In family violence work, the most dangerous stretch of time is often the gap between asking for help and receiving it. Relationships Australia South Australia (RASA) has decided to fill part of that gap with AI agents, and as a social worker who began my career in family violence and sexual assault services, I have some questions about how it works.
What RASA is doing
RASA's former CEO, Claire Ralfs, recently described the approach at Salesforce's Dreamforce conference. The organisation sees more than 21,000 clients a year, and people typically wait five to six weeks for an appointment. When someone books, they are sent the Family DOORS risk screening survey. An AI agent then analyses their answers against 15 years of DOORS data, and if it flags elevated risk, a trained staff member calls them within days, sometimes hours, to check they are safe until their appointment.
A second agent, which RASA calls Agent Atlas, produces a client summary so the practitioner can start the first session from where the person is now. RASA also uses the DOORS agent to help practitioners document sessions and check that every issue the client raised was covered.
RASA's own screening data shows why this matters. According to Ralfs, at the first appointment one in two clients is dealing with family violence, one in four is struggling with their own unsafe behaviour, and one in ten needs a suicide prevention plan.
What RASA is getting right
A six-week wait is a genuine risk, and RASA is trying to stop people "falling through the cracks" while they wait. Calling people quickly when a screen raises concern is good practice.
DOORS is also a serious, well-researched tool. Ralfs co-developed it with Professor Jennifer McIntosh, it has been validated with thousands of separated parents, and RASA has used it across its services for more than a decade. My concern is with the layer of AI now being added on top.
What happens to the risks the AI misses?
From what has been described, a person steps in when the agent flags someone. Nothing has been said about what happens when it doesn't.
In risk screening, a missed risk is the error that matters most. A false alarm means an unnecessary phone call, while a false all-clear can leave someone living with violence, or thinking about ending their life, waiting five weeks with nobody checking in. The question I most want answered is whether a trained practitioner reviews every completed screen, or only the ones the AI sends their way.
There is also a difference between a validated screening tool and an AI agent that interprets it. DOORS has been tested and retested. An AI layer analysing DOORS responses is a new system, and its accuracy needs its own testing. Fifteen years of good data does not automatically make the agent reading it reliable.
When a summary gets it wrong
The client summaries worry me too. Generative AI is good at producing text that sounds confident and complete, including when it is wrong.
We have already seen what this can look like in Australian social services. In 2024, the Office of the Victorian Information Commissioner reported on a child protection worker who had used the free version of ChatGPT to help draft a report for the Children's Court. The report contained inaccurate information that downplayed the risks to the child. One worker using a public tool is very different from an organisation deploying a configured system, but the failure is one I would worry about in any setting: a summary that quietly softens risk.
If practitioners are starting sessions from an AI summary, what is the process for checking it against what the client actually wrote? When people are busy and a summary reads well, it is easy to trust it. The Robodebt Royal Commission showed how much harm an automated system in human services can do when people stop checking what it produces.
What I would want to know
These questions may well have good answers. If they do, I would like RASA to share them publicly, because other services are likely to follow its lead. In particular, I would want to know:
whether every screen is reviewed by a person, flagged or not
how the agent's results have been tested against standard DOORS scoring, and how often it misses elevated risk
how practitioners check AI summaries and session notes for accuracy before relying on them
whether clients are told that AI is reading their disclosures, and whether they can choose a human-only pathway
where client data is held, and who can access it
The real problem is the wait
At PTC, I have chosen to keep AI out of clinical work entirely. I recognise that is an easier call for a small practice than for an organisation supporting 21,000 people a year, and I don't think every service has to make the same choice.
I do think we need to be honest about what is driving this. A five-to-six-week wait for relationship and family violence support is a funding problem, and AI agents should not become a quiet substitute for the staff these services need. If AI is going to help carry that load, people deserve to know that a qualified human is still checking its work, especially for the clients it decides are fine.
Sarah Newbold is a social worker and the founder of Progressive Therapeutic Counselling, an online practice in Australia. She writes and comments on AI ethics and responsible AI governance.
