Insurance
PolicyCompareIQWordingIQClassifierIQTriageIQD&O AnalyzerIQDiagnostics
AI-Ready BlueprintPlatform
FlowForgePartner practices
Databricks partnershipPlatforms we build on
Microsoft AzureGoogle CloudAnthropic ClaudeOpenAISnowflakeHow we choose
Regulatory frameworksVendor neutralityPlatform
FlowForge turns a submission, a claim or a bordereau into a flow you can see on one screen — every step named, every decision recorded, and a person in the loop wherever judgement belongs.
Submission intake & triage — as it appears on the canvas
How it works
A flow is built by the people who know the class of business, not by the people who know the platform. That is the whole design.
Open a worked flow for the job in front of you and change what is different about your desk. Nobody starts at an empty canvas, because nobody has time to.
Template gallery
Drag the steps you need, connect them, and set what each one is looking for. Rules that used to live in a spreadsheet and somebody’s memory become a diagram the whole desk can argue about.
What you place
Every run keeps its steps, its inputs, its outputs and the reason for each decision. When something is referred, the reviewer sees the evidence beside the question rather than being asked to trust a score.
What a run records
Architecture
Work arrives from the places it already arrives from, is handled by agents you defined, and is written back into the system that holds the record. Nothing is asked to move.
02 What handles it
Agents your team defined, running on your data platformDatabricks AI Functions · guardrails checked on every answer
03 Where the answer goes
The market platforms appear at both ends because they belong at both ends — a risk comes in and a response goes back out over the same connection, configured once.
Turn what arrives into fields you can act on, whatever shape it came in.
Agents you define, in your words, answering in a shape the next step can use.
Referral is a designed step, not a fallback for when the machine is unsure.
Where the work goes when it is done, and what happens when it cannot get there.
Implementation
Build, prove and run are separate places with separate access. A flow moves forward when somebody decides it should, and the move is recorded.
Authors work against sandbox connections and test data. Nothing here can reach a live system, because the live connections do not exist in this environment.
A promoted copy, run against real-shaped volumes. The behaviour under test is the behaviour in production, because the settings that govern behaviour travel with the flow.
Live traffic, live systems of record. Promotion into it is a decision taken by somebody who holds that right here, not one inherited from being an author elsewhere.
Integration
The market platforms, the mailboxes and the file drops your business already runs on — plus an open connector for everything we have not met yet.
Work arrives
Work is delivered
Source and target, configured once per platform.
Polled, watched, or pushed to — whichever the sender supports.
Written back as clean records, not as attachments.
Every screen in the product is an API call you can make yourself.
Security & governance
Each business area and each environment is its own space. A flow, an agent or a connection built in one is not visible from another, and cannot be reached across the boundary by anything.
Nothing you run trains a shared model.
People are given access to an organisation, to a class of business, or to both — and the role can differ by environment. Author in development and reviewer in production is an ordinary arrangement here.
Single sign-on through your existing identity provider.
Every credential lives in a managed vault and is fetched at the moment it is used. They are set per environment and never travel between them — a production secret cannot be read from a development space, whoever is asking.
Encrypted in transit and at rest.
A run keeps its steps, the documents as received, each decision and the person who took it. When an agent produced a finding, the finding is advice attached to the evidence — the decision is still recorded against a name.
History belongs to the environment it happened in and is never moved.
FAQ
No. The people who configure a flow are the people who know the class of business. An agent is given its instruction in plain English by an underwriter or a claims handler; there is nothing to train and no model to maintain.
It is caught the same way a junior’s work is caught. Guardrails are set per agent and checked on every answer, low-confidence work is referred to a named reviewer, and the reviewer sees the source document beside the question rather than being asked to trust a score. The decision on the record is the person’s, not the agent’s.
If it has an API, yes. There are built connectors for the market platforms and for Microsoft 365, and an open REST connector for everything else. Where a system only accepts files, a flow can write them to SFTP or a folder in the shape it expects.
It stays in your tenancy. The platform deploys into your own Azure subscription, and documents and runs are held in storage you control. Nothing you process contributes to a shared model.
That depends on the flow and on how quickly the connections can be agreed with the systems either side — which is usually the longer half. We would rather scope it against your actual submissions than quote a number here.
Yourselves, and that is the intent. We build the first with you, on your data, and hand over the canvas. The template library exists so the second one does not start from nothing.
Platforms we build on
Databricks Bronze partner, with production delivery on every platform named here.
