Hello, this is CUBIG, the company behind Syntitan, the AI-Ready Data Platform for enterprise AI. 💎
A procurement team can have every vendor contract stored, classified, and access-controlled, yet still be unable to use those contracts in an approved AI workflow.
The team wants AI to extract renewal dates, notice periods, obligations, and related clauses for human review. But the same contracts contain names, commercial terms, account details, and other values that cannot travel unchanged through every model path. Removing those values may reduce exposure while also breaking the relationships the review depends on.
This is an enterprise data usability problem. The question is whether data can move through the approved AI path, preserve the meaning required by the task, and produce a result that satisfies an explicit acceptance rule and risk boundary.
Cisco's 2026 Data and Privacy Benchmark Study shows how widely the broader tension is felt. Ninety percent of surveyed organizations say AI has expanded the scope of their privacy programs, while 65% struggle to access relevant, high-quality data efficiently. The survey—covering more than 5,200 IT, technology, and security professionals with data privacy responsibilities across 12 markets—does not establish that privacy programs caused the access problem. It shows that stronger governance and persistent data-access friction now coexist.
Storage, ownership, and access approval remain necessary. The next test is operational: what information must survive transformation for a particular AI task to work?
Usability Is a Data-Path Question
The procurement workflow gives three groups distinct decisions to make. Governance approves the path. The workflow owner defines the required fields and relationship checks. The data and AI team compares the transformed workflow with a controlled, human-verified reference set and records both utility and remaining disclosure risk.
The questions become concrete:
- Which values must not travel unchanged?
- Which relationships and sequences must the extraction preserve?
- Where does transformation occur?
- What representation crosses each system boundary?
- Which result will the workflow owner accept, reject, or escalate for human review?
Governance answers who may do what. A usability test shows whether that approved path can still produce the required work.
Every Method Preserves Something and Changes Something
There is no universal operation called “anonymize the data.” Removal, de-identification, synthesis, and substitution change different properties. Their value depends on the intended use and on what the team measures afterward.
NIST's Privacy-Enhancing Technologies Testbed evaluates de-identification and synthetic-data methods across privacy, fidelity, and utility. NIST notes that de-identification mechanisms can introduce artifacts and bias, while synthetic-data methods create new individuals intended to reproduce selected distributions. The UK's Information Commissioner's Office similarly advises teams to assess whether synthetic data is an accurate proxy for the intended purpose and warns that source-data bias can carry through.
The table is not a ranking. NIST and ICO provide independent guidance for evaluating de-identification and synthetic data. The context-preserving substitution row reflects current CUBIG product framing and still requires validation on the intended workflow.
For the procurement example, redaction may be sufficient for a field that never affects contract review. It is not sufficient if it removes a date or breaks a reference between a party and an obligation. Synthetic data may suit a separate testing need when the generated properties are accurate enough for that use. Context-preserving substitutes may suit the live document path when consistent references must remain understandable without sending original values unchanged.
The method name does not settle the decision. The recorded test does.
The Missing Layer Between Storage and Execution
In a common enterprise architecture, one system stores the contracts, governance controls access, and an AI service performs extraction. The difficult handoff sits between those functions.
The procurement workflow needs a transformation step that is specific to its output. That step must preserve renewal dates, notice periods, clause order, and counterparty relationships while enforcing the approved data path. The team must then test the transformed contracts before the process becomes routine.
This is why a technically available file can remain operationally unusable. A catalog can locate it. A policy can permit a user to access it. A model can accept it as input. None of those events proves that the transformed representation still supports the decision the procurement reviewer needs to make.
Usability has three parts:
- Path: The original values, transformed working data, model path, and any reconstruction step have explicit boundaries.
- Task: The workflow owner defines the structure, context, and signals the AI job cannot lose.
- Evidence: The team tests the transformed workflow for utility and remaining risk instead of accepting it by assumption.
Controlled contracts
Original values stay governed.
Task boundary
Preserve clause order and relationships.
AI review output
Required fields remain linked to their source.
Evidence gate
Compare with the verified reference and record what happens next.
- Accept
- Revise
- Human review
A Five-Step Enterprise Data Usability Test
The test begins with one blocked workflow, not a platform inventory.
1. Name the output that matters
The procurement team is not asking AI to “understand our contracts.” It wants a structured review record containing renewal dates, notice periods, obligations, related clauses, and the source location for each extracted item.
The workflow owner defines acceptance before testing. Required fields must be present, relationships must point to the correct counterparty and clause, and uncertain cases must route to human review. The rule does not need an invented universal score. It needs a result that this team can consistently accept, reject, or escalate.
2. Identify the information the task cannot lose
List the values, relationships, order, distributions, and rare cases that influence the output. Separate original identifiers from task-relevant meaning.
In the contract workflow, the model may not need a party's original name. It does need to recognize that the same party appears in a pricing clause, a renewal clause, and a termination clause. It also needs exact dates, the order of related provisions, and the connection between an obligation and the entity responsible for it.
This inventory becomes the utility contract for the transformation.
3. Draw the approved data path
Mark where original values begin, where transformation occurs, what travels to the AI system, what returns, and where any reconstruction happens. Include prompts, connectors, logs, caches, and human handoffs that touch the working data.
For procurement, the path owner should be able to point to the representation that leaves the controlled source, the representation the model receives, and the record the reviewer sees. If the team cannot name what crosses a boundary, it cannot evaluate that boundary or reproduce the workflow later.
4. Choose a method for the task
Select removal, de-identification, synthetic data, context-preserving substitution, a controlled execution architecture, or a combination because it fits the job.
Document what the method intentionally changes. If it suppresses a field, state why the contract review does not need that field. If it generates new records for testing, state which distributions and relationships must be accurate enough. If it substitutes terms, state how references to the same counterparty remain consistent across clauses.
This record keeps the transformation tied to the acceptance rule instead of to a generic promise of safety.
5. Test utility and remaining risk together
Run the transformed contracts through the intended extraction workflow. Compare the output with the controlled, human-verified reference set. Check the fields and relationships defined in step two, then record the privacy or disclosure risks that remain.
NIST's framework is useful because privacy, fidelity, and utility are separate measurements. A method can perform well on one dimension and poorly on another. The procurement team therefore records a bounded decision: fit for this workflow under these conditions, rejected for a specific failure, or escalated because the evidence is incomplete.
Where CUBIG Fits
Finish the method decision before choosing a product path. CUBIG positions Syntitan as the AI-ready data platform spanning diagnosis, refinement, release, proof, and verification between existing data infrastructure and AI execution.
Two capabilities address different data-path needs.
LLM Capsule is the context-preserving data layer for AI workflows. In the current CUBIG framing, original values and the protected mapping remain inside the customer-controlled environment. Context-preserving substitutes travel through the approved model path, and reconstruction happens through the mapping inside that environment. For the contract workflow, the relevant test is whether those substitutes retain consistent references and the clause relationships the extraction requires.
DTS addresses transformed or synthetic AI-ready data. Its role differs from substituting terms inside a live document workflow. A team may need context-preserving execution, a transformed dataset, synthetic data for a separate use, or another controlled approach. The five-step test supplies the decision criteria.
These descriptions are product positioning, not a promise that every dataset, model, or regulatory setting will produce the same result. Each team still needs evidence from its own workflow.
Usability Completes the Six-Axis Test
Traceability asks whether a team can follow the path behind a result. Reproducibility asks whether another operator can restore the recorded conditions and run the workflow again.
These axes depend on one another rather than forming a universal sequence. A traceable, reproducible data state cannot support the procurement task if the approved transformation removed an essential clause relationship. Useful data without traceability or reproducibility leaves the team unable to explain or repeat the result.
Usability, Integrity, Context, Consistency, Reproducibility, and Traceability describe one operating state from different angles. W09 closes the series with a practical decision: can this data participate in this work under an approved path and a documented acceptance rule?
Choose one high-value workflow that is currently blocked. Name the output, identify what the transformation cannot lose, map the path, and record the result against a controlled reference.
Run a sample proof on your own data. Start with Syntitan.
