A Support Copilot That Does Not Need the Whole Company

Consider a support engineer answering a deployment question. The answer may be in an approved runbook, while a nearby incident note contains credentials or information belonging to another customer. “Connect the assistant to all our documents” is not a sufficiently precise access policy.

The more useful proposal is a bounded support copilot: approved sources for a defined group, a permitted model destination and an answer that retains its supporting references. ECAI's documented private-index and LLM-bridge work provides a basis for evaluating that workflow.

OWASP's sensitive-information guidance emphasises access controls and restrictions on the data sources an LLM application can reach. That supports evaluating the disclosure boundary, not assuming that a reassuring prompt provides one.

The first question is whether the requester may read the corpus. The second is whether its contents may be sent to the selected model service. A user who may read an internal runbook should not automatically gain permission to export it to every model provider.

The private-index implementation guide describes separate readers and writers, corpus-wide permissions, encrypted segments and an allowlisted model-dispatch path. It also makes an important limitation explicit: this is not per-record authorisation. Collections with different readers should not be combined merely to simplify indexing.

For a pilot, create two synthetic collections with intentionally different readers. Give each a permitted destination and a set of questions with known supporting records. Keep the policy small enough that a human can review it without interpreting a large permission matrix.

The documented retrieval worker decrypts scanned segments in memory. It does not search ciphertext without decryption, and the host remains inside the trust boundary. A compromise of that worker is a different threat from a person reading an encrypted disk file.

The same guide describes bounded excerpts and a no-source path that does not call the model. Local model routing constrains the immediate destination; it does not prove that every downstream component, log, crash dump or proxy is private. Remote model use is a deliberate disclosure decision.

These limits are useful purchasing information. The buyer can decide whether the design fits its deployment boundary, rather than discovering after a pilot that “private AI” meant something different to each participant.

Ask a permitted user a question whose answer is present. Inspect the source reference and compare the generated answer with the actual record. Then ask an unrelated user the same question, select a forbidden destination and ask for material that is not in the collection.

The proposed acceptance conditions are no unauthorised disclosure, no dispatch to a forbidden destination and an explicit no-evidence outcome. Also test a permission change after a successful response. A refused request should not be disguised by showing the previous answer as though it were new.

Use synthetic content to examine prompts and logs. Put an instruction-like string inside a source document and ensure the surrounding application treats it as evidence to inspect, not authority to grant new tools or permissions. This is a test requirement, not a claim of complete prompt-injection defence.

Track the proportion of questions with the expected supporting record, the rate of appropriate abstention and the time a support engineer spends checking the answer. Record a wrong but well-cited answer as wrong. A source label is not evidence that every sentence follows from the source.

Compare retrieval against the team's existing search on the same corpus and questions. The documented private path uses term matching; do not claim a semantic-search advantage or a model benchmark victory without a controlled comparison. The advantage to demonstrate first is an inspectable access and disclosure workflow.

The implementation description comes from the dated private-index reference; its latest source and real cryptography/model backends were not revalidated for this article. The guide also records lifecycle limits around rotation, updates, deletion and rebuilding that belong in any deployment evaluation.

Start with the permission and destination model and a deliberately versioned corpus. Sell a useful answer inside a tested boundary, not unrestricted access with privacy language attached afterward.