Worked comparison · reference set reference-set-1.0 · draft

Draft. The answers below were collected from vendor documentation with AI assistance and have not yet been re-checked by the maintainer. Every cell links its source; treat the table as a reading list until the review is recorded.

Claims and document-intake agents that touch personal data

Which claims-intake agents document how they handle PII and PHI, and where a human stays in the loop? Insurance intake agents read policyholder records by design, so the questions are about data handling rather than access. Vendor trust pages describe certifications well and agent behaviour poorly; most answers below are "unknown" for that reason.

Answers were read from the vendor pages linked in each cell on 2026-10-09 — maintainer review pending. "Not documented" means the pages checked do not say; it is a question for the vendor, not a finding against the agent.

	Agentforce Insurance Salesforce	Tractable Tractable	Shift Technology AI Shift Technology	Guidewire AI Guidewire	UiPath Agentic Automation UiPath
Unresolved questions	1 of 5	5 of 5	4 of 5	5 of 5	1 of 5
Masks or minimises personal data before it reaches the model?	✔ documented Einstein Trust Layer: "Sensitive data is detected and masked before sending the prompt to the LLM"; admins choose what is masked · source	? not documented The security page could not be retrieved at collection time; the privacy page names SOC 2 Type II and ISO 27001 but not agent data handling · source	? not documented Not covered on the security page · source	? not documented The trust page lists certifications (SOC 1/2 Type 2, ISO 27001, ISO 27701, PCI DSS) and makes no statement about AI features · source	? not documented Not covered on the governance page · source
Data residency or regional hosting documented?	? not documented Not covered on the pages checked · source	? not documented Not found on the pages checked · source	✔ documented "No transfer of European personal health data to a country outside the European Economic Area" · source	? not documented Not mentioned on the trust page · source	✔ documented "Regional data residency is available across Automation Cloud regions"; "air-gapped and on-premises deployments" via Automation Suite · source
Customer data excluded from model training (documented)?	✔ documented "Your data isn't retained by third-party LLMs" and is not used for their training · source	? not documented Not found on the pages checked · source	? not documented Not covered on the security page · source	? not documented Not mentioned; only a general commitment to process data "as contractually permitted" · source	✔ documented "Your data never trains a model and never leaves approved boundaries" · source
Human review before a decision or record change is executed?	◐ configurable Per action: "enable the Require user confirmation setting" so the agent asks before a record change · source	? not documented Not found on the pages checked · source	? not documented The page describes agent actions as "human-guided" without saying what a handler approves · source	? not documented Not mentioned on the trust page · source	◐ configurable Audit logs capture the "human approver" of agent actions, which implies an approval step; the mechanism itself is not described on this page · source
Audit log of agent actions documented?	✔ documented "Prompts, responses, and trust signals are logged and stored in Data 360" · source	? not documented Not found on the pages checked · source	? not documented Described as "auditable" without documentation of the log · source	? not documented Not mentioned on the trust page · source	✔ documented "Tamper-evident logs capture every agent action, prompt, response, tool call, model version, and human approver" · source
Evidence about the agent	T1 · FedRAMP Moderate +6 vendor-level (inherited): STAR for AI, STAR Level 1, ISO 42001, EU AI Act Art. 50 statement, GDPR DPA available, GDPR DPF active	T5 · none found in the sources checked +3 vendor-level (inherited): GDPR, ISO 27001, SOC 2 Type II	T5 · none found in the sources checked +4 vendor-level (inherited): EU AI Act Art. 50 statement, GDPR, GDPR DPA available, GDPR DPF active	T5 · none found in the sources checked +2 vendor-level (inherited): GDPR DPF active, HIPAA	T5 · none found in the sources checked +14 vendor-level (inherited): AIUC-1, FedRAMP Moderate, Cyber Essentials, Cyber Essentials Plus, EU AI Act Art. 50 statement, FedRAMP, HIPAA, HITRUST, IRAP, ISO 27001, ISO 42001, NIST AI RMF aligned, SOC 2 Type I, SOC 2 Type II
Pilot restrictions that apply	Confirm the processing region in writing before personal data of EU or UK residents is in scope. Ask for the SOC 2 report under NDA before the pilot touches customer data; none was found in the sources checked. No PHI in the pilot until a HIPAA BAA is in hand; no HIPAA evidence was found in the sources checked. A public trace shows data leaving the environment under adversarial input. Confirm the vendor's fix is in the version you deploy, and keep an egress allowlist around the agent during the pilot.	Keep PHI and PII out of pilot inputs, or apply your own redaction before the agent sees a record. Confirm in contract that customer data is excluded from model training before the pilot starts. Keep the agent advisory: no automated claim decisions or system writes during the pilot. Log inputs and outputs on your side; the vendor's audit trail is not documented. Confirm the processing region in writing before personal data of EU or UK residents is in scope. No PHI in the pilot until a HIPAA BAA is in hand; no HIPAA evidence was found in the sources checked. No public trace of this agent under adversarial input was found. Run it in a sandbox against your own prompt-injection and tool-abuse scenarios before the pilot, and keep the transcript; the enterprise edition records one.	Keep PHI and PII out of pilot inputs, or apply your own redaction before the agent sees a record. Confirm in contract that customer data is excluded from model training before the pilot starts. Keep the agent advisory: no automated claim decisions or system writes during the pilot. Log inputs and outputs on your side; the vendor's audit trail is not documented. Ask for the SOC 2 report under NDA before the pilot touches customer data; none was found in the sources checked. No PHI in the pilot until a HIPAA BAA is in hand; no HIPAA evidence was found in the sources checked. No public trace of this agent under adversarial input was found. Run it in a sandbox against your own prompt-injection and tool-abuse scenarios before the pilot, and keep the transcript; the enterprise edition records one.	Keep PHI and PII out of pilot inputs, or apply your own redaction before the agent sees a record. Confirm in contract that customer data is excluded from model training before the pilot starts. Keep the agent advisory: no automated claim decisions or system writes during the pilot. Log inputs and outputs on your side; the vendor's audit trail is not documented. Confirm the processing region in writing before personal data of EU or UK residents is in scope. Ask for the SOC 2 report under NDA before the pilot touches customer data; none was found in the sources checked. No public trace of this agent under adversarial input was found. Run it in a sandbox against your own prompt-injection and tool-abuse scenarios before the pilot, and keep the transcript; the enterprise edition records one.	Keep PHI and PII out of pilot inputs, or apply your own redaction before the agent sees a record. No public trace of this agent under adversarial input was found. Run it in a sandbox against your own prompt-injection and tool-abuse scenarios before the pilot, and keep the transcript; the enterprise edition records one.

What the evidence says

Agentforce Insurance: 4 of 5 questions documented, 1 to ask the vendor; a bounded pilot is possible under 4 restrictions.
Tractable: the documentation checked answers 0 of 5 questions. The evidence is not sufficient to scope a pilot without answers from the vendor.
Shift Technology AI: the documentation checked answers 1 of 5 questions. The evidence is not sufficient to scope a pilot without answers from the vendor.
Guidewire AI: the documentation checked answers 0 of 5 questions. The evidence is not sufficient to scope a pilot without answers from the vendor.
UiPath Agentic Automation: 4 of 5 questions documented, 1 to ask the vendor; a bounded pilot is possible under 2 restrictions.

None of this ranks the agents. A table with more ✔ marks documents more; it does not deploy more safely. Open each dossier for the full ledger, the sources and the dispute link, and put the "not documented" rows to the vendor before scoping a pilot.