Australian cloud for business software — built and run in Sydney since 2002

NetLeverage
NetLeverage

Using realistic training data without exposing live customer records

Technical article
Hosting

About this article: This article is reviewed against the cited sources. This is source research, not lab validation. Check the stated versions and your own requirements before making changes.

Direct answer: Train customers on hosted training tenants fed by synthetic data, officially licensed public datasets, or de-identified extracts that pass a documented risk assessment—not by giving trainers a copy of yesterday’s production database “because it feels real.”

Why it matters

Customer enablement fails with toy data that does not match field lengths, tax rules or workflow branches. It also fails when a trainer’s laptop holds identifiable health, payroll or child-related records copied from a live tenant. If your business is an APP entity covered by the Privacy Act 1988—for example turnover thresholds, health-service providers and other categories in sections 6D and 6E of the Act—the Australian Privacy Principles apply to personal information you handle; training misuse is an avoidable reputational and compliance problem.

This article addresses training-data process and environment design, not a data-residency questionnaire or where servers sit geographically. Residency and training ethics intersect—e.g. offshore trainers viewing Australian personal information needs contractual and technical controls—but the immediate win is not putting live records in the classroom.

How it works

Classify before you copy

Label fields: public, internal, sensitive personal information, sensitive (health, etc.). Training may use public and some synthetic sensitive categories without real identities. Anything that can identify an individual needs a de-identification path or must stay out of training entirely.

The OAIC’s guidelines on APP 11 — Security of personal information quote the Privacy Act test that personal information is de-identified if it is no longer about an identifiable individual or an individual who is reasonably identifiable, and state that, whatever de-identification technique is chosen, the risk of re-identification must be actively assessed and managed. In smaller regional datasets, that risk rises—an editorial planning point, not a statistical proof.

Three approved sources

  1. Generated data matching schema constraints (valid ABN formats, plausible dates) via scripts or tools.
  2. Customer-approved scrubbed subset produced by an automated pipeline with sign-off from the customer’s privacy officer and your organisation’s privacy contact.
  3. Vendor-supplied sample packs you maintain and version like software—updated when the schema changes.

Environment controls

Training tenants sit in the same non-production class as sandboxes: separate URLs, no production webhooks, mail routed to test inboxes, shortened retention, no backups shipped to long-term archives unless policy allows. Trainers work in the training tenant with seeded data; avoid standing read access to customer production except rare, logged break-glass support windows agreed with the customer.

Curriculum linkage

Each course module lists required data shape (“10 invoices with credit notes”) and the seed pack version. When production adds a mandatory field, update the seed before marketing the course.

Attendance and recording

Recorded webinars blur screens that show seeded data but still capture trainee faces and voices—classify recordings as personal information with their own retention rule.

What to choose and why

Need Prefer Avoid
Credible workflows Golden seed packs per vertical Prod snapshot to USB
Custom customer process Scrubbed subset contract Ad-hoc Excel exports from CS
Certification exam Fixed dataset with known answers Live tenant “practice mode”

Run a periodic re-identification risk review when you add new columns (geolocation, device IDs) or when you merge training data from multiple customers for a generic course.

Risks and limits

  • Re-identification via rare job titles, transaction timestamps or free-text notes is subtle; involve statistical or privacy expertise for marginal cases.
  • Trainers pasting customer screenshots into slides leaks data outside your controlled tenant.
  • Machine-learning features trained on production exports can memorise identifiers; training ML belongs in a separate policy.
  • This is not legal advice; APP 6 (use and disclosure) and notifiable breach obligations need qualified interpretation for your facts.
  • De-identification tooling was not lab-tested in this article.

Sources

Related reading

Learning notes

  1. Realistic training needs engineered data, not production copies.
  2. Smaller populations can raise re-identification risk—plan OAIC-style risk review before copying data.
  3. Training tenants deserve the same integration fencing as sales sandboxes.
  4. Seed packs are versioned artefacts like installer binaries.
  5. Legal and privacy review belongs in the training-data standard, not after an incident.