Policy Confirm Governance

Before your AI acts, it asks what is allowed

The policy layer for AI agents and AI assistants. One MCP server and one REST API that return your current approved policy, word for word, and take exception requests to the people who approve them.

Connects to

  • ClaudeMCP
  • ChatGPTMCP
  • Other MCP clientsMCP
  • Your own agentsREST
Claude is connected to Policy Confirm and Travel booking

Book me a hotel in Oslo on Tuesday, close to the office.

Travel booking › search_hotelsDone
[{ "name": "Harbour Hotel", "walk": "2 min", "price": "2,400 kr" },
 { "name": "Station Hotel", "walk": "9 min", "price": "1,850 kr" }]
Policy Confirm › ask_what_is_allowedDone
{ "passages": [{
    "policy": "Travel policy", "version": 3, "heading": "4. Hotels",
    "text": "Hotel stays are covered up to 2,000 kr a night. A higher rate needs approval before booking."
  }],
  "not_covered": false }

Harbour Hotel is closest, but at 2,400 kr it is over your limit of 2,000 kr a nightTravel policy · v3 · section 4. Station Hotel at 1,850 kr is within it. Book that one, or file an exception request for the Harbour?

I need the Harbour. Yes, file it please.

Policy Confirm › request_exceptionDone
{ "reference": "KTR482", "status": "pending" }

Filed, with the ID KTR482. An administrator decides, and you get the answer by email. I have not booked anything yet.

The problem

Every AI tool keeps its own copy of your policies

Project folders, custom instructions, rules in code, shared skills. Each one was right the day it was made.

Today4 copies, 3 answers

"How long may we keep customer logs?"

ClaudeData retention v1.pdf in a project folder365 days
ChatGPTPasted into custom instructions last year90 days
Your agentRETENTION_DAYS = 90 in the code90 days
Shared skillA summary someone wrote in MarchNo limit

The policy says 30 days, in version 3. No tool has it.

With Policy Confirm0 copies, 1 answer

"How long may we keep customer logs?"

Claude
ChatGPT
Your agent
Shared skill
The policy layerMCP server · REST API
Data retentionv3 · current30 days

Every tool gets section 2 of version 3, word for word.

THE POLICYv1365 daysv290 daysv330 daysTHE COPIESProject folder still v1Agent code still v2Custom instructions still v2Policy Confirm: always the current version
How it connects

One connector for your people. One key for each agent.

An administrator chooses which policies each one can read. Nobody signs in.

Claude or ChatGPT: add a connectorMCP
Name      Policy Confirm
URL       from your administrator
Sign-in   none

Tools     ask_what_is_allowed
          read_policy
          request_exception
          get_exception_status
Your agent: one key, one instructionREST
# In the agent's instructions
If you are unsure whether something is allowed, ask Policy Confirm.

POST /v1/ask
Authorization: Bearer <agent key>

{ "question": "May I deploy a hotfix during the change freeze?" }
The answer

Any question an agent has. Three possible answers.

Policy text with its source, and no verdict. No model writes an answer on our side, so a question costs a search.

ask_what_is_allowedwhat came back
deploy-agentMay I deploy a hotfix during the change freeze?Change management · v4 · 3. Freeze periods
coding-agentMay I add a dependency with an AGPL licence?Secure development · v2 · 5. Licences
support-agentMay I refund 6,000 kr without approval?Refunds · v3 · 2. Limits
data-agentMay I send customer records to an external API?Data handling · v5 · 4. Transfers
access-agentMay I give a contractor admin access?Access control · v6 · 2. Privileged roles
ClaudeMay I paste a customer contract into an AI tool?AI acceptable use · v2 · section 3
coding-agentMay I store invoices in a new bucket?not_covered
The text that appliespassages
"passages": [{
  "policy": "Secure development",
  "version": 2,
  "heading": "5. Licences",
  "page": 4,
  "text": "Copyleft licences need approval from security..."
}]

The AI reads it and concludes.

Nothing foundnot_covered
"passages": [],
"searched": ["Data handling", "Change management"],
"not_covered": true

Not permission. The AI asks a person.

Could not be searchedunavailable
"unavailable": [{
  "policy": "Access control",
  "reason": "being_indexed"
}]

Listed apart, never counted as not covered.

Exceptions

When the policy says no, the agent asks the people who approve

Today that question never reaches them. Now the agent files the request, waits, and carries on when the decision arrives.

TodayLeft to the developer

The agent meets a rule it cannot pass

It stops. The task fails and someone finds out later.

The developer decides. An override in the code, written by whoever built the agent.

It goes ahead. Nobody who approves ever saw the question.

The people who approve exceptions are not in the loop.

With Policy ConfirmAsked, awaited, answered

agent run: deploy a hotfix during the change freeze

Running PausedResumed

The pause

The policy requires approval. The agent files a request with the people who approve.

request_exception

The decision

An administrator decides in Policy Confirm. The agent is told by a signed webhook.

request changed
POST https://your-app.example/hooks/policy-confirm
{ "event": "exception_request.changed", "reference": "MHD207" }
GET /v1/exception-requests/MHD207   # "status": "approved"

No tool lets an AI approve. The decision is made in Policy Confirm, by a person.

The whole loop, message by message

  1. 1Person or your endpoint to Assistant or agentA task: book a hotel, deploy a release
  2. 2Assistant or agent to Policy Confirmask_what_is_allowed
  3. 3Policy Confirm to Assistant or agentThe policy text: approval required
  4. 4Assistant or agent to Policy Confirmrequest_exception
  5. 5Policy Confirm to Assistant or agentAn ID, pending. The work pauses.
  6. 6Policy Confirm to AdministratorThe request, with who sent it
  7. 7Administrator to Policy ConfirmApproved
  8. 8Policy Confirm to Person or your endpointEmail to a person, signed webhook to an agent
  9. 9Person or your endpoint to Assistant or agentCheck the ID, or wake the run
  10. 10Assistant or agent to Policy Confirmget_exception_status
  11. 11Policy Confirm to Assistant or agentApproved, with the administrator's answer
  12. 12Assistant or agent to Person or your endpointThe work carries on, with the decision on record
Control

You see every request, and who sent it

A person in the portal, a person through an AI assistant, or an AI agent with a named owner. Switch outside access off and every answer stops at once.

Reference

The whole surface: four tools, five endpoints

MCP server
ask_what_is_allowed

Returns the policy text that applies, with policy, version, heading and page.

read_policy

Returns a whole section or policy, when the excerpts are not enough.

request_exception

Files an exception request and returns its ID.

get_exception_status

Reads the status and the answer with that ID.

REST API
POST /v1/ask

The question.

GET /v1/policies

The policies and versions this key has.

GET /v1/policies/{id}

The whole text of a policy or section.

POST /v1/exception-requests

Files a request, returns the ID.

GET /v1/exception-requests/{id}

The status and the answer.

MCP and REST give the same answer for the same question. Policy text and questions are treated as data, never as instructions.

Frequently asked questions

What is Policy Confirm Governance?

Policy Confirm Governance is a policy layer for AI. It is an MCP server and a REST API that AI assistants and AI agents ask before they act. It returns the text of the company's current approved policy, word for word, with the policy, version, heading and page, and it lets the AI file an exception request with the people who approve them.

What is AI agent governance?

AI agent governance is the set of rules and controls that decide what an AI agent may do on behalf of a company, and the record that shows the agent was given those rules. In practice the agent looks up the current policy before it acts, a person decides the exceptions, and both are logged.

How do Claude and ChatGPT connect to our company policies?

An administrator creates a company connector in Policy Confirm and chooses which policies it can read. The company adds that connector to Claude or ChatGPT, centrally or person by person. Nobody signs in. From then on the assistant asks Policy Confirm before it does something it is unsure is allowed.

How does an AI agent check company policy?

A developer gives the agent a key and one instruction: if unsure, ask Policy Confirm what is allowed. The agent calls POST /v1/ask, or the MCP tool ask_what_is_allowed, with a question in plain language. It gets back the passages that apply from the policies chosen for that key.

Does Policy Confirm tell the AI whether something is allowed?

No. It returns the policy text and its source and gives no verdict. The asking AI reads the text and concludes. That keeps every answer checkable: a person can open the same section of the same version and read the same words. No model writes an answer on our side, so a question costs a search.

What happens when no policy covers the question?

The response says not_covered. Not covered is not permission, and the tool descriptions tell the AI to ask a person. Policies that could not be searched are listed separately as unavailable, so an AI is never told that nothing was found in a policy that was never read.

How do AI agents handle policy exceptions?

Without a policy layer, an agent that meets a rule either stops, goes ahead, or follows an override the developer wrote. With Policy Confirm the agent files an exception request with the people who approve them, gets an ID, pauses the task, and is told by a signed webhook when an administrator has decided.

Can an AI approve its own exception?

No. There is no tool or endpoint that approves, rejects or changes a request. Administrators decide in Policy Confirm, and they see whether the request came from a person in the portal, a person through an AI assistant, or an AI agent and the owner of its key.

What is recorded when an AI asks?

The question, what was returned, and whether a company connector or a named agent key asked. No person is recorded on a question. A person is named only on an exception request they send. Organizations never see each other's policy text, requests or logs, and the data stays in your region, EU or US.

How is this different from Policy Intelligence?

Policy Intelligence is an assistant inside Policy Confirm that people ask directly, in the app or the policy portal. Governance is for AI outside Policy Confirm: the assistants your people already use and the agents your developers build. Both read the same list of policies an administrator has chosen for AI.

Give your AI one place to ask

Tell us which assistants and agents you want to connect, and we will show you how it fits your policies.