Partly. Caseworkers are accountable for every decision they approve about a person's case, whether it is a support assessment, a reduction or a closure. They are not accountable for errors the agent made out of their sight, or for bounds they never set. The deciding factor is what the caseworker approved and what the record shows they could see.
Why is the caseworker's approval the point of accountability?
Casework decides things that matter to people: support, housing, entitlements, safeguarding. Public bodies also face extra duties of transparency and fairness, and under the EU AI Act some public uses of AI are classed as high-risk. Under GDPR Article 22, people have protections against decisions based solely on automated processing that significantly affect them.
So a well-run service keeps the decision with a person. The agent gathers evidence, checks it against policy and drafts the letter. The caseworker approves or changes it. That approval is the moment accountability attaches.
What may the agent do, and what stays with the caseworker?
| The agent may | The caseworker keeps |
|---|---|
| Collect records from the case file and partner systems | Judging what the evidence means for this person |
| Check eligibility rules against the evidence | Any decision that reduces, ends or refuses support |
| Draft letters, case notes and referrals | Approving anything the person will receive |
| Flag missing documents and deadlines | Safeguarding concerns and urgent referrals |
| Hold anything outside its bounds | Explaining a decision to the person affected |
The last row is easy to miss. If a caseworker cannot explain why the agent's draft reached its conclusion, they should not approve it.
How does the record show who decided?
Every action the agent takes produces a verdict: allowed, held, refused or withheld, with the bound that decided it. For a held decision, the record shows the sources the agent used, stamped by provenance. An income figure from the official payroll record is sealed. An income figure the applicant wrote on a form is declared. A letter from a landlord is signed if the landlord is identifiable.
It also records the caseworker who approved, the time, and any change they made to the draft. If a person later challenges the decision, the service can reconstruct it step by step, including whether the caseworker saw a conflict between two sources.
An example: a reassessment that reduced support
Take a local authority reviewing a household's support. The agent finds a higher income figure in a declared update from six months ago and drafts a letter reducing support. A sealed payroll record from last month shows the income fell again, but the agent's bounds did not include that source. The draft notes only the declared figure.
The caseworker approves it. The household appeals. The record shows the agent's sources, the missing payroll feed, and the caseworker's approval. Accountability is shared: the caseworker approved a reduction without checking current income, and the service owner gave the agent too narrow a set of sources. Both are visible, which is the point.
What controls make accountability fair?
- Train reviewers on real cases first. A two-week practicum for caseworkers puts them through supervised decisions before they approve alone.
- Hold every adverse decision. Reductions, refusals and closures should always wait for a named person.
- Make conflicts visible. If two sources disagree, the agent should hold and show both, not choose one.
- Keep competence current. As the agent's certificate is renewed every 90 days and its bounds change, caseworkers need to keep up; see how to prepare caseworkers for 90-day recertification.
Payments teams face a similar split between preparation and approval; compare payments operations staff and agent mistakes. The full topic is covered in AI literacy and certification for people who work with agents.
Questions people ask
Can a caseworker refuse to approve an agent's draft?
Yes, and they should whenever they cannot explain it. A rejected draft is filed in the record with the caseworker's reason, which helps the agent's builders improve its sources and bounds.
Who is accountable if the agent sends a letter on its own?
If the service allowed the agent to send certain letters within bounds, the people who set those bounds own the result. Adverse decisions should never sit inside those bounds.
Does a case management system's audit log do the same job?
Usually not. A system log records who clicked. A governed record also shows what the agent proposed, why, from which sources, and which bound held it.