/

AI assistants & agents

For Operations teams

AI that works your queues, inside your rules

Add a model step to any app, or give an Operator a queue to work: triage claims, draft replies, flag payouts. Each one reads only what its role allows, logs every step and hands anything outside its limits to a person.

Halden Mutual Claims Triage: four KPI tiles, a claims table with risk-score bars, an open Assign to dropdown and an adjuster load list

Scoped like a new hire

An Operator acts on a named person's authority, limited to specific tables and actions. It can't see anything that person couldn't.

A person decides the rest

Anything outside the Operator's limits, or above a threshold you set, goes to a person for approval with the reasoning attached.

Every step on record

Prompts, tool calls, reads and writes are in the Logbook, and you can revoke an Operator's access in one click.

01 — In production

How Halden Mutual cut claims triage from 3 days to 4 hours

“Our auditor asked who could see claims data. We exported the answer from the Logbook before the meeting ended.”

Marit Solberg

Head of IT Risk, Halden Mutual

Read the story

Close crop of hands reviewing printed claim documents at a dim office table
Halden Mutual logo

Claims triage cut from 3 days to 4 hours

Close crop of hands reviewing printed claim documents at a dim office table
Halden Mutual logo

Claims triage cut from 3 days to 4 hours

  • Kettleby Freight logo
  • Halden Mutual logo
  • Tessaly logo
  • Brackwater Energy logo
  • Quenby logo
  • Fennimore Health logo
  • Saltmarsh & Co. logo
  • Brennock Precision logo

03 — What it includes

What keeps an AI step accountable

Your models, your keys

Use Anthropic or OpenAI models on your own API keys, and switch models without rebuilding the app.

Permission-aware context

Prompts include only the rows and fields the user or Operator is allowed to see.

Fields models never see

Write "never send diagnoses to a model" once as a Spec, and every prompt in every app leaves them out.

Structured output

Model responses fill forms and tables in a set schema, ready for a person to check and accept.

Budgets and limits

Cap tokens, cost and actions per Operator per day, and stop it automatically at the limit.

Replay any run

Open any Operator run to see each step, the data it read and the change it proposed or made.

05 — FAQ

What teams ask about AI agents

What's the difference between an AI step and an Operator?

Which models can we use?

Can an Operator make changes on its own?

How do we keep sensitive data from a model?

How do we turn an Operator off?

Is every model call recorded?

Give an Operator its first queue, and keep the final say

Create a free website with Framer, the website builder loved by startups, designers and agencies.