Analyst firms track market size. Vendors publish marketing. This database tracks what authority production agents were actually granted: every verified deployment, mapped against the same authority fields, edition after edition.
The fields
What every record captures
The PHINAMEK Agentic Protocol is the evaluation framework behind the Radar. Each deployment is judged against the same six fields, so records compound into something queryable.
Deployment
Who put what agent into production. The "who" behind the deployment.
Allowed actions
What the agent is permitted to do. The authority grant, the actual permission slip.
Forbidden actions
What it is barred from. The boundary line, where governance lives or dies.
Escalation
What happens when the agent hits its limit. Human review, thresholds, handoff paths.
Kill switch
Whether the agent can be stopped, and whether that has ever been tested.
Identity model
Whether the agent acts under its own scoped identity or on somebody's credentials.
The records
Search the database
Every record below is drawn from a verified Radar edition and carries its sources. Fields that cannot be sourced are marked "not disclosed" — absence of disclosure is an information gap, not proof of failure.
Confirmed Primary-source evidence of a production deployment with a named customer. Reported Secondary reporting, vendor claims, or partially verified activity — see the record’s note.
31 records
No records match your search.
R-001 · Edition 01 Reported
DEPLOYMENTOpenAI x U.S. federal agencies (IRS, SSA, VA, DHS) — conversational AI on federal websites answering questions, completing forms, guiding procedures
INDUSTRYPublic sector — federal civilian agencies
ALLOWED ACTIONSretrieve verified government records, recommend guidance, complete and submit forms
FORBIDDEN ACTIONSnot disclosed
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONDefine exactly which forms the agent may submit and what verification precedes each submission — form submission is a write authority on citizen records. Log every submission with attribution, and publish the boundary before an oversight body asks for it.
NOTEVerification note: press-reported via secondary summary of NYT reporting; no official agency announcement. Stage and scope reported, not confirmed.
DEPLOYMENTAccor ALL Concierge — conversational travel companion for inspiration, hotel comparison, booking support, reservation questions
INDUSTRYTravel / hospitality
ALLOWED ACTIONSretrieve guest preferences and property info, recommend hotels and itineraries, communicate via web, app, and chat
FORBIDDEN ACTIONSnot disclosed
ESCALATIONhuman support teams take over when requests are too complex or require direct assistance
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONName the booking threshold in writing: what the concierge may reserve on its own versus what kicks back to the guest or a human agent. Attach the guest's explicit confirmation to every booking action in the log.
ALLOWED ACTIONSretrieve telemetry and demand forecasts, execute real-time network adjustments
FORBIDDEN ACTIONSnot disclosed
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONWrite authority over live network infrastructure demands a defined rollback path. Specify which adjustments execute autonomously, which require human approval, and how you reverse an autonomous change that degrades service.
ALLOWED ACTIONSretrieve knowledge base and telemetry, recommend repair guidance
FORBIDDEN ACTIONSnot disclosed
ESCALATIONhuman technicians remain the actors; copilot guides, technicians execute repairs
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONKeep the guide-versus-execute boundary explicit as capabilities grow: the copilot guides, technicians act. Define the trigger for when guidance becomes a recommendation a technician cannot easily second-guess, and log the guidance given.
ALLOWED ACTIONSretrieve approved external sources, prepare assessments for human review
FORBIDDEN ACTIONSnot disclosed
ESCALATIONassessments presented for human review; no independent AI compliance determination
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONWrite the research-versus-decide division into the control documentation — agent researches and drafts, human decides — because examiners will ask where the determination was made. Treat the approved-source list as the agent's access boundary and audit against it.
DEPLOYMENTLititz Mutual x Guidewire ProNavigator — conversational guidance from policies and procedures in underwriting and claims
INDUSTRYInsurance — P&C mutual
ALLOWED ACTIONSretrieve policy and procedure corpus, recommend guided answers
FORBIDDEN ACTIONSnot disclosed
ESCALATIONhuman-in-the-loop support
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONDefine what counts as guidance versus advice: when an answer could be relied upon as a coverage determination, route it to a licensed human. Log the corpus version behind every answer so guidance traces to the policy in force.
DEPLOYMENTDanske Bank x Mastercard Agent Pay — AI agent booked and paid for an experience on live card rails
INDUSTRYBanking / payments
ALLOWED ACTIONSsearch and select experiences, complete booking, settle payment on card rails
FORBIDDEN ACTIONSnot disclosed
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELPayment Passkey authentication; issuer sees agent-initiated transaction flag
RECOMMENDATIONPer-transaction approval is the missing control — define the spend threshold above which the cardholder approves each transaction, not just the enrollment. Reconcile agent-initiated transaction flags against cardholder confirmations.
NOTEVerification note: Customer consent evidence present; no evidence of per-transaction approval.
DEPLOYMENTTrip.biz Agent ONE — four connected agents for business-travel planning, booking, approval, and insight
INDUSTRYTravel — corporate travel management
ALLOWED ACTIONSretrieve travel content and policy, recommend itineraries, book travel, auto-approve low-risk requests, communicate disruptions
FORBIDDEN ACTIONSnot disclosed
ESCALATIONexceptions routed to human review; only low-risk requests auto-approved
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONWrite down the definition of 'low-risk' the auto-approval uses, because that definition is the real authority boundary. Route everything above it to human review and log the decision with the policy version applied.
DEPLOYMENTDuke Energy x AWS — agents automating grid interconnection studies
INDUSTRYEnergy / electric utilities
ALLOWED ACTIONSretrieve grid models and standards, run and coordinate study workflows, recommend upgrade decisions
FORBIDDEN ACTIONSnot disclosed
ESCALATIONengineers direct the study and make final engineering decisions
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONKeep final engineering decisions with engineers in writing as agent capabilities expand. Define which study outputs the agent may finalize versus recommend, and retain the study inputs so any decision can be reconstructed.
DEPLOYMENTPenn Medicine x OpenEvidence — AI clinical decision support deployed to 10,000+ clinicians
INDUSTRYHealthcare — academic health system
ALLOWED ACTIONSretrieve EHR data and medical literature, recommend evidence-based insights
FORBIDDEN ACTIONSnot disclosed
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONClinical recommendations demand a verification control: define how a clinician confirms or overrides the agent's insight and log the override. Scope the EHR data the agent can read to the minimum necessary — that scope is its access boundary.
DEPLOYMENTDefense Logistics Agency — agentic AI digital employees for logistics operations
INDUSTRYPublic sector — defense logistics
ALLOWED ACTIONSwork from agency data, execute logistics workflows
FORBIDDEN ACTIONSnot disclosed
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELpersona-based agent accounts mirroring human job roles
RECOMMENDATIONPersona-based accounts are the right identity model — now define what each persona may not do. A job-role mirror without forbidden actions is an incomplete boundary; document off-limits workflows per persona and review them when roles change.
DEPLOYMENTCrocs x Salesforce Agentforce Rivet — superagent for shopping and customer support
INDUSTRYRetail — footwear/apparel
ALLOWED ACTIONSretrieve product and order data, recommend products, create service cases, execute checkout, communicate via chat and voice
FORBIDDEN ACTIONSexposing PII such as shipping addresses; going outside approved knowledge
ESCALATIONhuman handoff on repeated requests for a person or frustration signals, with full history
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONThe PII prohibition is stated — now enforce it technically, not just in the prompt: define the data fields the agent can never retrieve, and audit checkout executions against customer confirmation.
DEPLOYMENTRedwood Logistics x Augment Augie — AI teammate automating freight quote-to-cash workflows
INDUSTRYLogistics / supply chain
ALLOWED ACTIONSretrieve load, carrier, and shipment data; write updates to TMS/ERP; execute freight workflows; communicate via phone, email, portals
FORBIDDEN ACTIONSnot disclosed
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONWrite access to systems of record is the highest-risk permission here. Define which TMS/ERP writes the agent may commit versus stage for human approval, and make every write attributable and reversible.
DEPLOYMENTAXA x Publicis Sapient — Global AI Hub deploying AI agents group-wide (motor claims, email, knowledge)
INDUSTRYInsurance
ALLOWED ACTIONSretrieve claims and knowledge data, recommend actions
FORBIDDEN ACTIONSnot disclosed
ESCALATIONhuman oversight embedded throughout the AI lifecycle
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATION'Human oversight embedded throughout' needs a concrete definition per deployment: who reviews what, at which step, with what authority to stop the agent. Write the oversight down per use case or it will not survive an audit.
FORBIDDEN ACTIONSsetting rates or altering rating tables, locked in deterministic rating engine
ESCALATIONlicensed broker retains authority over the binding trigger
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONThe licensed broker's binding trigger is the control that matters — define the exact handoff: what the agent prepares, what the broker verifies, and where that verification is recorded. Keep rate-setting locked outside the agent's reach as the deployment scales.
DEPLOYMENTBNP Paribas x Google Cloud — five-year partnership; agents for credit memos and CIB workflows
INDUSTRYBanking
ALLOWED ACTIONSretrieve credit documentation and procedures, draft credit memos and research
FORBIDDEN ACTIONSprocessing barred data categories in public cloud
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELper-agent authentication with least-privilege resource grants
RECOMMENDATIONThe barred-data-categories prohibition is the boundary to enforce technically: define the classification check that runs before any retrieval, and audit the agent's actual data access against its least-privilege grants.
NOTEVerification note: Rests on one secondary source; the original vendor URL was not captured.
RECOMMENDATIONPersonalization at 10M accounts is a data-access boundary question: define which profile attributes the agent may use for offers and which are off-limits, and give members a visible way to know an offer came from the agent.
INDUSTRYHealthcare — international private hospital
ALLOWED ACTIONSretrieve emails and FAQs, write case summaries, route to departments, answer messaging queries
FORBIDDEN ACTIONSnot disclosed
ESCALATIONhuman verification of every categorization before automation; human handoff when agent is unsure
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONHuman verification of every categorization is the control — define the review rule explicitly and defend it as automation scales, because the pressure will be to relax it. Log the verifier and the verification decision.
DEPLOYMENTSiemens x Salesforce — Agentforce for industrial sales, service, and lead qualification
INDUSTRYManufacturing — industrial technology
ALLOWED ACTIONSretrieve engineering and CRM data, recommend parts and pitches, send outreach emails, save lead details
FORBIDDEN ACTIONSnot disclosed
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONOutbound email from an agent needs a send authority: define what the agent may send without review versus what requires approval, and keep a human-readable record of every send.
ALLOWED ACTIONStool use and multi-step workflows on customer data in-region
FORBIDDEN ACTIONSnot disclosed
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELper-agent identity with action logging
RECOMMENDATIONPer-agent identity with action logging is the foundation — now define the retention and the review: who reads the logs, on what cadence, and what pattern triggers an investigation. Identity without review is inventory, not governance.
DEPLOYMENTAllstar Services x Revin — AI voice/SMS agents as first contact for homeowners, expanded to 10 roofing brands
INDUSTRYConsumer — residential home services
ALLOWED ACTIONSretrieve brand scripts, answer calls and texts, book appointments into calendars
FORBIDDEN ACTIONSnot disclosed
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONDefine the autonomous-booking threshold before you expand: appointments booked with zero human involvement are a write authority. Name what the agent may commit on its own and what kicks back to a human, per brand.
DEPLOYMENTAltery x Vivox AI — explainable agents mapping corporate ownership for due-diligence onboarding
INDUSTRYFintech / payments
ALLOWED ACTIONSretrieve registry data, produce ownership maps, recommend escalations
FORBIDDEN ACTIONSnot disclosed
ESCALATIONanalysts keep review responsibility and final decisions
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONIn regulated workflows, demand the audit trail before the automation: agent investigates and recommends, human decides, every step logged. Treat the vendor's explainability claims as claims until independently verified.
NOTEVerification note: Vendor-produced case study; no customer-published confirmation.
DEPLOYMENTHackerRank Chakra — AI agent conducting technical coding interviews end-to-end
INDUSTRYHR tech / talent acquisition
ALLOWED ACTIONSretrieve candidate work context, ask follow-up questions, write scores and narrative reports
FORBIDDEN ACTIONSmaking final hiring decisions, remain human
ESCALATIONhuman makes the final hiring decision
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONCopy the proven pattern: the agent scores and writes the narrative, the human makes the hiring decision — and log every step in between. In hiring, the audit trail is the product's license to operate.
NOTEVerification note: Rollout scope and metrics are vendor-reported.
DEPLOYMENTBlend Autopilot × Blend Labs — agentic home-lending layer: document review, borrower conditioning, pre-underwriting on live loans
INDUSTRYFinancial services / home lending
ALLOWED ACTIONSretrieve loan files, review documents and income, follow up with borrowers, condition borrowers, drive pre-underwriting, run closing quality control
FORBIDDEN ACTIONSnot disclosed
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONIf your agents touch regulated decisions, scope the audit to include agent decisions, not just agent outputs. Define the decision records the agent retains before the examiner asks — what the agent decided matters as much as what it surfaced.
DEPLOYMENTTeleAgent × China Telecom — end-to-end workplace AI agent for knowledge work, private-deployment edition for sovereign environments
INDUSTRYTelecom
ALLOWED ACTIONSretrieve enterprise file and system data, draft documents, generate reports and charts, execute end-to-end knowledge tasks
FORBIDDEN ACTIONSnot disclosed
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONGovern the skill marketplace like a software supply chain: vet the skills, scope data permissions per skill, and monitor for exfiltration. At user-built scale, the skills are the attack surface.
NOTEVerification note: Wire release timed out on fetch; facts cross-verified via two independent republications. Vendor-reported scale figures; vendor and deploying enterprise are the same organization.
DEPLOYMENTPegasus Airlines × Zingly.ai — agentic voice AI handling guest-service calls and resolving requests in-call
INDUSTRYTravel / airline
ALLOWED ACTIONSretrieve passenger service data, communicate with callers, execute in-call resolution actions, escalate to human agents
FORBIDDEN ACTIONSnot disclosed
ESCALATIONescalate to human agents with interaction history intact
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONTreat voice agents as write-authority agents, not chat upgrades. Audit the in-call actions, define the escalation threshold, and check consent and disclosure practice for AI-handled voice calls.
DEPLOYMENTING × Google (Gemini Enterprise) — mortgage AI agent reviewing applications with human final decisions
INDUSTRYBanking
ALLOWED ACTIONSretrieve mortgage applications, recommend outcomes and next steps, communicate explanations
FORBIDDEN ACTIONSmake final mortgage decisions, reserved for human employees
ESCALATIONhuman employee makes final assessments
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONStage the autonomy crossing and announce the staging: mortgage decisions capped at human final approval is the right posture. Publish where the boundary sits today and you own the narrative when you move it tomorrow.
NOTEVerification note: Oct 8 press-event reporting, not an ING press release; underlying pilot announced June 2026.
RECOMMENDATIONAn agent positioned as the authoritative reference needs its authority fenced: define which determinations it may make versus recommend, require human sign-off on classification decisions, and log every determination with attribution.
NOTEVerification note: Pilot at directive stage; initial small-scale deployment begins within six months. Authority-transfer end state is aspirational per the memorandum.
DEPLOYMENTNebraska Cancer Specialists × Predoc — AI agents for inbound fax ingestion and outside medical-records retrieval
INDUSTRYHealthcare / oncology
ALLOWED ACTIONSretrieve outside medical records, communicate with provider offices, ingest inbound faxes, manage documents
FORBIDDEN ACTIONSnot disclosed
ESCALATIONnot disclosed
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONPHI through the agent layer demands explicit data boundaries: define what the agent may read and share, prohibit EHR write-back without human review, and audit the document ingestion chain end to end.
NOTEVerification note: Deployment selected and announced, not yet live; no customer-side executive named in coverage reviewed.
DEPLOYMENTFreight Right × in-house MCP — MCP server exposing quoting, tracking, booking to AI assistants like Claude
INDUSTRYLogistics / freight
ALLOWED ACTIONSretrieve rates and shipment status, recommend quotes, transact bookings with customer confirmation, communicate updates
FORBIDDEN ACTIONScommit bookings without customer confirmation through Freight Right
ESCALATIONcustomer confirms booking requests through Freight Right before commit; manual-pricing cases routed to operational teams
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONMap your inbound agents the way you map your vendors: list every third-party assistant that can reach your booking surfaces, and define what each may do, whose credentials it uses, and how you revoke it. Keep human confirmation at the booking commit.
NOTEVerification note: Company press-release announcement; Freight Right is both the named enterprise and the agent builder.
FORBIDDEN ACTIONSbook flights, transact authority deferred to a later phase
ESCALATIONbooking deferred to a later phase; deliberately staged at retrieve and recommend
KILL SWITCHnot disclosed
IDENTITY MODELnot disclosed
RECOMMENDATIONStage the autonomy crossing and announce it: booking deliberately deferred to a later phase is the right boundary posture. When the phase arrives, define the commit authority, the credentials, and the revocation path before the first booking executes.
The database answers the question every governance lead is asking: which deployed agents hold which authority, and where are the gaps? "Show me every support agent with refund authority and no tested kill switch" is the kind of query this makes possible.