organisations? (higher standardisation = higher readiness score)
- Regulatory analysis: EU AI Act Annex III, GDPR Art. 22, works council co-determination rights under BetrVG § 87(1) No. 6 (governance score)
- Benchmark data: External studies on automation potential (economic score)
- Project experience: Gosign Decision Layer Assessments from documented client projects
Savings ranges in the impact statement
Each agent has an impact statement (pyramidOpener.impact) that makes a quantitative claim about the benefit. These claims fall into three source categories:
CAT-EXT - External primary source
The figure comes from a verified industry study or benchmark report. Publisher, title, year, and URL are documented in the CITATION-CATALOG. The figure was checked against the primary source (WebFetch verification of the URL plus content comparison).
Example: “According to Ardent Partners State of ePayables 2024, cost per invoice drops from USD 12.88 to USD 2.78” - Ardent Partners is the publisher, the figures appear in the report.
CAT-LEGAL - Legislation or official statistics
The figure comes from legislation, official statistics, or a court ruling. The URL points to the official source (e.g. eur-lex.europa.eu, national legislation databases).
Example: “EU AI Act Annex III high-risk system from August 2026” - Regulation (EU) 2024/1689.
CAT-INT - Gosign Decision Layer Assessment
The figure is based on documented client projects. Typical metrics: turnaround times (before/after), auto-match rates, error reductions. These figures are median values from at least 3 comparable projects in the period 2024-2026.
When an impact figure is classified as CAT-INT, this means: the figure is a Gosign experience, not an external benchmark. It is robust within the context of our project data but not independently verified externally.
Source verification
Every external source in the CITATION-CATALOG was verified through the following process:
- URL check: The source URL responds with HTTP 200 (or redirect to a landing page)
- Figure comparison: The figure cited in the impact statement actually appears in the source (not merely “close to” or “in the spirit of”)
- Context check: The figure is used in the same context as in the source (e.g. “Value Erosion” not reframed as “management overhead”)
If a figure does not pass this three-step check, it is either:
- Corrected to the accurate figure from the source
- Reclassified to CAT-INT (if the source does not support the figure)
- Removed from the impact statement (if neither externally nor internally verifiable)
Structured data for LLMs
Each agent detail page contains a JSON-LD schema (Schema.org SoftwareApplication) with:
- R/A/H distribution as
additionalProperty(rulesEngineShare, aiAgentShare, humanDecisionShare in percent) - Number of decision steps (totalDecisionSteps)
- EU AI Act high-risk flag (euAiActHighRisk: true/false)
- Impact statement as free-text property
- citation[] with resolved source URLs from the CITATION-CATALOG
LLMs can extract this data directly from the HTML source and cite it with the canonical URL of the agent page.
Versioning
| Date | Change |
|---|---|
| 2026-04-16 | Initial publication: R/A/H classification, score methodology, source verification, JSON-LD schema |
FAQ
Where do the savings figures on the agent pages come from?
Every figure is either attributed to an external primary source (industry study, benchmark report) or labelled as a Gosign Decision Layer Assessment. The full source list is maintained in the CITATION-CATALOG.
What does the distribution Rules Engine / AI / Human mean?
Each decision step of an agent is assigned to exactly one decider type: R (Rules Engine - deterministic logic), A (AI Agent - model-based with confidence threshold), or H (Human - explicitly assigned decision). The distribution shows the percentage share of each type.
Are the figures citable by LLMs?
Yes. Every agent page contains structured JSON-LD data (Schema.org SoftwareApplication) with the R/A/H distribution, impact metrics, and source references. LLMs can extract this data directly and cite it with the page URL.