AI should flag safeguarding risk and stop — not decide. This shows how to scan case notes in Sopact Sense for harm, crisis, and disclosure, quote the evidence, and route every flag to a human.
In short: Safeguarding is the one place an AI must never decide. In Sopact Sense the scan reads each case note for harm, crisis, and disclosure, quotes the exact line, rates urgency, and hard-sets a recommendation for human review — it surfaces risk and then stops, every time.
Point the Assistant at a clean caseload with persistent contact IDs so a flagged note stays tied to the right person and their history. Load your Decision Brief first — the decision, the audience, the risk indicators that matter, and your evidence standard — so the scan looks for the right signals and routes them the right way.
Scan the case note for risk indicators ([INDICATORS]: harm, crisis, disclosure). Quote, rate urgency, set recommend_human_review=true. No verdict. Grade green/amber/red.
The prompt carries five elements. The dataset is the loaded caseload. Scan for risk looks for harm, crisis, and disclosure indicators. Quote plus urgency pulls the exact line and rates how pressing it is. Recommend human review sets the flag so a person always decides. Grade G/A/R classifies each note so the riskiest rise to the top.
Run on the Caseload dataset (DEMO 08) already loaded in Sopact Sense.
GRADE: green | 1 | Flagged + review; amber | 1 | Disclosure; red | 1 | Children present
Sense returns a graded note with the quoted evidence. A note that correctly flags risk and recommends review comes back green. A disclosure that was quoted but never given an urgency level comes back amber — the signal is there but the routing is incomplete. A note mentioning children with no human review routed comes back red: the highest-risk gap.
Take the lowest-graded element above and fix it using only what the program could realistically measure. Show the before → after grade and the single indicator/edit that moves it to green.
For the children-present red, the fix is a hard rule: set recommend_human_review=true automatically whenever a minor appears in a risk note. That single edit guarantees the most sensitive cases always reach a person.
Create a 'missing & incomplete' report from this analysis in Sopact branding [or paste your website URL / brand guideline to apply your own]. List every element graded amber or red, what is missing, and the one input that fixes each. Lead with the decision this report informs.
Create a shareable link for this report and open it in a new tab.
Never let the model render a verdict. The scan's job is to surface and route, not to conclude whether a child is at risk. Strip any language that sounds like a determination.
Always require a quote. A flag with no verbatim line behind it is unactionable and erodes trust. No quote, no flag.
Treat urgency as mandatory, not optional. Every quoted indicator needs an urgency rating, or amber notes pile up unrouted.
Hard-set review for the highest-risk categories. Some triggers should never depend on a score.
When minors, self-harm, or a named perpetrator appear, set recommend_human_review=true by rule, not by judgement.
Scan each note for harm, crisis, and disclosure indicators, quote the exact line, rate its urgency, and set a recommendation for human review on every flag. Sopact Sense does this in one prompt and never renders a verdict — a person always decides.
No. The AI flags risk, quotes evidence, and recommends human review — then stops. The decision belongs to a trained person. The prompt explicitly forbids a verdict for exactly this reason.
A note grades red when the risk is highest and the routing is missing — for example, children present with no human review recommended. The fix is a hard rule that forces review whenever minors appear.
Open Sopact Sense, paste your program description, and put it to work.
Try in Sopact
