Salesforce's autonomous AI agent platform built on the Einstein 1 Platform. Independently assessed by BenchLytix using proxy methodology derived from publicly available SLA, certification, and pricing documentation.
Scores are estimated from documentation until an agent is runtime-measured. Methodology v2.6.0 rev 1.27.
Week of 2026-05-04 · Manually assessed
Methodology v2.6.0Public-evidence proxy across four pillars. Not directly comparable to indie agent scores.
Manually assessed by BenchLytix · Week of 2026-05-04
Analyst est.Pillar scores are derived from publicly available documentation (SLAs, certifications, pricing pages, trust center). They do not represent live API performance measurements. Maximum composite: 79/100.
Public sources used to derive the proxy score. Last reviewed . Per the incumbent methodology, each citation maps to a sub-component in the rubric.
SLA tier classified as "99_5_to_99_9" because Salesforce default contractual SLA is ~99.0-99.5%; enterprise tier reaches 99.9% but is not the published default. Conservative classification per BOUNDARY CASE #1 in lib/incumbent/types.ts. Status page classified as "with_history" — trust.salesforce.com shows per-instance availability >12 months back. Incident history: trust.salesforce.com shows 2-4 P1/P2 incidents in trailing 12 months depending on instance. Conservative classification "three_to_five_p1". Reliability case study: multiple cases reference uptime/scale (Wiley 40% deflection, Accenture deployment). Latency vendor benchmark: Salesforce Engineering blog documents P90 end-to-end latency 15s→11s, sub-second TTFA for voice. Counts. Latency third-party benchmark: NONE found for Agentforce specifically. Cost: $0.10/conversation public, transparent unit pricing. Free trial requires Salesforce org setup but no sales call — qualifies under BOUNDARY CASE #6. Security: SOC 2 Type II, ISO 27001, GDPR via DPA, HIPAA via BAA, trust.salesforce.com, Salesforce HackerOne bug bounty. Pentest is embedded in SOC 2 scope but no STANDALONE third-party attestation is published — FALSE per strict-reading policy.