Sarah and I have been foster parents for over two decades. We adopted four of our children after many years of temporary care.
I mention that up front because everything that follows comes out of those experiences. We have participated in this system, lived with its decisions, and watched the people working within it try to meet demands that would exhaust almost anyone.
We have a perspective on what technology should do for child welfare. And it starts with the people doing the work.
The Real Constraint
We have sat across from a lot of social workers.
They arrive at your kitchen table carrying a caseload, a folder for your family, and a phone full of other families who need something just as urgently.
They are expected to make careful judgments about some of the most consequential circumstances in a person’s life. They must document what happened, meet deadlines, coordinate services, and account for decisions that can change a child’s future.
All while the next family is waiting.
The good social workers are extraordinary, and we have known many of them. Almost all are tired in a particular way that comes from caring about more people than any one person can reasonably hold.
There is a human limit to how much attention and emotional energy a person can give. And yet the work continues beyond that limit. This is the constraint technology should be helping us address.
Two Problems That Look Alike from a Distance
When people ask what artificial intelligence should do for social work, we think the answer starts by distinguishing two very different problems.
The first is judgment: understanding what is happening in a family, evaluating its circumstances, and deciding what should happen next.
The second is administrative overhead: documenting conversations, entering the same information into multiple systems, locating records, preparing reports, and assembling case histories from information that already exists.
Both consume a caseworker’s time, but they are not interchangeable responsibilities.
The first demands professional expertise, context, empathy, and accountability. The second contains substantial work that technology may be able to simplify.
Too much of the conversation about AI in child welfare has focused on predicting risk and informing decisions about families.
We believe the more immediate opportunity is to give social workers back the time they need to exercise their own judgment.
What the Data Actually Learned
Predictive analytics in child welfare has a complicated history.
Allegheny County, Pennsylvania, introduced its Family Screening Tool in 2016 to help inform decisions about which reports of suspected child abuse or neglect should be investigated. Its early models used outcomes including future referrals and out-of-home placements to estimate risk.
Those choices raise an important question: What is the model actually predicting?
A future referral reflects another interaction with the child welfare system. An out-of-home placement reflects an intervention made through that system.
Neither is a direct, independent measurement of everything happening inside a family.
Historical records can contain patterns shaped by poverty, unequal access to resources, reporting practices, and previous agency decisions.
When those records become training data, organizations need to understand which patterns a model may reproduce. The concern is particularly important because poverty and child welfare involvement are closely connected.
Research summarized by Washington University in St. Louis estimates that approximately 85 percent of families investigated for neglect have incomes below 200 percent of the federal poverty line.
A University of Washington study also found that a 10 percent increase in Earned Income Tax Credit benefits was associated with a 9 percent decrease in annual reports of child neglect.
These findings do not mean every neglect report is explained by poverty. They do demonstrate why economic circumstances, family needs, and system involvement must be interpreted carefully.
Predictive tools have also produced mixed findings. Some jurisdictions have discontinued or reconsidered their approaches amid concerns about reliability, equity, and appropriate use.
Allegheny County continues to use its screening tool. Research has raised concerns about disparities and the limitations of administrative data, while a 2026 study found that the tool’s implementation was associated with reductions in certain racial disparities in screening and removal decisions.
The evidence deserves careful consideration.
Our position is that no prediction should become a substitute for understanding a family, documenting the basis for a decision, and maintaining meaningful human accountability.
Evaluations, Not Scores
Long before we began building AI technology, we had a rule at home: Evaluations, not scores.
An evaluation can hold context, uncertainty, history, and disagreement. It can explain what a family is experiencing, what information is available, what remains unknown, and why a professional reached a particular conclusion.
A score compresses information into a number. That can be useful for certain analytical tasks. But when the subject is a child or a family, the limitations of that compression matter enormously.
A parent sitting across from a caseworker deserves to understand why a decision was made.
What information was considered? What concerns were identified? Who reviewed the evidence? What circumstances might change the decision?
A numerical prediction cannot answer those questions on its own. And a caseworker should never be placed in the position of defending a consequential recommendation they cannot explain or independently evaluate.
The standard should be meaningful human judgment supported by information that can be examined, questioned, and traced to its source.
Where AI Should Actually Help
There is another place to direct AI investment, and it is considerably less glamorous.
Take the paperwork, the repeated data entry, and the hours spent searching for information scattered across disconnected systems. Take the court report assembled late at night from records that already exist.
Help the social worker produce a draft, find the relevant documents, organize the chronology, and identify information that still needs review. Give the professional more time to sit across from a family and pay attention. This is where we see a meaningful opportunity for responsible AI in child welfare.
Consider a few practical applications:
- Document preparation: Assemble draft reports from authorized case records, with source references and professional review.
- Record retrieval: Help workers locate relevant information across approved document repositories without manually searching every file.
- Case chronology: Organize documented events into a timeline that a caseworker can verify and correct.
- Administrative summaries: Prepare drafts of meeting notes, service updates, and routine documentation for human approval.
- Information consistency: Identify missing fields or discrepancies across records so employees can resolve them.
These applications still require careful design.
Case records contain deeply sensitive information. AI-generated summaries can omit context, introduce errors, or misrepresent what a source actually says.
Access controls, data protection, source traceability, and human verification are essential.
The goal is to reduce administrative demands while preserving the professional’s authority over the work.
Human Oversight Has to Mean Something
There is a difference between having a person involved and giving that person meaningful control.
If a system produces an answer and the worker is expected to approve it quickly, human oversight can become little more than a formality.
Real oversight means the professional has the information, time, and authority to question the output. They should be able to inspect the source material, correct mistakes, reject recommendations, and document their reasoning. And they should understand where the information came from.
These principles shape how we think about building secure AI systems at Pensato and helping organizations adopt them responsibly through Varium.
We focus on protecting sensitive information, preserving traceability to source documents, and establishing governance that keeps people accountable for consequential decisions.
For organizations managing highly sensitive records, local processing and controlled deployment environments may also be important considerations.
Technology should fit the responsibilities of the people using it and the obligations of the organization deploying it.
The person responsible for the decision must remain able to understand, challenge, and defend it.
The Fifth Voice
In barbershop singing, when four voices align with remarkable precision, listeners can sometimes hear what sounds like a fifth voice. No fifth singer has joined the group. The sound emerges from the interaction of the voices already there.
A foster home has something of that quality.
People show up with different experiences, histories, and hopes. Over time, through attention and commitment, something develops that none of them could have created alone. It becomes a family.
That is worth protecting in social work: relationships, and trust. The willingness to keep showing up when circumstances are difficult and answers aren’t simple.
When compassion becomes a template, it stops being compassion.
Good technology can carry more of the administrative burden. It can organize records, prepare drafts, and help professionals find the information they need. But it cannot replace the human relationship at the center of the work.
Tell Us What You See
If you work in child welfare or social services, we’d like to hear from you.
How much of your week goes toward working directly with people, and how much goes toward documenting and proving the work you’ve already done?
Where could better tools make a meaningful difference without compromising privacy, professional judgment, or accountability?
At Varium and Pensato, we’re interested in what it takes to build AI around the realities of human services: secure information handling, traceable documentation, practical governance, and technology that supports the people doing the work.
Let’s talk about how to take paperwork off a caseworker’s plate while keeping people, not predictions, at the center of decisions.
Comments are closed