Inside the controlled environment
Prepare a follow-up for Priya Nordin, priya.nordin@example.invalid, account AC-4827, about the Leuven delivery.
Roadmap direction
Give agents the context. Keep identity closer.
Agents need enough business context to do useful work. They rarely need every real name, address, account reference, or personal detail attached to it. Our edge-privacy direction places a transformation boundary before external model processors.
The customer problem
Sensitive work should not force a choice between useful context and responsible disclosure.
An agent may need to connect an email to a customer record, follow a case across documents, or prepare the next action. Sending every source value to an external processor gives the model more identity than the task may require.
The planned boundary will replace eligible identifying details inside the PitchAI-controlled environment while preserving the relationships the model needs. The processor receives a minimised working version. Authorised results return through the same local boundary for restoration.
Planned data flow
Transform before the request leaves. Restore after the result returns.
The mapping stays apart from the model-provider payload, so the agent can keep useful relationships without sending the restoration key.
01
Assemble approved context
The agent gathers the prompt, conversation, and tool data needed for the agreed process.
02
Find identifying details
Policy and detection layers identify eligible names, contact details, record references, locations, and contextual identifiers.
03
Substitute locally
Real values become consistent synthetic references. The protected mapping and restoration authority stay local.
04
Process minimised context
OpenAI, Anthropic, or another approved processor receives the transformed context needed for the task.
05
Return transformed output
The response comes back with synthetic references still in place and no restoration map attached.
06
Restore for authorised use
The local boundary restores recognised references and returns useful output to the authorised workflow.
Synthetic example
The agent can follow the case without the processor receiving the source identity.
External processor view
Prepare a follow-up for Elena Vos, elena.vos@example.invalid, account AC-7314, about the Mechelen delivery.
Authorised local result
Follow-up prepared for Priya Nordin about account AC-4827 and the Leuven delivery.
Every person and organisation in this example is fictional.
Intended benefits
A smaller disclosure surface without flattening the work.
01
Minimise what leaves
Transform eligible identifiers before the provider request while retaining the context the agent needs.
02
Preserve useful relationships
Keep repeated people, records, and references consistent across the working context.
03
Keep restoration authority local
Separate the mapping from the provider payload and restore recognised values only inside the authorised environment.
The whole route
A privacy boundary is wider than the prompt box.
The roadmap covers each place sensitive context can travel, from the working request to returned results, tools, files, operational traces, and memory.
01
Instructions and conversation
Dynamic prompts, conversation history, summaries, and memory passed into a request.
02
Tools and connected systems
Tool arguments, returned records, connector payloads, and structured fields.
03
Provider responses
Streamed text, structured output, and model-generated tool calls on the way back.
04
Files and media
Document text, images, handwriting, faces, QR codes, hidden objects, and file metadata.
05
Operational traces
Logs, caches, retries, telemetry, error paths, backups, and evaluation records.
06
Retrieval and memory
Search indexes, embeddings, vector stores, long-term memory, and linked source material.
What the terms mean
Pseudonymisation reduces exposure. Identifiability still depends on context.
When PitchAI retains a protected mapping that can restore a value, the transformed data remains pseudonymised personal data within that trust boundary. Replacing a name or address alone does not make a record anonymous. Narrative details, rare combinations, and other available information can still single someone out or link records.
For a recipient, identifiability depends on the complete payload and the means that recipient or connected parties could realistically use to identify someone. That assessment is specific to the processor, purpose, access, and surrounding controls.
Edge privacy
Build useful agents with a more deliberate data boundary.
This capability is an established Agent Engine roadmap direction, now in design and validation across the complete agent route.
Discuss your first process