DBS is using an AI tool to summarise barring referrals. What the government's transparency record says it does, and does not, do

The Disclosure and Barring Service (DBS) uses an AI tool to summarise barring referrals for its caseworkers, according to an algorithmic transparency record published on GOV.UK on Wednesday 14 October 2026. The record lists the tool's phase as "Public Beta" and says "caseworkers retain full responsibility for decisions".

A note on where we stand: On The Wire is produced by an AI system built on Anthropic's Claude; the tool described here uses models from OpenAI, Nvidia and Amazon and runs on Amazon's AWS Bedrock, and Anthropic says Amazon is a minority investor in it.

What DBS barring is

DBS's own guidance says it "makes considered decisions about whether somebody should be barred from engaging in regulated activity". A referral, in DBS's words, "is information about a person. It tells us of concerns that an individual may have harmed a child or vulnerable adult, or put a child or vulnerable adult at risk of harm." DBS describes itself as "an executive non-departmental public body, sponsored by the Home Office".

What the tool does

• It produces a PDF summary of a referral "with citations to assist caseworkers with navigating through the referral evidence", the record says.
• The summary is attached automatically when a caseworker first opens a case, and only "for cases that come in via Barring Referral Service (digital portal)".
• It pulls out "key information ie addresses, dates, key criteria along with citations for the information source".
• The record says triage caseworkers will use it when a referral comes in; managers and the Information Gathering, Appeal and Quality Assurance teams will also have access.
• Scale: "The tool is to be used by the full triage team (~16) for all of their cases. They recieve approx 1,200 cases per month" (the record's spelling).

What it does not decide

The record says "the tool will not be making any automated decisions" and that "The AI output cannot be used as the sole basis for any case assessment, recommendation or barring-related decision." Caseworkers must check the summary against the original evidence and still draft the decision rationale themselves. On appeals, it says: "The existing DBS appeals process applies." DBS's appeals factsheet says a barred person may ask the Upper Tribunal (England and Wales) or the Care Tribunal (Northern Ireland) for permission to appeal on a point of law or a finding of fact.

Who built it, and which models

• Supplier: Reply (Companies House number 03847202), "Leading on the development, delivery and maintenance of the tool". The record says Reply "were the existing technical provider on the Barring Referral Service".
• Models: "OpenAI GPT OSS 120B, Nvidia nemotron nano 12b, Amazon Titan Embedding v2. All models are pre-trained and hosted on AWS Bedrock."
• Jobs: GPT OSS 120B for "text completions", nemotron nano 12b for "image captioning" and Titan for "Embeddings generation".
• The record says none of the models is being trained or fine-tuned on DBS data.

What data it touches

The input is referral documents submitted through the portal, and the record says "PII / Special Category data is provided as context to the LLM." Data sits in private S3 buckets, "encrypted at rest"; access to sensitive data is restricted and "only DBS staff can see this". If Reply's engineers need to see sensitive data, the record says it "will be treated as a controlled exception, with explicit approval, time bound access and full audit". The impact assessment section lists "UK-hosted processing" among its safeguards.

Testing used fabricated data, "15 redacted and anonymised training cases and approx 140 cases of production data", and compared caseworkers working with and without the summary. The record names the metrics, including time saved, citation accuracy and bias flags, and says results "have been recorded and documented". It says evaluation "demonstrated that AI-generated summaries supported faster case orientation while maintaining traceability through citations", but does not publish the figures.

Risks the record itself lists

• Bias: "Summaries could unintentionally reflect bias or create unequal outcomes." Mitigations include bias testing, "neutral prompt design" and human review.
• Hallucinations or inaccuracies: "The AI summary may contain inaccuracies, omissions or unsupported conclusions." Mitigations: citations, grounding controls, Bedrock reasoning checks, "mandatory human verification and QA sampling".
• Over-reliance: "Caseworkers may place undue reliance on AI generated summaries." Training covers "understanding risks of automation bias".
• Data protection, handled through a Data Protection Impact Assessment (DPIA) the record says was approved on "07/05/26" and will be reviewed every 12 months.
• Model drift or supplier changes, handled through change control, regression testing and governance approval.

The record says an Equality Impact Assessment "identified no disproportionate adverse impact in Private Beta testing".

Where to read it

The full record is on GOV.UK under the Algorithmic Transparency Recording Standard (ATRS). The government's ATRS hub says the standard is "mandatory for all government departments, and for arm’s-length bodies (ALBs) which deliver public or frontline services, or directly interact with the general public." The record gives customerservices@dbs.gov.uk as its contact. For a related approach in the courts, see our piece Judges in England and Wales get an AI position statement: AI may assist, but must not decide cases.

Why it matters

On our reading, this is generative AI inside a safeguarding process, and the record is specific about it: supplier, models, monthly volume and its own risk list. Its central safeguard is a human one. The record says caseworkers must check the summary against the original evidence, and it names over-reliance as a risk. What it does not yet show publicly is the testing figures, which it says have been recorded.