CaseFlow
The governed data and decision platform.
CaseFlow connects the systems your organization already runs, turns their records into the objects your teams talk about, and runs the models and agents that help them decide — with a person approving every action that has a consequence, and a record of every step.
01 · Overview
Four parts. One owner for every resource.
Each part of CaseFlow owns one kind of resource, and every change goes to the part that owns it — so two copies never disagree about what is true. Access follows each person’s roles and clearance.
| Resource | Owned by | What it means |
|---|---|---|
| Sources, files, datasets and their history | Data Plane | Raw data and previews stay here; the ontology receives what you choose to publish. |
| Business objects, links, actions, decisions and duplicate matches | Ontology | One place of record. Matches are proposed by an engine and accepted as decisions. |
| Models and how they are served | Model Studio | Every model is deployed the same way, so you can see and undo any change. |
| Agents, their tools, runs and workflows | Agents and Workflows | Each agent version is fixed to a model and to approved connector versions. |
02 · Data Plane
Bring data in once, with its owner and its history.
The Data Plane connects your sources, prepares what they send and commits versioned datasets with their lineage. Nothing reaches the ontology without passing through it.
Connections
Every source, with its owner and its credentials kept safe
- Databases
- Files and spreadsheets
- Documents and scans
- REST APIs
- Google Workspace
- Email and SMS
Data Builder
One pipeline, several views, reviewed like code
- Visual graph
- SQL
- Code
- Diff between versions
- Change timeline
- Branches and review
Intake and quality
Messy files in, checked datasets out
- Table and header discovery
- Validation
- Error quarantine
- Field extraction
- Text recognition
- Managed tables
Lineage and operations
Know where every value came from, and how builds are running
- Dataset lineage
- Access map
- Schedules and chains
- Build progress
- Slow-run alerts
- Evidence search
Bad rows are quarantined, never silently dropped, and every dataset can name its source and every step since.
03 · Ontology
Model the things your organization talks about.
The ontology turns datasets into objects — cases, sites, people, assets — and the links between them. It is the one authority for business objects, actions and accepted identity decisions.
Objects and links
The things your teams talk about, mapped from datasets
- Object types
- Link types
- Property mappings
- Object views
- Graph explorer
- Path search
Versions
One published head; every run pinned to it
- Drafts and review
- Publishing
- Compatibility checks
- Loads tied to one version
- Ontology lineage
- Mapping checks
Actions and functions
Governed verbs, with a person on the consequential ones
- Action types
- Forms and rules
- Effects
- Approvals
- Functions
- Testing before release
Duplicate matching
Tell two records of one person apart from two people
- Candidate pairs
- Explained comparisons
- Review inbox
- Runs you start
- Merge decisions
- No automatic merging by default
Every load is tied to one published version. If the data model changes during a load, the load stops and asks to be run again.
04 · Model Studio
Every model in use, deployed one way.
Model Studio deploys every model in use — extraction, classification, detection, ranking, embeddings and the language models agents rely on — and every one is deployed the same way.
Registry
Each model with its versions and contract
- Versions
- Input and output schema
- Documentation
- Governed catalogue
Serving and change
One way to serve, changed in steps you can undo
- Deployments
- Staged versions
- Promotion
- One-step rollback
Oversight
See who depends on a model, and everything that changed
- Version comparison
- Who depends on it
- Access grants
- Deployment audit
You can always tell which model version produced an answer, and return to the previous one in a single step.
05 · Agents
Agents that do real work, and ask before they act.
Managed agents read through the ontology, use only the tools you attach and draft work for people. Every tool is an approved, pinned connector release — including our Google Workspace connector, described in full in our Privacy Policy.
Build
Each agent version is reviewed and fixed in place
- Model from your catalogue
- System policy
- Skills
- Revision history
Tools
Only the connectors you attach, at versions you approved
- MCP connectors
- Verified connector versions
- Google Workspace
- Your own APIs
Control
A person approves every effect
- Approvals queue
- Which objects it may read
- Spend limits
- Baseline policies
Observe
Replay any run, step by step
- Run debugger
- Linked sources
- Longer research runs
- Isolated code execution
Agents propose; people approve. Keys go to a connector when it starts and are never shown to the AI model.
06 · Workflows
Automate the routine. Keep people on the decisions.
The Workflow Builder lays out triggers, checks, approvals, waits and notices on one canvas, and Explore turns the results into analysis and dashboards that name their sources.
Workflows
Automate the routine; keep people on the decisions
- Visual builder
- Triggers and intake
- Human steps
- Email and SMS notices
- Run history
- Stuck-run alerts
Explore
Analysis where every figure names its source
- Analysis
- Charts and maps
- Dashboards
- Compare
- Files and media
- Search
No automation switches itself on: publishing or activating one is an authenticated person’s decision, recorded.
07 · Security
It runs where your data lives.
Sensitive operations should not have to move their records somewhere new to get better decisions. Read more on our Security page.
Deployment
Runs where your data lives
- Your servers
- Your cloud account
- A dedicated environment we run for you
- One codebase for every option
Access
Mirrors how your organization is structured
- Roles
- Per-person clearance
- Workflow duties
- Sharing groups
Assurance
What you can verify, not just what we say
- Audit log
- Release records
- No standing access
- No training on your data
Every hosting option runs the same code, and each organization’s data is kept separate from every other’s.
08 · Questions
Common questions.
Straight answers. If yours is not here, ask us.
Do you train AI models on our data?
No. We never use your data to train or improve models for us or anyone else. The one exception is yours to make: on your written instruction, a model can be adapted with your own data inside your deployment, for your use only — and Google user data is never used for that.
Where does our data live?
Wherever you choose: on your own servers, in your cloud account, or in a dedicated environment we run for you. In every case our staff hold no standing access to it.
Can an agent act without a person?
An agent can read what it is cleared to read and draft work. Anything that sends, changes, publishes or approves waits for a named person, and the run trace records who approved it.
Can we use our own models?
Yes. Deploy them through Model Studio, alongside any external provider your administrator approves. Agents can use only the models in your governed catalogue.
Does CaseFlow replace the systems we already run?
No. It connects to them through governed connections and keeps each source’s owner and permissions. We start by improving one decision, not by replacing your tools.
What happens when a model or an agent is wrong?
Its output is a proposal with its sources attached, so a reviewer can see why. A person approves before anything happens, a model version can be rolled back in a step, and every run can be replayed.
How do we start?
Bring one recurring decision. We agree the scope and how success will be measured, run a pilot on synthetic data unless you choose otherwise, and decide on production together afterwards.
Is CaseFlow certified?
Not yet. Our Security page lists every third-party attestation, with its scope and date, once it is issued.
See CaseFlow working on your own question.
Send us the decision, the sources it draws on and who signs it off today. We will come back with how CaseFlow would handle it, and what a pilot would measure.