Methodology

Eight factors. Fixed weights. Published thresholds.

Every score is produced by one algorithm from operational metrics: response time, override rate, throughput against benchmark, customer satisfaction, uptime, tenure, data volume, and outcome events. The factor weights are fixed constants in a versioned model — the same for every customer, changed only through a versioned release, never by hand and never per account.

Model asam_v2.1Read-only accessFirst scores within 48 hours
HUBSPOTJOBBERVAPICSV EXPORTREAD-ONLY SCOPESINGESTidentifiers strippedHIERARCHICAL RIDGEsector-learned weightsAS0-1000IS0-100
Fig. 1 · Scoring pipeline. AS = AgentScore, per agent. IS = ImplementationScore, per company.

Factors

1Efficiency
2Quality
3Business impact
4Autonomy
5Reliability
6Durability
7Learning
8Governance

Signals read

Response time
Override rate
Throughput against benchmark
Customer satisfaction
Uptime
Tenure
Data volume
Outcome events

Metrics are read for the AI agents a customer runs, and for nothing else. Customer personal data is never read.

Published thresholds

Provisional flagBelow 60 percent confidence
Not scoredFewer than 50 recorded events
First scoresWithin 48 hours of connection
Benchmarked sectors3 of 10 verticals

Thresholds are part of the methodology. A score that cannot clear them is flagged or withheld, never rounded up.

Read-only by design

AgentScore connects to the systems already in production through read-only scopes, or scores from a CSV export. AgentScore does not write to customer systems.

Confidence thresholds

Below 60 percent confidence, a score is marked provisional. Under 50 recorded events, an agent is not scored at all. Scores that do not meet these thresholds are withheld.

No overrides

Scores cannot be adjusted by customers or by Orbitex. Recompute re-runs the algorithm; the product contains no manual score input.

Next: Coverage