No. An AI agent can register a benefit claim, gather the evidence, check it against the eligibility rules and draft a decision, but a caseworker should make and sign it. A wrong refusal can leave someone without money for rent before any appeal is heard, and a wrong award becomes a debt recovered from a person on a low income.
Why are benefit decisions so hard to reverse?
On paper, almost every benefit decision can be changed. There are reconsiderations, appeals and corrections. In practice, the person lives through the gap between the wrong decision and the right one.
A refusal that is overturned months later still meant missed rent in the meantime. Back-payment repairs the money but not the arrears letter, the borrowing or the strain. A wrong award is no better. The overpayment becomes a debt, often recovered from future payments, from someone who spent the money on essentials in good faith.
That is why public services sit at the strict end of the scale. Our piece on irreversibility in public services sets out the pattern, and irreversibility in insurance claims shows how much a private remedy differs from a public one. The general argument is in why autonomy in regulated industries is a question about reversal.
What can the agent do, and what stays with a caseworker?
| Step | The agent | Stays with a caseworker |
|---|---|---|
| Registering the claim | Creates the record from the application at depth 3 | Identity doubts, suspected duplicate claims |
| Gathering evidence | Requests missing documents, reads what arrives, notes gaps | Deciding what is sufficient when evidence conflicts |
| Applying the rules | Checks each eligibility condition and marks it as met, not met or unclear | Interpreting unclear conditions |
| Calculating entitlement | Drafts the calculation with every input shown | Approving the figure |
| Deciding | Drafts the decision and the reasons | Making and signing the decision |
| Telling the claimant | Sends routine acknowledgements; drafts decision letters in plain language | Sending the decision |
| Reconsideration | Assembles the file and any new evidence | A different caseworker decides |
Registration is the step most often safe to automate. The guidance on letting an AI agent register claims in your claims system applies, with one difference: a public claim's date of receipt can affect entitlement, so the agent must record the date the claim arrived, not the date it processed it.
A worked example
Take a local authority that runs a housing support payment. An applicant uploads a tenancy agreement, three payslips and a bank statement through an online form.
The agent registers the claim with the upload date. It reads the payslips, finds a month missing, and sends a standard request for it. When the payslip arrives, the agent checks the income condition, the residence condition and the household condition. Two are met. The household condition is unclear: the bank statement shows regular payments from an adult who is not named on the application.
The agent does not resolve the question. It drafts the decision as far as it can, sets out the unclear condition with the transactions that raised it, and holds the case for the named caseworker. The caseworker phones the applicant, learns the payments come from a parent who lives elsewhere, records the reason, and decides.
What controls make it safe?
- Draft-only on decisions. The agent's role scope includes drafting decisions, never issuing them.
- Every condition shown. Drafts list each rule and the evidence used, so the caseworker checks the reasoning, not just the outcome.
- No silent inference. Where the evidence suggests something the applicant did not declare, the agent raises it for a person. It never reduces an award on its own reading of a bank statement.
- A full record. Each verdict, source and person is filed, so a reconsideration or a complaint can see exactly what the agent did.
- Clear ownership of scope. Widening what the agent may do is a governance decision in its own right; see who should decide what an AI agent is allowed to do.
Public bodies carry additional transparency and fairness duties, and under the EU AI Act some public uses of AI are high-risk. GDPR Article 22 and its UK equivalent restrict decisions based solely on automated processing that have legal or similarly significant effects. A benefit decision is one of the clearest examples of such an effect.
Where Pryme fits
In Pryme, registering a claim, requesting evidence and drafting a decision are separate capabilities, each certified on the authority's own past cases. Issuing a decision is not among them, so any attempt is held for a named caseworker. Every verdict, with the bound that decided it, is filed in an append-only record that a reconsideration or complaint can read.
Questions people ask
Can the agent approve claims that clearly meet every rule?
It can draft them and mark them as clean, which lets a caseworker approve them faster. The signature stays with a person, because an award turns into a debt if the evidence later proves wrong.
Can an agent tell an applicant their claim will probably succeed?
No. A prediction is heard as a promise. The agent can explain the conditions and say what evidence is still needed.
Does the agent help with fraud?
It can flag inconsistencies, with the evidence that raised them. Deciding that a claim is fraudulent, and any sanction that follows, stays with trained staff.