Public sector · Playbook 19

I've worked in this sector

Benefits and grants application review

Every application read, checked against the scheme rules and made ready for an officer within a day. Missing evidence asked for early, fraud and error checks done fairly, and an officer deciding every award and every refusal.

  • Document AI
  • Agents with tools
  • Retrieval (RAG)
  • Human in the loop
  • UK GDPR (as amended by the DUAA)
  • Public Sector Equality Duty
  • ATRS
  • Subsidy Control Act 2022
  • GovS 013 / GovS 015
  • PFER Act 2025

At a glance

The problem
Applications arrive with a pile of evidence, and officers spend most of their time reading, chasing and re-keying before anyone assesses anything
Typical volume
Around 2,500 applications a month across Council Tax Reduction, crisis support and three local grant schemes, rising sharply when a new scheme opens
Who's involved
Revenues and benefits, the grants team, counter fraud, finance (the section 151 officer), information governance and the DPO, equalities, IT
Pilot
10–12 weeks on one scheme, after a 3–4 week discovery
Platforms
AWS, Azure, or a sovereign or on-prem stack

An illustrative scenario. The numbers are typical examples, not a client case study. The cloud services and UK rules were checked on 9 November 2026.

The business problem

People applying for local support send what they have: online forms with free-text answers, phone photos of payslips and bank statements, tenancy agreements, letters about other benefits, and sometimes a GP letter about a disability. Businesses and community groups applying for grants send accounts, quotes, constitutions and project plans. Before an officer can assess anything, they have to read it all, work out what is missing, chase it, and re-key the figures into the benefits or grants system.

In a typical council, that reading and chasing takes more of an officer's day than the assessment itself. Applications wait days for a first look. When a new grant scheme opens, hundreds arrive in a week and the queue grows faster than the team can read. Applicants who struggle with forms, often the people the scheme is for, are the ones most likely to submit incomplete evidence and wait longest.

The decisions matter, and so does public money. A wrong refusal can leave a household in arrears or a small charity without its project. A wrong award is money the council can't spend elsewhere, and fraud against grant schemes is a known risk. The council has to treat people fairly, show it has thought about equality, and be able to explain any decision if it is challenged, while spending less time on admin and more on the cases that need judgement.

What AI changes

  • Every application is read and its evidence extracted within minutes, whatever form it arrives in.
  • The scheme's eligibility rules run as tested, versioned code, not as a model's opinion. The same facts always give the same answer, and the officer sees which rule produced it.
  • Gaps are found on day one, and a clear request for the missing evidence is ready for an officer to send, in plain English.
  • Each application reaches an officer with a draft assessment: the facts found, the rules applied, the calculation, and anything that needs their judgement.
  • Fraud and error checks run on every application in the same way, and produce leads for a trained person to look at, never an automatic refusal or stopped payment.
  • An officer decides every award and every refusal, and every step is logged so the council can explain and defend each decision.

How success is measured

Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.

Time to first look

How long an applicant waits before anyone acts on their application.

ExampleFrom 5 working days to under 1

Evidence requests

The share of applications needing more evidence, and how many rounds it takes.

ExampleOne round, not three

Officer time per application

Time moved from reading and re-keying to assessment and judgement.

ExampleDown by 30–40%

Override rate

How often officers change the draft assessment. Very high means it isn't helping; near zero can mean people have stopped checking.

ExampleStable, and reviewed weekly

Fraud and error

Overpayments and wrong awards found after payment, and the share of fraud flags that turn out to be genuine.

ExampleFewer wrong awards, no rise in false flags

Fairness

Award rates, refusal rates and flag rates by protected characteristic and by channel, where the data allows.

ExampleNo group worse off

UK regulation and controls

Cross-sector rules are on the UK regulation map.

  • ICOAutomated decisions (UK GDPR, as amended by the Data (Use and Access) Act 2025)

    Since 5 February 2026, the strict limit on solely automated significant decisions applies mainly where special category data is involved, and applications often include health or disability evidence. Every other significant automated decision needs safeguards: telling people, letting them make representations and contest it, and human intervention. An officer making every award and refusal keeps the design clearly within the rules.

  • ICOUK GDPR and the Data Protection Act 2018

    The lawful basis is the council's public task, not consent. That means a DPIA before the pilot, using evidence only for the scheme it was given for unless the law allows more, processing in the UK, a clear retention period, and a contract that stops providers training their models on applicants' data.

  • Equality and Human Rights CommissionPublic Sector Equality Duty (Equality Act 2010, s149)

    The council must have due regard to equality when it designs and runs the service. Record an equality impact assessment before the pilot, test outcomes and fraud flags by group, and keep a non-digital route for people who can't use the online form.

  • GDSAlgorithmic Transparency Recording Standard (ATRS)

    The ATRS is mandatory for central government departments and arm's-length bodies that serve the public, for tools that significantly influence decisions about people. Councils aren't yet required to use it, but many publish records voluntarily. Publish one before go-live: it's the simplest way to show residents what the tool does and doesn't do.

  • Cabinet OfficeCounter fraud and grants standards (GovS 013 and GovS 015)

    The counter fraud standard expects fraud risk to be assessed and managed. The grants standard requires a proportionate fraud risk assessment and due diligence for every central government scheme. Councils use the same approach as good practice. Checks run on every application, flags are explained, and a trained person decides what happens next.

  • Cabinet Office and DWPPublic Authorities (Fraud, Error and Recovery) Act 2025

    Royal Assent on 2 December 2025, with most provisions commenced in stages during 2026. It gives the Public Sector Fraud Authority powers to investigate and recover fraud against public bodies outside tax and social security, with codes of practice and independent oversight. Its benefits measures keep a member of staff involved in every entitlement decision. Design the same way: data can raise a question, but a person answers it.

  • BIST and the CMASubsidy Control Act 2022 (business grants)

    A grant to a business or a trading charity can be a subsidy. The scheme's subsidy basis, any cumulative allowance per enterprise, and the duty to publish larger awards on the subsidy database (currently those over £100,000) all become rules in code and fields in the case record.

  • HM TreasuryManaging Public Money (central government) and the section 151 officer (councils)

    Public money must be spent with regularity, propriety and value for money. Keep a clear trail from each award to the scheme rules and the evidence, and keep the approval limits in the platform's permissions, not in a prompt.

The solution

Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.

Benefits and grants application review: reference architecture
  1. 01Channels

    • Online application form
    • Email and post (scanned)
    • Partners and advice agencies
  2. 02Intake

    • API GatewayAPI ManagementKong, or your existing API gatewayAPI for forms and uploads
    • Amazon S3Blob StorageCeph, or IONOS S3 Object StorageEvidence store
  3. 03Intelligence

    • Amazon TextractContent Understanding (Document Intelligence)Docling plus an open-weight vision modelRead evidence
    • AWS LambdaAzure FunctionsA rules engine (e.g. OpenFisca or Drools) on KubernetesScheme rules as code
    • Bedrock Knowledge BasesFoundry IQ (Azure AI Search)OpenSearch, or PostgreSQL with pgvectorScheme guidance
    • Agent
      Bedrock AgentCoreFoundry Agent ServiceLangGraph on Kubernetes, with open-weight models via vLLMAssessment agent
    • Agent
      Bedrock AgentCoreFoundry Agent ServiceLangGraph on Kubernetes, with open-weight models via vLLMFraud and error check agent
  4. 04Orchestration

    • Step FunctionsLogic Apps / Durable FunctionsTemporalApplication workflow
    • EventBridge + SQSEvent Grid + Service BusApache Kafka or RabbitMQEvents and queues
  5. 05Systems of record

    • Revenues and benefits system
    • Grants management system
    • Document management
    • Data matching and counter fraud tools
  6. 06People

    • Person
      Assessment officer: decides
    • Person
      Senior officer: refusals and large awards
    • Person
      Counter fraud investigator
  • Cloud or platform service (AWSAzuresovereign / on-prem)
  • AI agent
  • A person decides
  • Your existing systems
  • Outside your organisation

Components, side by side

Benefits and grants application review: components on AWS, Azure and a sovereign or on-prem stack
LayerAWSAzureSovereign / on-prem
Intake: forms, uploads, scanned postKeep the paper and phone routes. Scan post into the same flow so nobody is disadvantaged by the channel they use.API Gateway, S3, Amazon SESAPI Management, Blob Storage, Logic Apps email connectorKong, Ceph or IONOS S3 Object Storage, your mail server
Reading evidenceTest on real phone photos of payslips and bank statements, not clean PDFs. Every extracted figure links back to the page it came from.Amazon Textract, or Bedrock Data AutomationAzure Content Understanding (includes Document Intelligence)Open-source parsing (e.g. Docling) plus an open-weight vision model
Eligibility rulesThe model never decides eligibility. The scheme owner signs off the rules and a set of worked examples, and every change is tested against them.Rules as code in AWS Lambda, versioned and testedRules as code in Azure Functions, versioned and testedA rules engine (e.g. OpenFisca or Drools) on Kubernetes
Reasoning and draftingUsed to read, summarise and draft, not to decide. AWS and Azure both offer a choice of models, including Claude.Amazon BedrockMicrosoft FoundryOpen-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub
Scheme guidance (retrieval)So drafts quote the council's own scheme policy and guidance, at the version in force on the date of the application.Bedrock Knowledge BasesFoundry IQ, built on Azure AI SearchOpenSearch, or PostgreSQL with pgvector
Agent runtimeNew builds use AgentCore; the original Bedrock Agents closed to new customers on 30 July 2026.Amazon Bedrock AgentCoreFoundry Agent ServiceAn open-source agent framework (e.g. LangGraph) on Kubernetes
WorkflowLong-running: an application can wait weeks for evidence without losing its place.AWS Step FunctionsLogic Apps or Durable FunctionsTemporal, or n8n
Middleware and eventsAmazon EventBridge, SQSEvent Grid, Service BusApache Kafka or RabbitMQ
Data matchingFor the council's own matching rules, alongside national exercises it already takes part in. Results are leads for a person, never decisions.AWS Glue and Amazon AthenaMicrosoft FabricApache Spark or PostgreSQL on Kubernetes
Human reviewOfficers review and decide where they already work, not in a new screen.In the officers' existing benefits or grants systemIn the officers' existing benefits or grants systemIn the officers' existing benefits or grants system
GuardrailsBedrock GuardrailsFoundry guardrails (Azure AI Content Safety)Open-source guardrails (e.g. NeMo Guardrails, Llama Guard)
Identity, keys, data protectionApproval limits by role (who can award what) are enforced here.IAM Identity Center, KMS, MacieEntra ID, Key Vault, PurviewKeycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM
Monitoring and auditLog every extraction, rule result, draft and officer decision, with the reason.CloudWatch, CloudTrailAzure Monitor, Log AnalyticsPrometheus, Grafana, OpenTelemetry
UK data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency.eu-west-2 (London)UK South / UK WestYour own UK data centre, or a UK or EU provider

On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →

Hybrid and on-prem

Many organisations keep core systems and sensitive data on their own servers. This is how the same design works across both.

Stays on-prem

The revenues and benefits system, the grants system and the document store stay where they are, whether in the council's own data centre or hosted by its supplier. Nothing in this design needs them to move.

Uses cloud AI

Evidence reading, scheme guidance retrieval and the two agents can run in a UK cloud region, or on the council's own GPUs with open-weight models if evidence must not leave its estate. The rules engine can run in either place.

How they connect

A private link (AWS Direct Connect or Azure ExpressRoute) or the council's existing network connection, through the API layer. Only the evidence and fields the assessment needs are sent, and results come back as events onto the council's queue.

If the link drops

Applications keep arriving and queue safely on the council's side. Officers assess them exactly as they did before, and the backlog is read and checked when the link returns.

The agents

What each agent may do through its tools, what it may not, and where a person signs off. The limits are enforced by the platform, not the prompt.

Assessment agent

May do

  • Read the application and every piece of evidence
  • Run the scheme's eligibility rules and calculation with the extracted facts
  • Look up the scheme policy and guidance in force on the application date
  • List missing or unclear evidence and draft a plain-English request for it
  • Draft the assessment for the officer, with each fact linked to its source
  • Create and route a task in the benefits or grants system

May not do

  • Award, refuse, reduce or stop anything
  • Send any letter or message to the applicant
  • Change the eligibility rules, the calculation or a figure an officer has confirmed
  • Use health or disability evidence for anything except the purpose it was given for
  • Close or withdraw an application

Where a person signs offAn officer decides every award and every refusal, and sends every evidence request. Refusals and awards above a set value go to a senior officer. The agent's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.

Fraud and error check agent

May do

  • Run the agreed data-matching and consistency checks on every application
  • Check for duplicate applications, bank details and documents across the scheme
  • Request due diligence results for organisations applying for grants
  • Explain each flag in plain English, with the evidence behind it
  • Route flagged applications to the counter fraud team

May not do

  • Refuse, suspend or delay an application on its own
  • Label anyone as fraudulent
  • Use protected characteristics, or obvious proxies for them, as a reason to flag
  • Contact the applicant or any third party

Where a person signs offA trained counter fraud officer reviews every flag and decides what happens next. A flag is a question, not a finding, and the assessment officer still makes the award decision.

Middleware and integration

The benefits and grants systems stay the record. The AI reaches them only through their published APIs or the integration layer the council already runs, never directly into the database.

Events decouple the AI from the systems of record. Each new application or piece of evidence becomes a message on a queue. If the benefits system is slow or down for its overnight run, work waits safely, nothing is lost, and every message carries an ID so a retry never creates a duplicate case.

The rules engine is its own service with its own tests. When a scheme changes in April, the new rules go live on a date, and applications are assessed against the rules in force when they were made.

The AI can be switched off without stopping the service. If it's paused, applications go straight to officers, exactly as they did before.

Process and people

  • A named owner for each scheme's rules, with sign-off of the rules and worked examples before any change goes live.
  • An exception queue for applications the AI can't read or isn't sure about, with a clear owner and a target time.
  • Override rules, so officers know when and how to disagree with a draft, and their reasons are captured.
  • A weekly review of a sample of drafts and flags, including refused applications and flagged ones that turned out to be genuine.
  • An equality impact assessment before the pilot and quarterly outcome reports by group, owned by the head of service.
  • Officers and counter fraud staff trained before go-live, including what the AI can't see and how to explain a decision.
  • An ATRS record published before go-live and updated when the tool changes.

Delivery plan

  1. 3–4 weeks

    Discovery

    Map today's process for each scheme, measure the baseline, check data access, write down the rules, and pick the first scheme.

  2. Gate · must be true to passBaseline agreed, rules and worked examples signed off, DPIA and equality impact assessment started, counter fraud and finance involved.

  3. 8–10 weeks

    Pilot

    One scheme, in shadow mode: the AI drafts, officers work as normal, and the two are compared.

  4. Gate · must be true to passExtraction accuracy and rule results within agreed limits, no group worse off, fraud flags explainable, DPO and section 151 officer content.

  5. 4–6 weeks

    Production

    Drafts go live for the first scheme, with monitoring, support, a rollback plan and the ATRS record published.

  6. Gate · must be true to passMonitoring, support model, rollback plan, officer training and ATRS record in place.

  7. Ongoing

    Scale

    Add schemes one at a time, starting with the next grant round, and reuse the rules engine and evidence reading.

  8. Gate · must be true to passEach new scheme passes the same pilot gate, with its own rules sign-off.

Roles and RACI

Who is responsible, accountable, consulted and informed, by activity
ActivityHead of serviceProduct ownerSolutions architectI can fill thisTechnical PMI can fill thisEngineersI can fill thisDPO & equalitiesCounter fraud lead
Business case and benefitsAccountableResponsibleConsultedConsultedInformedConsultedConsulted
Scheme rules and worked examplesAccountableResponsibleConsultedInformedResponsibleConsultedConsulted
Target architectureInformedConsultedAccountable and ResponsibleConsultedConsultedConsultedConsulted
DPIA, equality impact and ATRS recordAccountableConsultedConsultedInformedInformedResponsibleConsulted
Build and integrationInformedConsultedAccountableResponsibleResponsibleInformedInformed
Accuracy and fairness testing on real applicationsInformedAccountableResponsibleConsultedResponsibleConsultedConsulted
Go-live decisionAccountableResponsibleConsultedResponsibleInformedConsultedConsulted
Run, monitor and improveInformedAccountableConsultedConsultedResponsibleConsultedConsulted

Business case and benefits

Head of service
A · Accountable
Product owner
R · Responsible
Solutions architect
C · Consulted
Technical PM
C · Consulted
Engineers
I · Informed
DPO & equalities
C · Consulted
Counter fraud lead
C · Consulted

Scheme rules and worked examples

Head of service
A · Accountable
Product owner
R · Responsible
Solutions architect
C · Consulted
Technical PM
I · Informed
Engineers
R · Responsible
DPO & equalities
C · Consulted
Counter fraud lead
C · Consulted

Target architecture

Head of service
I · Informed
Product owner
C · Consulted
Solutions architect
A/R · Accountable and Responsible
Technical PM
C · Consulted
Engineers
C · Consulted
DPO & equalities
C · Consulted
Counter fraud lead
C · Consulted

DPIA, equality impact and ATRS record

Head of service
A · Accountable
Product owner
C · Consulted
Solutions architect
C · Consulted
Technical PM
I · Informed
Engineers
I · Informed
DPO & equalities
R · Responsible
Counter fraud lead
C · Consulted

Build and integration

Head of service
I · Informed
Product owner
C · Consulted
Solutions architect
A · Accountable
Technical PM
R · Responsible
Engineers
R · Responsible
DPO & equalities
I · Informed
Counter fraud lead
I · Informed

Accuracy and fairness testing on real applications

Head of service
I · Informed
Product owner
A · Accountable
Solutions architect
R · Responsible
Technical PM
C · Consulted
Engineers
R · Responsible
DPO & equalities
C · Consulted
Counter fraud lead
C · Consulted

Go-live decision

Head of service
A · Accountable
Product owner
R · Responsible
Solutions architect
C · Consulted
Technical PM
R · Responsible
Engineers
I · Informed
DPO & equalities
C · Consulted
Counter fraud lead
C · Consulted

Run, monitor and improve

Head of service
I · Informed
Product owner
A · Accountable
Solutions architect
C · Consulted
Technical PM
C · Consulted
Engineers
R · Responsible
DPO & equalities
C · Consulted
Counter fraud lead
C · Consulted

R ResponsibleA AccountableC ConsultedI Informed· not involved

Where I'd fit

Option A

Lead the architecture

If there's no architect on the programme, I'd own the design end to end: the rules engine, the evidence reading, the integration with your benefits and grants systems, and the gates. The DPIA, the equality work and the ATRS record would sit in the plan from the start, not be added before go-live. I'd report to your programme board.

Option B

Support your architect

If your partner or your own team already has an architect, I'd take on the rules as code, the agents and the integration with your systems of record under their lead, and give them a second senior opinion on the rest, especially where the AI stops and an officer starts.

Where this goes wrong

  1. Goes wrong

    The model is asked whether someone is eligible, and gives a different answer on a different day.

    How the design avoids it

    Eligibility runs as tested code from signed-off rules. The model reads and drafts; it never decides.

  2. Goes wrong

    A data match is treated as proof, and a genuine applicant is refused or left waiting.

    How the design avoids it

    Every flag goes to a trained person with the evidence behind it. No refusal, suspension or delay happens because of a flag alone.

  3. Goes wrong

    The online route works well, so the phone and paper routes are quietly run down.

    How the design avoids it

    Keep every channel feeding the same flow, and report outcomes by channel as well as by group.

  4. Goes wrong

    A new grant scheme opens with rules nobody has written down precisely.

    How the design avoids it

    Write the rules and worked examples before the scheme opens, and test the code against them.

  5. Goes wrong

    Information governance and equalities see it a week before go-live.

    How the design avoids it

    DPIA, equality impact assessment and ATRS record at the discovery gate, not the production gate.

On the horizon

Changes already announced or under way that affect this design. More in the UK horizon scan.

  1. 2026–2027
    The ICO's code of practice on AI and automated decision-making

    Regulations in force since May 2026 require the ICO to produce a statutory code. Keep the decision log, the explanations and the officer review records in a form you could show the ICO.

  2. 2026–2027
    The Public Authorities (Fraud, Error and Recovery) Act 2025 bedding in

    The Public Sector Fraud Authority's new powers and codes of practice came into force in stages during 2026, with independent oversight and published reports. Expect closer scrutiny of how public bodies find and act on fraud and error. Keep every flag, its evidence and the officer's decision on record.

  3. Over time
    The ATRS extended beyond central government

    The government has said it intends to extend the mandatory standard across the wider public sector over time. Publishing a record now means the council is ready if the requirement reaches it.

  4. April 2027 to April 2029
    Local government reorganisation in England

    Most two-tier areas are being replaced by new unitary councils, with most vesting days in April 2028. Schemes, rules and systems will be merged. Keep each council's scheme rules as separate, versioned code so they can be compared and combined, not untangled.

  5. Late 2026–2028
    The Cyber Security and Resilience Bill

    Medium and large managed service providers come into regulation for the first time. Expect your IT and AI suppliers to face new duties, and your contracts and supplier assurance to need updating.

Questions to ask your partner or vendor

Good answers are specific. Vague answers to any of these are worth following up.

Copy them into your next supplier meeting

  1. Where exactly does the eligibility decision come from: tested rules, or a model's answer?
  2. What happens to an application when a fraud or data-matching flag is raised, and who decides?
  3. How will we show the outcomes are fair across groups, and what data will we need to do it?
  4. Which actions can the agents take without an officer, and where is that enforced?
  5. Where is applicants' evidence processed and stored, and is it used to train anyone's model?
  6. Who owns the scheme rules, test cases and prompts after you leave, and how do we change them in April?

Planning something like this?

Whether you need someone to lead the design or to support the architect you already have, let's talk it through.

Sam Naseem
Sam NaseemBirmingham, UK · UK & Europe
Available now