Why NIST AI RMF and AI agents end up in the same review
Somebody connects Claude to the CRM in March. By June the risk register has a line on it, and nobody can say which records answer it. NIST AI RMF and AI agents meet at exactly that point. The framework asks organizational questions, and the agent answers in tool calls.
NIST's AI Risk Management Framework (NIST AI 100-1) is voluntary guidance built on four functions: Govern, Map, Measure and Manage. The Generative AI Profile (NIST AI 600-1) is a companion. It lists risks specific to generative AI with suggested actions under those same functions. NIST also keeps a framework overview page and an AI RMF Playbook of suggested actions, both checked November 2026. None of them names a product.
What follows pairs each function with the record an access layer in front of the tool calls can produce. Elaichi is the access layer in the examples. No subcategory numbers appear here on purpose, because tying a subcategory to your own system is the work. MCP (Model Context Protocol) is the standard way an AI assistant calls tools in other apps.
Govern: one role each, and a named person who decides
Govern is about accountability. Who sets the policy, who holds it, and who can change it later.
In Elaichi a member holds exactly one role in an organization. Their permissions are their role's permissions, with nothing added on the side. A request to assign two roles is refused, and the message tells the admin to author a custom role that combines what they need. Custom roles come with the Gold and Black plans, and how roles and the audit trail line up covers that pairing in more detail.
Eight roles ship with every organization. Six form a strict chain from Guest up to Org Owner, where each tier holds everything the tier below it holds. Two sit off that chain: Billing Admin, and Auditor, which the product describes as free, read-only visibility for compliance and oversight. Auditor does not take a billable seat, so adding a reviewer costs nothing.
Auditor is narrower than its name suggests. Two of its view permissions are reserved and grant nothing. An Auditor does not see other members' connections, templates or toolboxes. What it opens is billing detail, the connector catalog, restrictions, API token metadata, single sign-on and SCIM settings, logging destinations and the audit log.
Who may change policy is a Govern answer too. Among the system roles, only Org Owner and Org Admin can create, edit or delete a restriction. Org Admin holds every permission except billing management and deleting the organization. Elaichi keeps deletion with the owner on purpose, so a compromised admin account cannot delete the workspace along with the evidence.
If your directory is the source of truth, wire it and say so. Single sign-on over SAML or OIDC, SCIM provisioning and group-to-role mappings come with Gold. A group mapping confers at most one role, for the same reason a person holds one.
One trap belongs in the written policy. Creating connectors is the only permission Elaichi flags as high trust. A restriction binds a connector's identity, not the host it calls, so blocking one connector does not stop a new connector built against the same API.
Map: which app, which tool, whose account
Map is the inventory. What is connected, in what context, and on whose authority. For a tool-calling agent that comes down to three lists.
The apps. Elaichi's catalog holds 650+ connectors of two kinds. Some Elaichi authors and runs itself. The rest are native MCP connectors, where the vendor builds and runs its own MCP server and Elaichi handles sign-in, access and audit. A team on Gold can bring its own remote MCP server under the same rules.
The tools. In Elaichi, connected tools are never listed one by one, however few there are. A client finds them with search_tools, then runs one with read_tool or execute_tool. Each connector's page in the console lists its tools and labels each one Read only, Makes changes or Can delete data. That labeling is the inventory a reviewer actually wants.
The identity. Every connected-tool call runs as one specific person, never the organization at large and never a shared service account, which is the argument in non-human identity for AI agents. Over MCP that person is whoever approved the OAuth grant behind the client's token. Record the exception as well: a connection shared with a team runs on its owner's account at the far end, so the call acts in that app as the owner, whoever asked.
Two more entries round out a Map record. Claude, ChatGPT, Cursor and Microsoft Copilot all sign into the same address, https://api.elaichi.ai/mcp, so there is one endpoint to list rather than one per client. For any other client, Elaichi requires OAuth dynamic client registration, PKCE with S256 and a browser sign-in. Test it before you write it down.
Each person can also list their own connected apps under Settings, then Connected apps. The columns are the app, the access it holds, the toolboxes it reaches, when it was connected and when it was last used.
Where data sits belongs here too. An organization picks EU, US or APAC when it is created, and that choice cannot be changed later. EU and US pin the organization's data store to that jurisdiction. APAC uses best-effort placement instead. Connector credentials are held by a separate credential service and encrypted at rest with AES-256-GCM.
Measure: one entry per call, and four things it will not tell you
Measure is what you track and how you know. Every connected-tool call that reaches execution writes one audit entry.
That entry names the person, the toolbox the call went through, the tool's advertised and catalog names, the connector and the connection the call actually reached. It also records the tool's classification, whether approval was required and how it was given, and whether a person edited the arguments. The surface is on the row as well, along with the MCP client when the surface is MCP, and the outcome. The in-product audit log is not plan-gated, so an organization on any active plan has it, Personal included.
Four limits matter more than the field list, because a careful reviewer finds them anyway.
A call refused before execution writes no entry. A tool a restriction withheld, a missing scope, an unknown tool name and a rate limit all stop ahead of the executor. The log records what ran, not every attempt.
Argument values are not recorded. The exception is the one path argument that names the object, kept as the target id. A row can therefore say which record was touched without holding the payload.
A connected-tool row carries no duration, so latency is not a measure you can take from it.
Listing tools and searching them are not audit events. "Full audit logging of the endpoint" is the wrong phrase for a policy document.
Retention is 90 days. For a longer record, forward it to your own log service: export to your own Datadog comes with the Black plan. How long each kind of history is worth keeping is the subject of audit log retention for AI actions. Read access is part of the same control. A member holding audit:view reads every event in the organization, and a member without it reads their own events plus events on connections they own or can edit.
Manage: restrictions, access requests and how fast a change lands
Manage is acting on what Map and Measure told you. Three controls carry most of it, plus one timing fact that belongs in the runbook.
Restrictions decide which connectors and tools a person can reach. A rule is allow or block. In Elaichi, restriction targets are role or user only; the organization default is the absence of a rule, which means allow everything. A team's limits are written on the role its members hold, or on each member. Rules on the role and rules aimed at the person both apply, and the personal layer can only narrow what the role allows. Within a layer, a block always beats an allow. One trap: an allow rule that names nothing blocks every connector and tool for its target, so an empty placeholder is a lockout.
Access requests are the exception path. Any member can file one, from the console or from their AI client. Only a person in the console can decide it, under Governance, then Access requests, holding member:manage and confirming their identity first. No AI surface can approve or deny a request, denials included. Approving records an access grant for that one requester and leaves the role's rules alone.
Approval before an action depends on where the call comes from. Over MCP there is no per-call approval step. The OAuth grant is the approval, chosen on Elaichi's consent screen for that named client. A confirmation prompt somebody sees in a chat is their client's own prompt, not Elaichi's. Approval cards inside Elaichi, and approve steps inside workflows, come with the Black plan.
Timing is the detail most runbooks get wrong. A role, team-membership, restriction or access-grant change takes effect within about two minutes. Revoking a grant, removing or suspending a member, revoking a share and disconnecting an account are effective on the next call.
Offboarding is the Manage control reviewers ask about first. Removing a member revokes their MCP grants and API tokens in that organization and ends their membership. It does not remove their accounts in the connected apps, which still have to be deprovisioned there. The order of operations is in what a departing member leaves behind.
Where the Generative AI Profile changes the answer
NIST AI 600-1 organizes suggested actions under the same four functions, for risks specific to or made worse by generative AI. For an agent that calls tools, one risk is structural. The model reads text out of a tool result and may act on what it read. The OWASP Top 10 for LLM applications (checked November 2026) catalogs the same class of risk for builders.
Say plainly what your layer does not do. Elaichi's MCP endpoint has no prompt-injection defense of its own. It never receives the person's prompt, only a tool name and arguments, so it has nothing to judge a call against. That defense belongs to the client. The in-app Elaichi Agent does scan tool results for injection patterns and holds its own writes behind a second check, and the Agent comes with the Black plan. Heuristic detection is not a guarantee, and a Manage entry that claims otherwise will not survive a second reading.
The controls that genuinely bound this risk are duller. A grant limited to chosen toolboxes. A role whose allow rules name only the read tools the work needs. A connection that is not shared more widely than the job requires.
When you do not need an access layer for this
If one team uses one vendor's own MCP server, each person signs in with their own account, and that vendor's admin console answers who did what, the Map and Measure records for that app already exist. Adding a second vendor to produce them again is work with no return. The framework asks for records, not for a product.
The case changes when the answer has to span apps from different vendors. Then the questions get harder. Is the same role honored everywhere, can a reviewer filter in one place, and does one offboarding action end access across all of it.
If nobody has connected an AI client to a company app yet, the honest Map entry is empty. Write the policy and the approval path first. MCP governance: who may do what, in which app sets out those decisions in order, and what an AI agent audit log must capture is the field list to hold any vendor to.
The crosswalk, in four lines
| Function | What a reviewer asks | The record |
|---|---|---|
| Govern | Who set this, and who can change it? | One role per member, custom roles on Gold, restriction management held by Org Owner and Org Admin |
| Map | What is connected, and as whom? | The connector catalog, per-person grants under Connected apps, each tool labeled Read only, Makes changes or Can delete data |
| Measure | What gets recorded? | One audit entry per tool call that runs, naming the person, connector, connection, client and outcome |
| Manage | What do you do about it? | Restrictions on a role or a member, access grants a person decides, 90 days of history with forwarding for longer |
Two neighbors to this post: ISO 42001 and AI agents covers the management-system version of the same evidence, and deployer duties under the EU AI Act covers obligations that are not voluntary at all. To see what a given app exposes as tools, browse the connector catalog, and the governance use case shows the setup end to end.