> ## Documentation Index
> Fetch the complete documentation index at: https://docs.inviolet.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Intent Detection and Response (IDR)

> The category Inviolet defines: access governed by why, not just who.

Identity Access Management (IAM) tells your systems *who* is allowed to read a
piece of data. **Intent Detection and Response (IDR)** tells your systems *why* that read is
allowed — and then proves it.

## The problem IAM doesn't solve

A senior support engineer is allowed to read the `customers` table. That's
their role; that's the IAM permission. But:

* Should they read `customers.ssn` while answering an Email subject line "where
  is my package"?
* Should they pull 50,000 rows in one query when their pattern is 1–5 rows?
* Should an LLM agent acting on their behalf see the same set of columns as
  the human, or a smaller set scoped to "respond to one customer query"?

IAM can't answer those questions. The role is "support engineer." The data is
"customers." The decision should be "yes, but only the columns you actually
need for *this* intention, and only for the duration of *this* request."

## What Intent Detection and Response (IDR) adds

Inviolet introduces a third dimension between identity and data: **intention**.

```
            ┌─────────────┐
            │  Intention  │  ← "Reset a customer password"
            └─────────────┘
                  │
       ┌──────────┴──────────┐
       ▼                     ▼
 ┌──────────┐          ┌──────────┐
 │ Identity │          │   Data   │
 └──────────┘          └──────────┘
```

Every tool call an agent (human or AI) makes carries a declared intention. The
intention is matched against policy: which data elements are allowed for *this*
intention, by *this* identity, at *this* time. The match is logged. The
credential issued to fulfill the request has a TTL bounded by the intention.

## Why now

Three trends collapse onto each other to make Intent Detection and Response (IDR) urgent:

1. **AI agents** can do things 10,000 times faster than humans, and they don't
   pause when something feels off.
2. **LLM-driven workflows** generate tool calls a human never wrote — so the
   "intention" is implicit in the prompt and gone the moment the LLM runs.
3. **Data sprawl** means the question "should this query happen" requires
   judgment your DBAs no longer have time for.

Intent Detection and Response (IDR) moves the judgment up the stack. The agent makes the
request, declares an intention, and the system says yes-with-scope or
no-with-reason. The human oversees the policy, not every query.

## What Inviolet ships

* **Intent extraction** — every LLM tool call is parsed into a structured
  intention record before it leaves the gateway
* **Policy engine** — intentions are matched against declarative rules
* **Intent tokens** — short-lived JWTs carry intention context downstream
* **Database proxy** — column-level enforcement at the query layer
* **Decision feed** — every allow/deny is auditable and queryable in real time

## Read next

* **[Mandates](/concepts/mandates)** — the credential format that carries intention
* **[The six enforcement layers](/concepts/six-enforcement-layers)** — where each layer sits in your stack
* **[The maturity model](/concepts/maturity-model)** — how teams move from observation to enforcement