Anything that decides a person’s rights, safety or standing. Caseworkers should never let an agent decide eligibility or entitlement, judge safeguarding or vulnerability, impose a sanction or recover money, close a case carrying a risk flag, share case data with another body, or act on an unverified allegation. Each needs a person who can be asked to explain it.
Why are these red lines rather than limits that rise with evidence?
Most agent limits are “not yet”. An agent drafts today and may send routine letters later, once certification on real cases supports it. Depth is earned on evidence, certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day.
Red lines in casework are different. They do not rise with evidence because performance is not the point. Public bodies carry duties of fairness and transparency to the people they serve. Data protection law gives people rights where a decision with significant effects is made solely by automated means. Under the EU AI Act, certain public uses of AI are high-risk. A citizen should be able to ask who decided and get a name.
What are the red lines, one by one?
Deciding eligibility or entitlement. Whether someone gets housing, a benefit, a licence or a grant. The agent prepares the case and may draft a recommendation; the caseworker decides and writes the reason.
Judging safeguarding or vulnerability. A message that mentions self-harm, abuse, a child at risk or an inability to cope must go straight to a person. The agent may flag it and acknowledge receipt with approved support details. It must not assess the risk or decide the response.
Imposing a sanction or recovering money. Reducing a payment, starting recovery of an overpayment, or issuing a penalty affects someone’s income directly. These require a human decision and often a right of review.
Closing a case with a risk flag. A case may look complete on paper while a flag records a concern raised elsewhere. Closure of flagged cases should always be held.
Sharing case data with another body. Disclosure to the police, another department or a third party needs a lawful basis and a judgement about necessity. The agent can draft the request log; a person decides.
Acting on an unverified allegation. An anonymous tip-off that someone is working while claiming, or living with a partner, is an external claim. It can open an enquiry; it cannot by itself change an outcome.
How do these lines appear in a working day?
Take a caseworker in a benefits team. A held item proposes a letter suspending a payment because a data match shows earnings. The letter is accurate on its face and the data match is sealed. But the caseworker opens the case and sees a vulnerability flag recorded last month after the claimant disclosed a serious illness.
The caseworker declines the suspension, calls the claimant, and finds the earnings were sick pay already reported. Without the flag check, the agent’s draft would have stopped someone’s income during treatment.
The lesson the team draws: suspensions on flagged cases should be refused by the agent outright, not held. The caseworker records the reason, and the agent owner changes the limit. Teaching caseworkers to record corrections this way is covered in how to teach caseworkers to bank corrections that improve an agent.
Who else holds similar lines?
Compliance officers and paralegals face their own versions; see what compliance officers should never let an agent do and what paralegals should never let an agent do. The lines also vary by country, because public law and data rules differ. Compare caseworker training in Germany and in the UAE. The role map is in AI agent skills by role.
How should a public body enforce these lines?
Write them into the authority’s own limits so the agent refuses or holds them. Brief every caseworker on each line with a real, anonymised example. Sample held decisions monthly from the record, focusing on flagged and vulnerable cases. And make sure every decision letter names a responsible service and explains how to challenge it.
Where the Academy fits
The Operator track in Pryme Intelligence Academy trains people who review agent work daily, and its practicum requires two weeks of real review decisions with no missed red line, signed off by a Pryme assessor. For caseworkers, that standard matches the stakes: one missed red line can affect someone’s safety or income.
Questions people ask
Can the agent tell an applicant they are not eligible?
It can draft the letter. The decision behind it, and the letter’s release, stay with a caseworker.
What if the agent is more consistent than people?
Consistency helps, and the agent’s drafts make decisions more uniform. But consistency is not accountability, and the law expects a person behind decisions about people.
Should the agent respond to safeguarding disclosures at all?
Only with an approved acknowledgement and support details, then immediate routing to a person.