Business case and benefits
- Director of services
- A · Accountable
- Head of impact
- R · Responsible
- Solutions architect
- C · Consulted
- Technical PM
- C · Consulted
- Engineers
- I · Informed
- DPO
- C · Consulted
- Trustees
- I · Informed
Outcome figures and themes drawn from caseworkers' notes for the trustees' report, funder reports and internal learning. De-identified, every figure traceable to records, nobody scored, and trustees approving what is published.
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 support charity's case management system holds the evidence of its impact: what people came with, what was done, and what changed. Most of it is in free-text notes. Structured fields capture some outcomes, but they are filled in unevenly, and the detail that funders and trustees want sits in the narrative.
So reporting is manual. Each funder wants its own measures, periods and definitions. An impact team pulls spreadsheets, service leads read samples of notes, and figures are adjusted until they look right. The same number appears differently in two reports. Small services report counts so low that a reader could work out who is being described. Learning that should change services, such as a rise in a particular kind of need, is noticed months late, if at all.
The bar has also gone up. Under Charities SORP 2026, for periods starting on or after 1 January 2026, charities above £500,000 income must review performance against their objectives and explain the impact they are making, and the largest must also review fundraising performance. Impact claims need to be backed by evidence, and trustees put their names to them.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
From the request to a report ready for sign-off.
ExampleFrom 2 weeks to 2 days
Share of published figures linked to a query, a snapshot date and a suppression rule.
ExampleEvery one
How often the model's outcome tags match a human-coded sample, reported by outcome and by service.
ExampleAbove an agreed threshold, reviewed quarterly
Published figures below the small-number threshold, or that could identify someone.
ExampleZero
How much the impact team changes the drafts. Very high means it isn't helping; near zero can mean drafts are approved unread.
ExampleStable, and reviewed monthly
Themes reviewed by service leads that led to a recorded decision.
ExampleReported to trustees each quarter
Cross-sector rules are on the UK regulation map.
Reporting is tiered by income: Tier 1 (up to £500,000) summarises main achievements; Tier 2 (to £15 million) must review performance against objectives and explain the charity's impact; Tier 3 adds a review of fundraising performance. Every figure in the impact section carries its source, period and method, and the report says what the data can't show.
Trustees approve the annual report and can't delegate that responsibility. The Commission expects reasonable independent checks before relying on AI output for an important decision. Trustees see the method, the tagging accuracy and the limitations alongside the draft, not just the numbers.
A significant data breach or loss is a reportable serious incident, reported as soon as reasonably possible, and trustees decide whether it is significant. That stays true when a supplier causes it, and the Commission has reminded trustees of this. A personal data breach may also need reporting to the ICO within 72 hours. The incident plan covers the analysis store and the AI suppliers.
Case notes hold special category data such as health, and often concern people in vulnerable circumstances. A DPIA settles the lawful basis and the Article 9 condition before the pilot. Since 5 February 2026 the UK GDPR defines statistical purposes: the output must be aggregate data that is not personal data, and the data must not be used for measures or decisions about a particular person. That is why nothing flows back into individual records.
Pseudonymised notes are still personal data; only properly anonymised outputs fall outside the law. Names, contact details and other identifiers are removed or replaced before analysis, the re-identification key stays with the case management team, and small numbers are suppressed. The ICO consulted on research-specific guidance from August to October 2026, so the rules are reviewed when it is final.
When impact figures and stories are reused in appeals or bids, they must not mislead and need appropriate evidence. Case studies must reflect real situations, say if details were changed, and have the person's clear and informed permission.
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: scheduled extractA dated snapshot, taken through the case management supplier's published API or export. Agree in discovery what the supplier allows. | API Gateway, S3 | API Management, Blob Storage | Kong or your existing API gateway; Ceph or IONOS S3 Object Storage |
| De-identificationOff-the-shelf detection misses local place names, nicknames and rare details. Add rules for them and test on real notes. | Amazon Comprehend PII detection, plus the charity's own rules | Azure Language in Foundry Tools (PII detection), plus the charity's own rules | Open-source PII detection (e.g. Microsoft Presidio), plus the charity's own rules |
| Analysis store and figuresEach published figure is a named, versioned query with its snapshot date and suppression rule. The language model never calculates a number. | Amazon S3 and Amazon Athena | Microsoft Fabric | PostgreSQL on Kubernetes |
| Outcome tagging (predictive ML)A classifier trained on notes coded by caseworkers against the charity's outcome framework. It tags notes, never people, and its accuracy is reported by service. | Amazon SageMaker AI | Azure Machine Learning | MLflow and Kubeflow on Kubernetes |
| Theme discoveryAmazon Comprehend topic modelling closed on 30 April 2026. Themes are prompts for service leads, not findings. | Open-source topic modelling (e.g. BERTopic) on SageMaker AI | Open-source topic modelling (e.g. BERTopic) on Azure Machine Learning | Open-source topic modelling (e.g. BERTopic) on Kubernetes |
| DraftingBoth clouds offer a choice of models, including Claude. Drafts insert approved figures; they don't generate them. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Methods, framework and case studies (retrieval)The outcome framework, definitions, last year's report and consented case studies, so drafts use the charity's own terms. | Bedrock Knowledge Bases | Foundry IQ, built on Azure AI Search | OpenSearch, or PostgreSQL with pgvector |
| Agent runtimeThe original Bedrock Agents closed to new customers on 30 July 2026; new builds use AgentCore. | Amazon Bedrock AgentCore | Foundry Agent Service | An open-source agent framework (e.g. LangGraph) on Kubernetes |
| WorkflowRuns each reporting cycle: snapshot, de-identify, tag, calculate, suppress, draft, review. | 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 |
| Human reviewReviewers see each figure beside its query, its snapshot and the tagging accuracy for that service. | In the impact team's existing reporting tools and documents | In the impact team's existing reporting tools and documents | In the impact team's existing reporting tools and documents |
| GuardrailsBlocks drafts containing a number not in the approved figures, a name, or a quote without a consent record. | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protectionThe re-identification key never leaves the case management side. Analysts see de-identified data only. | IAM Identity Center, KMS, Macie | Entra ID, Key Vault, Purview | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM |
| Monitoring and auditLog every snapshot, model version, query, draft, edit and approval for each report. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency. Commissioners' and funders' contracts may require UK processing. | eu-west-2 (London) | UK South / UK West | Your 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 →
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 case management system, identifiable notes and the re-identification key stay where they are, whether on the charity's own servers or with its supplier. De-identification runs on that side, so identifiable notes never reach the AI.
Uses cloud AI
Outcome tagging, theme discovery, figures and drafting run on de-identified data in a UK cloud region, or on the charity's own GPUs if the trustees decide notes shouldn't leave its estate even in de-identified form.
How they connect
A scheduled, encrypted transfer over a site-to-site VPN, or AWS Direct Connect or Azure ExpressRoute for larger charities. Only de-identified batches cross the link, and results come back as figures and drafts, never as changes to records.
If the link drops
Nothing a service user depends on stops: casework carries on as normal. Reporting is a batch job with deadlines weeks away, so the next snapshot simply runs when the link returns.
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.
Report drafting assistant
Where a person signs offThe head of impact checks every figure and draft. Trustees approve the annual report, and the named grant holder approves each funder report. The agent's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
The case management system stays the record. A scheduled job takes a dated snapshot through the supplier's published API or export, de-identifies it on the charity's side, and sends only the de-identified batch onward. Nothing is ever written back.
Figures have a lineage. Each one names its query, snapshot date, outcome framework version, model version and suppression rule, so any number in a published report can be reproduced or corrected later.
The AI is optional. If it's paused, the impact team runs the same queries on structured fields and writes the narrative themselves, exactly as before, with less of the notes' detail.
Map today's reporting, agree the outcome framework, audit note quality and consent records, and confirm what the case management supplier allows.
Gate · must be true to passOutcome framework agreed, DPIA started, extract route confirmed, suppression rule set.
One service, last year's notes: de-identify, tag, calculate and draft, then compare with a human-coded sample and last year's reports.
Gate · must be true to passDe-identification tested on real notes, tagging accuracy above the agreed threshold, every pilot figure reproducible, DPO sign-off.
Live for that service's funder reports, with monitoring, support and a rollback plan.
Gate · must be true to passMonitoring, support model, rollback plan and impact team training in place.
Add services one at a time, then the trustees' annual report impact section for the year.
Gate · must be true to passEach new service passes the same pilot gate; trustees see the method before the first annual report.
| Activity | Director of services | Head of impact | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | DPO | Trustees |
|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Informed | Consulted | Informed |
| Outcome framework and suppression rules | Consulted | Accountable and Responsible | Consulted | Informed | Consulted | Consulted | Consulted |
| Target architecture | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Informed |
| DPIA and privacy notice | Consulted | Consulted | Consulted | Informed | Informed | Accountable and Responsible | Informed |
| Build, de-identification and integration | Informed | Consulted | Accountable | Responsible | Responsible | Consulted | Informed |
| Pilot and accuracy testing | Consulted | Accountable | Consulted | Responsible | Responsible | Consulted | Informed |
| Trustees' report approval | Responsible | Responsible | Informed | Informed | Informed | Consulted | Accountable |
| Run, monitor and improve | Consulted | Accountable | Consulted | Consulted | Responsible | Consulted | Informed |
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 extract from your case management system, de-identification, the figures and their lineage, the models, the drafting and the gates. I'd report to your director of services and make sure trustees can see how every figure was made.
Option B
If your partner or your own team already has an architect, I'd take on the data pipeline, de-identification and the integration with your case management system under their lead, and give them a second senior opinion on data protection and the model testing.
Goes wrong
The language model writes numbers, and one is wrong in the trustees' report.
How the design avoids it
Figures come only from versioned queries. Guardrails block any number that isn't in the approved set.
Goes wrong
De-identification misses local details, and a small service's story identifies someone.
How the design avoids it
Charity-specific rules on top of standard detection, testing on real notes, and a small-number rule before anything leaves the team.
Goes wrong
Tagging works for one service and quietly fails for another.
How the design avoids it
Accuracy reported by service against a human-coded sample, and services added one at a time.
Goes wrong
Outcome data drifts into judging caseworkers.
How the design avoids it
A trustee-approved rule, no staff or team scoring in the design, and aggregate outputs only.
Goes wrong
Thin notes produce confident-looking figures.
How the design avoids it
Report coverage alongside every figure, and say plainly in the report what the data can't show.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Run the pipeline in shadow on this year's notes now, and agree the outcome measures with trustees before the year closes.
Keep de-identification and suppression rules as configuration, so they can be tightened without a rebuild.
Keep the records that show outputs are aggregate and never used for decisions about individuals.
Expect your IT, case management and AI suppliers' duties to change, and review contracts and the breach plan with them.
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.