Most carriers can get a first picture of employee AI use from logs they already own: identity provider sign-ins, DNS, and a CASB or secure web gateway. That picture shows which AI apps were visited. Per-user and per-department usage, and what data was sent, only exist when sanctioned AI runs through one governed gateway.
Discovery tools show which AI apps were visited
Network and identity discovery is a sensible first step, and Microsoft Defender for Cloud Apps, Netskope, Zscaler, Palo Alto Networks Prisma Access, CrowdStrike, and Cloudflare One all offer a version of it.
Microsoft Entra's shadow AI discovery in Global Secure Access is typical. According to Microsoft's documentation, admins can see discovered generative AI apps, the users who access them, how often, the bytes sent and received, and risk scores drawn from the Defender for Cloud Apps catalog.
Carriers without one of these platforms can pull the same signals from systems they already pay for. Our shadow AI detection playbook walks through six of them, including OAuth grants, DNS logs, expense data, and an amnesty survey, and notes where each one misleads.
Discovery misses personal devices, embedded AI, and prompt content
Microsoft describes shadow AI discovery as "application-level inventory and risk scoring." Capturing prompt content is a separate feature, Generative AI Insights, that depends on TLS inspection (Microsoft Learn). Like any network tool, discovery sees only traffic routed through it and scores only apps in its catalog.
An adjuster who pastes a claim note into a chatbot on a personal phone over cellular never touches the corporate proxy, and the same goes for a laptop on a home network outside the inspection path.
AI inside software the carrier already approved is harder to spot. Copilot in Microsoft 365, AI features in Salesforce, and AI that core system vendors such as Guidewire and Duck Creek build into their platforms all travel over domains that are already sanctioned, so they look like ordinary business traffic.
Discovery also says little about content or cost. A byte count doesn't say whether the upload was a policyholder's medical history or a meeting agenda, and it says nothing about spend.
Regulators now want a count of AI systems
The NAIC's Big Data and Artificial Intelligence (H) Working Group built the AI Systems Evaluation Tool, now called the AI Risk Evaluation Supplement, to give examiners a standard set of questions. Fenwick reports that the pilot ran from March to September 2026 in 12 states: California, Colorado, Connecticut, Florida, Iowa, Louisiana, Maryland, Pennsylvania, Rhode Island, Vermont, Virginia, and Wisconsin.
Exhibit A, "Quantify Regulated Entity's Use of AI Systems," asks for current counts and use cases across all lines of business, with each system's name, how it is used, its program area, and its inherent risk level. Discovery logs supply the system name at best.
As of November 2026, version 7.0 is out as an exposure draft. The working group meets at the Fall National Meeting on November 16, and the NAIC has said it anticipates adoption there. The NIST AI RMF asks for the same thing in GOVERN 1.6. Our NAIC bulletin readiness guide covers what examiners request and what to keep on file.
A governed gateway records who used AI, with what data, at what cost
When sanctioned AI runs through one gateway, every request is logged with the person, department, model, data classification, and cost. That covers browser chat and API calls alike, and it gives an Exhibit A response and a budget review the same source.
Discovery is still how a carrier finds the unapproved tools. In our experience that list gets shorter once the approved option is the one employees prefer.
How Swept AI handles it
Swept AI keeps a live inventory of every AI application in use, including unapproved ones, through web access log parsing and vendor AI monitoring. See how governance works. For AI that runs through Swept, spend is metered in real time by person, team, and model.
How can FAQs
Yes, within limits. CASBs and secure web gateways such as Microsoft Defender for Cloud Apps, Netskope, Zscaler, and Prisma Access match traffic against an app catalog and list the AI apps in use, who used them, how often, and how much data moved. They only see traffic routed through them, they can only score apps in their catalog, and they do not record what was typed into a prompt unless a separate inline inspection feature is turned on.
Most carriers combine two sources. Discovery from identity provider, DNS, and CASB logs finds which AI apps employees visit, including unapproved ones. A governed gateway for sanctioned AI then records each request with the person, department, model, data classification, and cost, which is the detail a regulator or budget owner asks for.
The NIST AI Risk Management Framework is voluntary, but its GOVERN 1.6 subcategory calls for mechanisms to inventory AI systems, resourced according to organizational risk priorities. The NIST Playbook prefers inventorying all models or systems, and at minimum the high-risk ones.
It is Exhibit A of the NAIC AI Risk Evaluation Supplement, formerly the AI Systems Evaluation Tool. Titled 'Quantify Regulated Entity's Use of AI Systems,' it asks carriers for current counts and use cases of AI systems across all lines of business, including each system's name, how it is used, the program area, and its inherent risk level.
Generally not. Microsoft describes Entra shadow AI discovery as application-level inventory and risk scoring, and offers prompt-level visibility as a separate feature that requires TLS inspection. Discovery records which tool was used and how much data moved, without the contents of the prompt.