Every letter, email and form read and routed the day it arrives, supporters' own instructions applied correctly, and every Gift Aid claim backed by a valid declaration. People handle complaints, vulnerable supporters and the claim itself.
At a glance
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.
A large fundraising charity's supporter care team handles everything that isn't a gift arriving cleanly online. Change-of-address letters, emails asking to stop the catalogue, forms returned with a Gift Aid tick, calls about a direct debit taken twice, enquiries about leaving a gift in a will, and complaints. Each one is read, matched to the right supporter record and keyed into the CRM by hand. At peak times, such as after a big appeal, the backlog runs to weeks.
Gift Aid sits behind that backlog. Paper declarations are scanned and keyed, online ones arrive from donation platforms in different formats, and each donation then has to be matched to a valid declaration before it can be claimed. A missing home address, a declaration that was cancelled, or a payment that was really a lottery ticket all make a claim wrong. Teams claim late to be safe, and some eligible Gift Aid is never claimed at all.
The risks run both ways. Applying a supporter's opt-out late, or contacting someone who asked to stop, breaks the law and the Code of Fundraising Practice and damages trust. Claiming on a donation without a valid declaration means repaying HMRC. And a letter from a supporter who seems confused, bereaved or unwell needs a person's care, not a template.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
From a letter or email arriving to the instruction being applied or a person picking it up.
ExampleSame working day
Opt-outs and Fundraising Preference Service suppressions applied within the agreed time, checked by audit.
ExampleEvery one
A weekly sample of AI-applied changes checked against the supporter's own words.
ExampleAbove 99%, every error reviewed
Share of eligible donations matched to a valid declaration before the claim deadline.
ExampleUp, measured against last year
Donations removed from a claim at the officer's check, and any adjustment after HMRC review.
ExampleFalling, and none after submission
Complaints, bereavements and possible vulnerability handed to a person, not answered automatically.
ExampleEvery one in the audit sample
Cross-sector rules are on the UK regulation map.
A valid declaration names the charity, the donor's full name and home address, says which donations it covers, and explains that the donor must pay enough UK income tax or capital gains tax to cover the claim. A verbal declaration needs written confirmation that the donor can cancel within 30 days. Declarations are kept for six years after the accounting period of the last donation they cover. The AI reads and checks declarations; it never creates, completes or backdates one.
Claims are made through HMRC's online service within 4 years of the end of the financial period the donation was received in, and claims of more than 1,000 donations must use software. Small cash and contactless gifts of £30 or less can be claimed under the Small Donations Scheme, up to £8,000 a tax year, within 2 years. The AI prepares the claim; the Gift Aid officer checks it and an authorised official submits it.
Complaints follow a clear, published procedure and are resolved without unnecessary delay. Fundraising involving people in vulnerable circumstances needs extra care, so any sign of it goes to a person. Requests from the Fundraising Preference Service are retrieved and actioned within 21 days.
Since 5 February 2026 charities can email or text people who showed interest in their purpose without asking for consent first, if they offered a clear opt-out when the details were collected and in every message (ICO guidance, 28 April 2026). PECR fines are now at UK GDPR levels. Replies to supporters stay service messages with no fundraising asks, and consent only changes on the supporter's own instruction.
A DPIA before the pilot, processing in UK regions, and no training of providers' models on supporter data. Since 19 June 2026, data protection complaints must be easy to make and acknowledged within 30 days, so the agent recognises them and routes them to the data protection team with a reference. Subject access and erasure requests go to people too.
Trustees stay responsible for how the charity uses AI with supporters' data and money. A significant data breach is a reportable serious incident, reported promptly; a person decides, and the system's logs show exactly what the AI changed and when.
Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.
01Channels
02Intake
03Intelligence
04Orchestration
05Systems of record
06People
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Intake: post, email, formsThe mailroom keeps scanning as now. Scans and emails land in one store, each with a reference back to the original. | API Gateway, Amazon S3 | API Management, Blob Storage | Kong or your existing API gateway; Ceph or IONOS S3 Object Storage |
| Reading documentsTest on real handwriting and ticked boxes. AWS documentation (checked November 2026) shows Bedrock Data Automation's London profile processing in eu-west-2 only; recheck before relying on it, or use Textract plus a model running in London. | Amazon Textract, or Bedrock Data Automation | Azure Content Understanding (includes Document Intelligence) | Open-source parsing (e.g. Docling) plus an open-weight vision model |
| Drafting and reasoningChoose models that run in the UK region; some are only offered through cross-region or global processing. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Policies and replies (retrieval)Only approved supporter care policies, reply templates and the charity's Gift Aid procedures, each with an owner and a version. | Bedrock Knowledge Bases | Foundry IQ, built on Azure AI Search | OpenSearch, or PostgreSQL with pgvector |
| Agent runtimeAgentCore Runtime, Gateway and Policy are listed for London. The original Bedrock Agents closed to new customers on 30 July 2026. | Amazon Bedrock AgentCore | Foundry Agent Service | An open-source agent framework (e.g. LangGraph) on Kubernetes |
| Permissions on CRM changesField-level rules outside the prompt: the agent's service account can change an address or set an opt-out, and nothing else. | Amazon Verified Permissions, IAM | Entra ID, Azure Policy and app roles | Keycloak plus Open Policy Agent |
| Personal data detectionMasks bank details and health information before text reaches logs or drafts. | Amazon Comprehend PII detection | Azure Language in Foundry Tools (PII detection) | Open-source PII detection (e.g. Presidio) |
| WorkflowHolds each item until a person acts, with timers for the Code's and HMRC's deadlines. | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| Middleware and events | Amazon EventBridge, SQS | Event Grid, Service Bus | Apache Kafka or RabbitMQ |
| Outbound confirmationsAmazon Pinpoint reached end of support on 30 October 2026; use AWS End User Messaging. Service messages only. | Amazon SES, AWS End User Messaging | Azure Communication Services (email, SMS) | Your existing email and SMS gateway |
| GuardrailsBlocks fundraising asks in service replies, and advice on tax or wills. | Bedrock Guardrails | Azure AI Content Safety | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protection | IAM Identity Center, KMS, Macie | Entra ID, Key Vault, Purview | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM |
| Monitoring and auditEvery CRM change logged with the source document, the rule that allowed it and the before and after values. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencySupporter and donation data. Many charities' own policies require UK hosting. | eu-west-2 (London) | UK South / UK West | Your own UK data centre, or a UK/EU provider (IONOS AI Model Hub runs in Germany) |
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 →
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 supporter CRM, the direct debit bureau link, the finance system and the paper archive stay where they are, whether in the charity's own data centre or hosted by the CRM supplier. The mailroom and its scanners stay too. The AI never becomes the supporter record.
Uses cloud AI
Reading scanned post, classifying messages, drafting replies and matching donations to declarations run in a UK cloud region, or on the charity's own GPUs for the sovereign option. Only the documents and fields each task needs are sent.
How they connect
Through the CRM's published API and an API gateway, over a private link such as AWS Direct Connect or Azure ExpressRoute where the CRM is on-prem. The direct debit bureau and HMRC are reached by the same routes as today, by people.
If the link drops
Post and email queue safely and the team works the CRM by hand, as now. The agent resumes from the queue in order, checking each record hasn't changed in the meantime, so nothing is applied twice or over a newer instruction.
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.
Supporter care agent
Where a person signs offA person checks every draft reply and every change the agent wasn't sure about. The agent's write access is limited by field and by action in the platform, not by instructions in a prompt.
Gift Aid matching agent
Where a person signs offThe Gift Aid officer checks the exceptions and a sample of matched donations. An authorised official submits the claim through HMRC's online service, as now.
The supporter CRM stays the record. The agent works through the CRM's published API with its own service account, and every write is checked by a policy service before it reaches the CRM. Agree in discovery what the CRM supplier's API allows, and whether it can record the source document against each change.
Direct debits keep their existing route. Instructions and amendments go to the direct debit bureau as now, made by a person, and Bacs reports of cancellations and failed collections come back into the CRM through the same integration. The agent reads those reports to answer queries; it never sends instructions.
Online donation platforms send declarations and donations in their own formats. A small mapping layer turns each into one internal shape, keeps the original file, and records who collected each declaration, so the audit trail runs from a claimed donation back to the declaration image or online record.
Events keep the AI decoupled. Each arriving item becomes a message on a queue. If the AI service is slow or down, items wait and the team works as before. Each item carries a reference so it is applied once, in order.
Map today's post, email and Gift Aid processes, measure the baseline backlog and claim errors, and agree what the CRM API allows.
Gate · must be true to passBaseline agreed, CRM route confirmed, DPIA started, permission rules written down.
Address changes, opt-outs and paper declarations first, in shadow mode: the agent proposes, the team applies, and the two are compared. Then one claim prepared in parallel with the usual process.
Gate · must be true to passAccuracy within agreed limits, no consent errors, a parallel claim that matches, supporter care lead and finance sign-off.
The agent applies permitted changes and prepares claims, with monitoring, daily sampling and a rollback plan.
Gate · must be true to passMonitoring, support, rollback plan and team training in place.
Add online platforms' files, call notes, and more instruction types one at a time.
Gate · must be true to passEach new instruction type passes the same shadow-mode gate.
| Activity | Director of supporter care | Product owner | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | Gift Aid and finance lead | DPO | SecurityI can fill this |
|---|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Informed | Consulted | Informed | Informed |
| Target architecture and CRM integration | Consulted | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted | Consulted |
| Permission rules for the agents | Accountable | Responsible | Responsible | Informed | Consulted | Consulted | Consulted | Consulted |
| DPIA and PECR review | Consulted | Consulted | Consulted | Informed | Informed | Informed | Accountable and Responsible | Consulted |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Consulted | Informed | Consulted |
| Gift Aid claim checks | Informed | Consulted | Consulted | Consulted | Consulted | Accountable and Responsible | Informed | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted | Consulted |
| Run, monitor and improve | Accountable | Responsible | Consulted | Consulted | Responsible | Consulted | Consulted | Consulted |
R ResponsibleA AccountableC ConsultedI Informed· not involved
Option A
If there's no architect on the programme, I'd own the design end to end: the document reading, the agents and the hard limits on what they can change in your CRM, the Gift Aid matching and the gates between pilot and production. I'd report to your programme board and work closely with your supporter care and finance leads.
Option B
If your partner, your CRM supplier or your own team already has an architect, I'd take on the agents, the permission rules and the Gift Aid matching under their lead, and give them a second senior opinion on CRM integration, data protection and security.
Goes wrong
The agent reads a vague message as consent and opts someone in.
How the design avoids it
It can only set opt-outs. Opt-ins come from the charity's own consent forms, and anything ambiguous goes to a person.
Goes wrong
Gaps in declarations get filled from other records to make claims work.
How the design avoids it
The agent can't write to declaration fields, and missing details are asked for by a person. The claim only includes donations backed by a complete declaration.
Goes wrong
A bereaved or confused supporter gets a template reply.
How the design avoids it
Routing rules for bereavement and vulnerability that the supporter care lead owns, tested on real letters, with replies drafted for a person, never sent.
Goes wrong
Changes land on the wrong supporter because of duplicate records.
How the design avoids it
Matches below a set confidence go to a person, the agent can't merge records, and each change is logged with its source for easy reversal.
Goes wrong
Lottery tickets and event fees are claimed as donations.
How the design avoids it
Payment types are mapped in discovery, ineligible ones are flagged, and the officer checks every exception before the claim.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Record, for each supporter, how they came to you and the opt-out they were offered, so the soft opt-in can be shown to apply. The agent never decides it does.
Run the pilot's parallel claim through the new screens, and keep the claim file in a format your software can submit.
Keep the trail from each claimed donation to its declaration complete and quick to produce.
Make sure the agent recognises a data protection complaint in any channel and routes it with a reference. It never closes one.
If the agent later handles shop orders, event bookings or lottery queries, design to it: the charity answers for its agent as it would for a member of staff.
Good answers are specific. Vague answers to any of these are worth following up.
Whether you need someone to lead the design or to support the architect you already have, let's talk it through.
