Relevant does not mean permitted.
A matching document still needs the caller’s tenant, purpose and source permissions checked before it enters the decision.
Governed data. Business
context. Connected evidence.
The right data.
The right scope.
Ready for action.
Every decision needs the right records, documents and numbers—checked for permission and freshness, with their source attached.
Bring those inputs together for existing agents, policies, human reviewers and applications built in Frozion. Governed tools and workflows are available through MCP, with workflows also callable over HTTPS.
BUSINESS REQUEST
Review a supplier bank change
Caller + delegated authorityTenant + purpose
01
Operational recordSupplier · legal entity · current case
Scoped
02
Retrieved evidenceDocuments + related business records
Eligible
03
Calculated fact2 bank changes in the last 30 days
Versioned
FROZIONExecution FabricGoverned from request to outcome
Governed context package
What this decision may use
Selected information, source references, time, scope and provenance.
PolicyHITL reviewAgent or model
Connected to the execution that uses it
One foundation for existing agents and native applications
Typed recordsKeyword + vectorFeature StoreScoped memoryProvenance & lineage
The AI Data Plane across your adoption path
Govern existing agents
Supply eligible records, evidence and calculated facts to a governed agent action.
Grow tools into workflows
Reuse scoped queries and facts across the steps of a business process.
Build native agents
Give native agents context with permissions, sources and freshness attached.
Build systems of engagement
Applications use the same governed business context and retained references.
More than relevant data
The decision needs context
your business can trust.
A useful search result is only part of the answer.
The agent, policy and reviewer need to know which records they may use, whether a number is current enough and where each input came from. Those checks become harder when every consumer reconstructs the context separately.
One number should have one meaning.
Give agents, policies and reviewers the same defined fact, with its entity, window, version and freshness attached.
The input needs a source.
Connect the record or document revision to the fact, decision and eventual outcome, so teams can inspect what was used.
From knowing to doing
A richer tool response.
A stronger business decision.
An agent may know which API to call. The business still needs to know which records, facts and evidence justify the next action.
When an existing agent invokes a Frozion workflow through MCP or HTTPS, that workflow can assemble permitted data, calculate or retrieve business facts, evaluate policy, involve a reviewer and continue durably. The Data Plane supplies context and evidence; the Control Plane governs the work that uses them.
The same foundation supports agents built in Frozion, App Builder experiences and tenant-specific SaaS applications. Native agent development is optional.
AI DATA PLANE
What may this decision use?
Business records, eligible documents, reusable facts and their provenance.
AI CONTROL PLANE
What may this execution do?
Per-step authority, policy, HITL with FIDO2 step-up, durable continuation and tool calls.
These controls apply to the data access and execution paths routed through Frozion. Reading an eligible fact does not, by itself, authorize a business change.
Inside an existing agent's tool call
Inspect the context
before the action.
This is the Data Plane’s part of CASE-041. Check which inputs this supplier change may use, then see what changes when a required fact is stale or the caller lacks access.
Supplier bank-change review
01 / CHECK THE INPUTSRead scope verified
GlobalCo · Supplier-change review · Delegated supplier access
-
Supplier recordACME-204 · current operational revisionEligible
-
Verification evidencePermitted document + source revisionEligible
-
Bank changes / 30 days2 · definition v3 · within freshness windowCurrent
Scope controls eligibility. Freshness determines whether
eligible information is current enough for this decision.
Ready for policy evaluation
One package. Distinct data roles.
The workflow can use the current supplier record, eligible evidence and the same versioned fact without losing what each represents.
- What is available
- Operational record + document reference + calculated fact.
- What travels with it
- Identity and tenant scope, business purpose, source revisions, fact definition, time and provenance.
- Next governed step
- Evaluate the supplier-change policy. In this example, repeated bank changes require an authorized finance reviewer.
Context is ready; the proposed write still needs its own
authorization.Conditions and records
01 / Data & context
Know which source owns each fact—and
how the decision uses it.
Keep ERP, CRM, databases and document systems authoritative for the information they own. Operational Collections provide typed business entities for applications; context assembly selects what a particular decision is permitted and ready to use.
A relevant document, a current balance and a remembered conversation can all be useful. They carry different meanings—and different rules for use.
Operational Collections & governed query
Query the business fact you need.
Work with typed entities, canonical business identities and explicit tenant and legal-entity scope. Define which sources own the requested fact and how joins, time, units and missing values should be interpreted.
Define governed queries over typed business entities, with an explicit authorization boundary and output contract. A visual query can produce the SQL or source-specific operation needed by the application.
Explore Operational Collections & Query Explore Collections & QueryCurrent supplier payment exposure
- Entity
- Supplier ACME-204 · legal entity US-01
- Authority
- Active tenant + permitted resources and fields
- Sources
- Supplier master + eligible payment obligations
- Meaning
- Open obligations · declared currency · current state
- Result contract
- Amount + unit + time + source references
The query definition makes the business meaning reviewable
before the result enters a decision.
Context for one decision
Create the package
this decision needs.
A context package is the set of inputs a particular decision is permitted and ready to use.
Your workflow combines eligible records, retrieved evidence, Feature Store values, instructions, user intent and permitted memory. Keep each item’s role and origin explicit through selection and transformation.
Explore Visual Orchestration-
01
Check eligibility and sufficiency.
Apply identity, tenant, purpose, taxonomy and source permissions. Check that the required facts are present and satisfy the decision's freshness rules.
-
02
Keep roles and precedence explicit.
Distinguish instructions, source evidence, calculated facts and memory. Define how conflicts and missing information are handled; retrieved content does not become an instruction merely by entering context.
-
03
Inspect what the decision actually used.
Trace the source references and transformations associated with the decision. Historical evidence explains what was used; inspection still follows current access permissions.
Seven kinds of data. Each with a clear job.
A current balance, a policy document and a remembered conversation are all data. They carry different authority, time and reuse rules. Frozion keeps their roles explicit.
01
Operational truth
The current business state held by the system responsible for it.
02
Retrieved evidence
Relevant, eligible material selected for the current question or decision.
03
Calculated fact
A derived value with an entity, definition, version, unit and time.
04
Decision context
The governed package of information selected for a particular use.
05
Memory
Prior information retained for possible reuse under explicit conditions.
06
Cache
A prior result eligible for reuse within its scope, versions and freshness.
07
Durable state
The persisted execution position and obligations that govern continuation.
02 / Knowledge & retrieval
Find the relevant information.
Keep its authority boundary.
Combine precise keyword matching with semantic relevance. Identity, tenant, purpose, taxonomy and source permissions constrain the candidates available to the request.
Retrieval finds eligible candidates. The workflow selects what belongs in the decision context, alongside authoritative records and calculated facts.
LEXICAL / BM25
Find the exact reference.
Match names, identifiers, policy terminology, phrases and codes.
Supplier numberACME-204
VECTOR / SEMANTIC
Find the related meaning.
Retrieve relevant passages when the question and source use different words.
Business question“Changed payment destination”
Hybrid retrievalKeyword and vector search together, with candidate access constrained by the requester’s tenant, purpose and source permissions before context is assembled.
Documents & object storage
Trace derived content
back to its source.
Follow the relationships among original files, chunks, embeddings, summaries and generated artifacts, with source identity, revision, classification, ownership, retention and deletion state in view.
Source changes matter downstream: a prior summary or index entry should be reconsidered when its source is revised, reclassified or no longer eligible. Specialized object storage can remain in place.
FROM RELEVANCE TO A BUSINESS DECISIONRelevant content is one input.
Relevant content is one input.
A decision needs the full context.
Retrieval supplies candidates. The application must also establish permission, current facts and the evidence that explains their use.
| Decision requirement | What relevance search answers | What governed context adds |
|---|---|---|
| Find information | Which passages match the question? | Combine eligible documents with the records and calculated facts the decision requires. |
| Check permission | A relevance score measures the match. | Check the caller’s authority, tenant, purpose and source permissions before assembly. |
| Interpret a number | A passage may quote a value. | Use a defined fact with its entity, window, unit, calculation version and freshness. |
| Handle stale inputs | A relevant passage may describe an earlier state. | Apply configured freshness rules: refresh, wait or follow the declared exception path. |
| Explain the decision | Identify which passages matched. | Connect source revisions, selected inputs and fact versions to policy, review and outcome. |
03 / Facts & reuse
Define a business fact once.
Use it across the decision.
A risk measure, payment exposure or recent-change count can inform a policy, model, application screen and human reviewer. The integrated Feature Store gives those consumers a defined fact with the information needed to interpret it.
Carry the calculation, entity, source, unit, time, version and provenance with the value. Keep the values used by an execution connected to that execution.
The integrated Feature Store
A number becomes useful
when its meaning travels.
Calculate facts from eligible data using defined formulas and windows. Reuse them in deterministic rules, deeply nested policies, routing, context packages, reviewer screens and permitted AI tasks.
Separate point-in-time meaning from freshness. A historical value may be correct for explaining yesterday's decision while being unsuitable for approving today's change.
Explore the Feature Store
Policy BuilderHITL InboxModels & agentsApp Builder
supplier.bank_changes_30d
2recorded changes
for supplier ACME-204Definition v3
for supplier ACME-204Definition v3
A defined calculation for a defined entity.
- Definition
- Count of recorded bank-account changes
- Window
- Rolling 30 days ending at the decision time
- Entity & scope
- ACME-204 · GlobalCo · legal entity US-01
- Unit
- Number of changes
The rule and the reviewer
can inspect the same versioned fact.
FACTS FROM GOVERNED ACTIVITYTurn what already happened
Turn what already happened
into the next decision’s context.
Define counts, totals and rates from recorded activity: refunds by an agent in an hour, payments to a supplier in seven days or records read by a user. Policy can use those features alongside facts from your systems of record.
Feature workflows run on a schedule or when invoked through an API. You define the materialization and freshness requirements.
Explore rules that know the historyA FACT THE RULE CAN USE
agent.refunds_count_1h10 recorded refundsEntity + one-hour window + definition version + materialization time
Policy checks the recorded count before the next governed action.
A materialized count reflects the recorded state at its last update. Concurrent decisions can read the same value; it does not reserve capacity.
04 / Lineage & continuity
Trace the evidence from
source to outcome.
Follow the relationships among source data, selected context, feature values, policy evaluations, human decisions, tool attempts and reported outcomes.
The Data Plane retains the evidence relationships. The Control Plane's inspectable execution graph and Behavior Intelligence use them to explain the process and support governed replay.
SOURCE RECORDSupplier revisionEntity · tenant · source
CALCULATED FACTBank-change countDefinition · time · value
RETRIEVED EVIDENCEVerification documentRevision · eligible passage
Context packageThe information selected for this decisionScope · transformations · version · provenance
01Policy evaluationConditions + fact references
02Human decisionReviewer + approval scope
03Governed tool attemptAuthority + operation identity
04Reported outcomeResponse + unresolved state
Inspect which information influenced the decision and how that decision connects to the attempted action.
Data provenance
Where information came from, which revision was used and how it was transformed.
Decision lineage
Which context, fact values, rules and human judgments shaped the decision.
Execution lineage
How authority, waits, attempts and reported effects connect to the same business operation.
Durable state is a distinct responsibility
Remember the execution.
Revalidate the next action.
Persist the position of a Frozion workflow: pending waits, event correlation, approval state, business operation identity, attempts, version references and recovery state.
This state supports the Control Plane as it waits for a person or system and resumes the same execution. When it continues, current authority, required data and policy must still support the next operation.
Historical evidenceWhat did this decision use?
Preserve the original context and fact versions so later inspection does not substitute today's values.
Live continuationWhat may happen now?
Use the persisted workflow position and revalidate the conditions required for the next governed step.
The value of an integrated Data Plane
The value of an integrated Data Plane.
across the application.
The value is in the handoffs: retrieval into context, facts into policy, context into review, and decisions into execution evidence. Your applications use these capabilities together.
| Business requirement | How Frozion connects it | Operational benefit |
|---|---|---|
| Ground an existing agent | Expose a workflow that returns governed records, knowledge and calculated facts. | Enrich the agent's tool call while retaining its current framework and language. |
| Make business rules contextual | Feed versioned Feature Store values into nested policies and the HITL Inbox. | The rule and reviewer can inspect the fact behind the escalation. |
| Retrieve within scope | Apply request authority to keyword and vector candidates before the workflow assembles decision context. | Relevant knowledge remains bounded by the application and caller's permissions. |
| Reuse expensive prior work | Evaluate cache and memory against current scope, sources, freshness and versions. | Avoid repeating eligible work while making invalid reuse explicit. |
| Explain a consequential action | Link source revisions, context, fact values, policy, approvals and tool attempts. | Inspect how the available information shaped the reported outcome. |
| Use activity history in policy | Counts, totals and rates from governed activity feed policy as defined features. | Recognize patterns that a single request cannot reveal. |
Keep your systems of record
Start with one decision.
Connect the data it needs.
Identify the operation, its authoritative sources and the facts that determine what may happen next.
-
01
Define the decision and its scope.
Name the caller, tenant, purpose, business entities and permitted data operations.
-
02
Assemble the records, knowledge and facts.
Set source authority, query and retrieval rules, feature definitions, freshness requirements and evidence references.
-
03
Connect context to governed work.
Use the package in policy, human review and durable execution. Inspect the source-to-outcome path before expanding.
Composable storage. Shared controls.
Use the right storage
for each responsibility.
Operational entities, source files, search indexes, calculated facts, evidence and execution state have different requirements. Frozion connects them through identity, scope, contracts, versions and provenance.
Relational recordsDocument & object storageKeyword & vector indexesFeature valuesCache & memoryEvidence & execution state
Keep the enterprise systems responsible for their own records. Build a governed application around permitted information from those systems.
KNOWLEDGE TO ACTION
One foundation.
One foundation.
Different applications.
Enterprise knowledge
Retrieve permitted evidence and connect answers to governed follow-up work.
Governed SaaS
Build tenant-specific context, policy and evidence into branded applications.
Questions platform teams ask
Before you connect
your data.
Retrieval finds relevant material. A business decision also needs authoritative records, defined calculations, permission checks, freshness and traceable inputs. The AI Data Plane connects those responsibilities so the same context can serve policy, human review, agents and the work that follows.
Yes. Existing agents can invoke governed tools and workflows through MCP, with workflows also callable over HTTPS. A workflow can retrieve the permitted records, documents and calculated facts the task needs while the agent keeps its framework. The same capabilities serve agents and applications built in Frozion.
Those systems can remain authoritative for the records
they own. The Data Plane provides application data roles,
governed access, context assembly and evidence
relationships around permitted information. Different
physical stores can serve different responsibilities.
Retrieval finds eligible candidates using keyword and vector search. The workflow assembles decision context by combining selected candidates with records, calculated facts and permitted memory, using explicit scope, freshness, precedence and sufficiency rules.
No. Defined facts can feed policy, routing, workflow context, App Builder screens, human review and models. They can also summarize recorded Frozion activity so rules can act on what has already happened. Feature materialization and freshness are configured for the use case.
Define the response in the context and workflow
requirements: refresh or recompute a fact, wait for a
source, route an unresolved case to an authorized
reviewer, or stop the operation. Missing or stale
information remains explicit. A human decision does not
automatically grant access the caller or reviewer lacks.
Memory retains prior information that may be useful again.
Durable state records the execution's position, waits,
obligations and recovery state. Historical evidence
preserves what informed a past decision. Resuming work
uses current authorization and required data checks while
keeping the original decision history inspectable.
Start with one consequential decision
Bring your data sources.
See the governed path to action.
We'll map the records, knowledge, calculated facts and
permissions
your application needs—and show how they connect to policy,
review and execution.