Custom LLM for Recruitment Agencies That Keeps Candidate Data Onshore

It is safe when the model runs on infrastructure you control and unsafe when it does not, and that distinction is the whole point. A resume is a dense package of personal information: contact details, employment history, sometimes date of birth, visa status or health disclosures. On a private deployment in an Australian region or your own server room, that file is processed inside your existing security boundary, subject to your access controls, and logged. Nothing is retained by a third party and nothing contributes to training an external model. The residual risks are the ones you already manage under APP 11: access control, encryption, retention and staff conduct. On a public consumer tool, you add an offshore recipient, an opaque retention policy and a training-rights question you cannot answer to a client or a regulator.

We are not your lawyers, but the exposure is real and specific rather than theoretical. Sending a candidate's personal information to an offshore AI provider is a cross-border disclosure engaging APP 8, which makes you accountable for that recipient's handling of the data and requires reasonable steps beforehand to ensure APP compliance. APP 1 requires your privacy policy to disclose overseas disclosure and the likely countries. APP 6 limits use to the primary purpose of collection or a related purpose the candidate would reasonably expect, and most candidates do not expect their file to be pasted into a consumer chatbot. If the file contains a police check or medical information, APP 3 consent requirements for sensitive information apply as well. The practical risk is that the gap surfaces during a client security review, a candidate access request, or a breach assessment, when it is too late to fix.

Yes. JobAdder and Bullhorn both expose documented REST APIs, and they are the two platforms we see most often in the Australian agency market. We integrate by reading candidate and job records through the API, indexing them for retrieval, and writing structured output such as screening summaries or submission drafts back onto the record so the work lives in your system of truth rather than a side tool. The same approach applies to Vincere and Recruit Wizard on the agency side, and PageUp, LiveHire, Workday and SuccessFactors for in-house talent teams. Two practical constraints matter: the integration respects your existing ATS user permissions, so a consultant cannot retrieve candidates they could not already see; and if your platform has no usable API or you are on a heavily customised legacy build, integration effort can become the dominant cost of the project. We assess that in the first stage of scoping.

For a single, well-defined use case such as resume screening on a cloud-hosted Australian deployment with one ATS integration, implementation is typically $25,000 to $60,000. A medium deployment covering two or three use cases with retrieval over your full candidate database and an integration lands in the $60,000 to $120,000 range. Infrastructure is separate and ongoing: a small model handling roughly 500,000 tokens a day runs $800 to $2,000 per month on Australian-region cloud, and a medium model at around 2 million tokens a day runs $2,500 to $6,000 per month. Managed operation covering monitoring, updates and support is $1,500 to $4,000 per month. Those are genuine ranges, not entry prices with the real number hidden behind a call. Where your agency falls depends on consultant headcount, database size and how clean your existing data is.

For most agencies, Australian-region cloud is the right answer. It keeps candidate data onshore in AWS Sydney or Melbourne, Azure Australia East, or Google australia-southeast1, satisfies the data residency question that clients actually ask in security reviews, requires no capital outlay, and scales with your volume. On-premises becomes the right answer in a narrower set of cases: when you recruit for defence, national security or cleared roles and the client contract mandates physical control of the infrastructure; when you already run your own data centre and the marginal cost is low; or when your query volume is high enough that fixed hardware beats ongoing cloud spend, which typically takes around twenty-four months to cross over. A hybrid split is common and often the most sensible outcome: sensitive candidate files processed on-premises, everything else on Australian-region cloud.

This is the risk most agencies underestimate, and it is separate from privacy. A screening model that learns from your historical placements will learn your historical biases, and it can reconstruct protected attributes from proxies: a graduation year implies age, a career gap can imply parental leave or disability, a postcode can correlate with race. Under the Fair Work Act and the federal age, sex, disability and racial discrimination Acts, plus state law such as Victoria's Equal Opportunity Act 2010, the exposure is yours regardless of what the software did. We exclude protected attributes and known proxies from the ranking inputs, require a written rationale for every ranking so decisions are reviewable, keep a human accountable for any adverse decision, and test for disparate impact across the shortlist. From December 2026, privacy policies must also disclose where personal information feeds computer programs making decisions that significantly affect individuals, so your process will need to be describable in plain English anyway.