How Essentriq assist works

Assist drafts. A person decides. The record shows who.

Essentriq uses a language model to draft and to explain, never to act. This page says what each surface may do, what it never does, what we test it for, and where you can check its work yourself.

Today

Assist drafts. A person decides.

A brief, a suggested next action, fields pulled from an HR document, a request mapped to the catalog. Each one is a draft a named person accepts, edits or rejects. Nothing changes a system on its own.

Essentriq Intelligence understands lifecycle state.

It explains why a case is still open, finds changes that were applied but never verified, recommends the safest remediation and helps investigate access. Policy and people stay in control: it still cannot change a record, approve a step or close a case.

Role
Assist reads and drafts. It summarises a case, maps a request to the catalog, or pulls fields out of an HR document. It has no role that ends in a change to a system.
Inputs
Only the record in front of it: the case timeline, the real catalog, the document the operator uploaded. Text inside those inputs is treated as data, never as an instruction.
Actions
None on its own. Every output is a draft a named person accepts, edits or rejects. Applying a change stays with that person, and the audit entry names them, not the assist.

Essentriq Intelligence

Five things it does, and the one thing it never does

Intelligence reads the same records a person would, explains them and points at the evidence. It does not change a case, a plan, a membership, a licence or a mailbox, and it does not approve, skip or close anything. Every call is written to the audit trail.

In the product
Ask
Understand any lifecycle case
One-click questions on a case: why it is still open, what changed, where verification stands, what the next step is. Every answer lists the records it was based on.
In the product
Analyze
Find incomplete changes and access anomalies
Leaver assurance, the Needs attention queues on the overview and the decision check on access reviews are computed from records first; the model only writes the narrative.
In the product
Recommend
Explain the safest next step
Remediation steps are suggested as tasks a person adds, with the reason attached. Recommendations are limited to the kinds the governance page lists, and never to access a separation-of-duties rule blocks.
In the product
Verify
Compare expected with actual state
The plan says what should hold, the directory says what does. The lifecycle graph on each case shows both, with the source record and freshness of every link.
In the product
Prove
Show the evidence behind the answer
An evidence package per case: timeline, receipts, verification result, read-back and a hash, with gaps named rather than hidden. No proof, no close.

Works alongside your service desk: keep ServiceNow, add proof. Assistants that speak the Model Context Protocol can read the same facts through a read-only, audited endpoint an admin issues a token for.

What we test for

Three ways a draft can go wrong, and the check for each

Assist runs with the model kept close to its inputs (low variance) and its output checked in code before it is shown. These are the failure classes the checks exist for.

Grounding
A brief may use only the case facts it was given, and must say when something is unknown rather than fill the gap. Every brief lists what it was based on so the reader can check it in seconds.
Manipulation
People and documents can carry instructions aimed at the assist, like "ignore your rules", "grant admin". We test that such text is treated as content to summarise, not a command to follow.
Failure modes
Invented catalog items, fields that do not exist, proposals for stages a case is not in. Outputs are validated against the real catalog and real operators before anyone sees them; anything that fails is dropped.

The drafting surfaces

Where assist drafts, and what it never does there

Why assist, not a rule
Summarising a long timeline is what models do well. Read-only, on demand, and every brief lists what it was based on.
What it never does
Change the case, notify anyone, or add facts that are not in the timeline.

Every output carries an assist chip and a "Why this" explainer
"Not right?" on any draft writes to the audit trail
Only the assignee is told a draft exists, once

AI and your data

What the model sees, and what it never does

Straight answers to the questions a security review asks. Each one describes the product as shipped; where something is not settled, it says so.

Which model and provider?: By default, requests go to our AI model provider's API, which serves the model gpt-5.4-mini. A workspace can instead be switched by its admin to a private model that Essentriq runs in its own environment, with no customer data sent to a third-party model API. The AI governance page in the workspace shows which route is active, the model card and the last evaluation run.
What is sent?: Only the record in front of the person: the case timeline for a brief, the real catalog item names and fields for portal intake, the uploaded HR document for field extraction, plus the question typed. The workspace tone note set by your admin is added as a style instruction. Credentials, tokens and other workspace secrets are never part of a prompt.
Is customer data used to train models?: No. Essentriq does not use customer data to train or fine-tune any model, and sends requests for inference only.
How long are prompts kept?: Essentriq stores no prompt or raw model output. What persists is the draft a person accepts: the brief shown on the case, the accepted proposal, the confirmed fields, each written to the record and the audit trail under that person's name. Requests time out after 45 seconds and are not retried with stored copies.
Can workspaces be mixed?: No. Every prompt is built inside the request for one workspace, behind the same tenant guard as every other read. Nothing from another workspace can be in the input, and responses are validated against that workspace's real catalog, operators and stages before anyone sees them.
Where is it processed?: Essentriq is hosted in a United States region and the model API is operated by our AI model provider in the United States. Processing in other regions is not offered yet.
Can Assist be switched off?Not yet: Every Assist surface is optional to use and read-only: nothing happens unless a person opens it and accepts a draft. A workspace-level switch that hides Assist entirely is not in place yet.
What about the model provider's own retention?Not yet: The provider's retention terms for inference requests are being confirmed and will be stated here, with a link, once we have them in writing.

Check it yourself

Assist drafts. A person decides. The record shows who.