← All insights

Insight / AI and infrastructure

Published

Confidential data handling with local AI

How local AI helps handle confidential correspondence: originals stay local while reviewed alternate copies support hosted search across business units.

Envelopes and a pencil arranged on a desk
Photo: Cht Gsml / Unsplash

A client needed to locate and protect confidential information across email and document histories from several business units without making the original correspondence available to an online search provider. The work combined two recurring tasks: preparing records for a searchable index and finding the trail from discussion to decision to supporting document. Local AI helped prepare alternate copies for review; the hosted provider only needed the approved versions, not the original names and other identifying details.

This anonymised case study describes the project’s data boundary. The client’s identity and original correspondence are not reproduced here. The time and labour figures below are illustrative assumptions, not measured client results.

The design separated the correspondence into three parts:

  1. Original emails and documents stay local. An authorised person can still inspect them when a search result needs its full context.
  2. A local process creates linked alternate copies. A local model identifies candidate sensitive details, such as names and locations, and replaces them in the alternates. The link between each source and its alternate, including any replacement map, stays under local control.
  3. Only the approved folder of alternates is exposed to the hosted provider. This is the material it can scan and index. The folder gives the team a defined place to review what is shared before it crosses the boundary.

Stable replacement tokens can preserve a person’s role across messages without sending their name. That helps a search follow the history of a conversation. It does not make the alternate copy safe by itself: a distinctive sequence of events, dates or roles might still reveal who is involved. The OAIC’s de-identification guidance says removing direct identifiers alone may be insufficient and that re-identification risk depends on the release context.

Review before sharing

Every alternate needs human review before it enters the hosted folder, with extra scrutiny when a new kind of record is introduced. Reviewers should check subjects, headers, quoted replies, attachments, filenames, document metadata and text recovered from scans, as well as the main body. They also need to catch project-specific phrases that reveal a client or commercial arrangement without naming either directly.

If an alternate cannot be sanitised with enough confidence, keep it out of the hosted index and search it locally or by hand. The provider boundary needs a separate check: which folder it can read, what copies, snippets, embeddings or logs it retains, who can search them, and when those materials are removed. Restricting the source folder does not control every later copy.

Illustrative business case / One employee

The work behind a confidential search

This is a worked example built entirely from assumptions. It is not a measured client result or a forecast. The manual comparison prepares the same protected copies to the same review standard.

Assumed workload

New records to locate and protect
160 / active week
Cross-unit correspondence searches
40 / active week
Working weeks
46 / year
Loaded labour rate for the model
AUD $60 / hour

Assumed time per task

Prepare one protected copy
8 min manual 3 min with local AI and human review
Search and verify one source
12 min manual 5 min with the reviewed index
Index checks and exception oversight
1 hour / active week

Weekly time account

Prepare 160 protected copies 21h 20m manual / 8h assisted
13h 20m released
Complete 40 searches 8h manual / 3h 20m indexed
4h 40m released
Additional index oversight
−1h
Total 29h 20m manual / 12h 20m assisted
17h / week

Modelled annual labour-capacity value for one employee handling this workload

AUD $46,920

17 hours per active week × 46 weeks = 782 hours per year. At the assumed AUD $60 loaded hourly rate, that is AUD $46,920 of potential staff capacity.

Illustration, not an achieved saving. This is not a payroll reduction or net return. It excludes archive preparation, setup, hosting, model operation, storage, security, governance, training and unusual correction costs. More review, lower search quality or less work would reduce the modelled value.

Each business unit needs its own measured volume and review burden before this calculation is applied to another employee. Staff must still verify an original before acting on an indexed result; time saved on a wrong answer is no benefit.

What to measure next

Use comparable samples from each business unit to record manual and locally assisted preparation time including human review, plus manual and indexed search time including source verification. Count new records, searches, rejected alternates, corrections and cases where a result misses the needed context. Record each employee’s actual workload, active weeks, loaded cost and the setup and running costs separately. Those measurements can replace the assumptions above and show whether the index frees useful capacity for this client.

The lasting lesson from this project is the data flow: keep the original correspondence and its source mapping local, and let the hosted provider work only with the approved alternate set. That narrows what is shared, while leaving re-identification risk and human review as ongoing responsibilities. Our data sovereignty guide covers the wider choice between local and hosted processing.

External project website

Open this project site?

is hosted at . It will open in a new tab.