Several Australian recruitment firms have asked us the same fair question: if we download an open-weight model and run it on our own servers, does our candidate data problem go away? It can. For most agencies under 200 staff, though, it is not the cheapest way to get there.
Recruitment and HR agencies hold some of the most sensitive personal information any Australian business touches: resumes, salary history, references, visa status and, more and more often, recorded video interviews. The Privacy Act obligations attached to that data do not relax because the model processing it is free to download.
Does self-hosting an open-source model make recruitment AI Privacy Act compliant?
Not on its own. Self-hosting an open-weight model such as Qwen or GLM keeps candidate data off third-party APIs, which removes one disclosure question. It does nothing about collection notices, consent for reference checks, access logging, retention, or how a screening decision is explained to a candidate. Compliance comes from the documented data flow and controls around the model, and those have to be built whichever model you run.
That is the trap we see most often. An agency hears that open source means the data never leaves the building and treats the model choice as the compliance project. The model choice is maybe a fifth of it. The rest is governance work that someone has to design, build and keep running.
Which recruitment tasks suit an open-weight model
Some jobs fit a self-hosted model well. Others carry risk that no licence or hosting decision resolves. This is how we sort them for a typical agency:
| Task | Self-hosted fit | Where the risk sits |
|---|---|---|
| Bulk resume parsing into structured fields | Strong: an 8B to 30B model handles it well | Low if access is logged and retention is set |
| Interview transcript summaries for internal notes | Strong: transcript never leaves your infrastructure | Low to moderate, depending on who can read the notes |
| Reference checks and background screening | Model choice is not the issue | Third-party personal information and consent |
| Video or voice analysis for culture-fit scoring | Model choice is not the issue | Contested ground under Australian anti-discrimination law |
| Shortlist ranking that affects who gets a call | Possible, with human review | Automated decision transparency and bias testing |
The pattern is clear enough. Where the work is extraction and summarisation, self-hosting genuinely helps. Where the work involves people who are not your candidate, or judgements about a candidate's personality, the hard questions are legal ones. We covered the screening side in detail in our guide to recruitment screening under the Privacy Act.
What each path costs a mid-sized agency
For a mid-sized agency, a self-hosted resume-screening pipeline on an open-weight model, with proper access logging and a documented data flow ready for a Privacy Act audit, typically costs $12,000 to $25,000 to build. Running and maintaining it adds roughly $1,000 to $2,500 a month.
A Claude-based pipeline under enterprise terms usually cuts that build cost by about a third. The reason is not that Claude is cheaper per token. It is that the data handling terms are already documented and audited, so you are not also building the infrastructure governance layer yourself. On a $20,000 self-hosted build, that difference is in the order of $6,000 to $7,000 before the monthly running cost is counted.
Self-hosted wins when candidate volume is high, the agency has internal engineering capacity, or a client contract forbids any external processing.
Claude wins when the agency has no one to patch, monitor and log a model server, and needs a documented position quickly.
Either path still needs a written data flow, a retention rule and a named person accountable for it.
For the three-year view rather than the build quote, our total cost of ownership breakdown for open-source AI walks through the line items agencies usually leave out, such as security patching and the staff time spent keeping the server healthy.
Mistakes we see in agency pilots
Most failed pilots in this sector fail in the same three places. The first is loading a folder of historical CVs into a test environment without checking whether the original collection notice covered that use. The second is letting consultants paste reference check notes into whatever tool is open, which spreads third-party information across systems nobody is tracking. The third is switching on a scoring feature because the vendor offered it, before anyone has asked whether the agency could explain a low score to the candidate.
None of these are model problems. All three happen as readily with a self-hosted model as with a hosted one. A short data flow map, done before the pilot rather than after, avoids all of them.
How we approach it
We do not default every recruitment client to open source or to Claude. We map the actual data flow first, agency by agency, because the right answer depends on candidate volume and on how much internal engineering capacity exists to maintain a self-hosted system. As at September 2026 that is still the right order of operations: data flow first, model second. A common outcome is Claude for drafting and summarising, with a self-hosted parser added only where CV intake volume justifies running a server. If you are still comparing products, our round-up of AI tools for recruitment agencies is a useful companion, and our services page explains how a data flow review runs.
If you run a recruitment or HR agency and want a straight answer on where your risk actually sits, book a session with us.
