Business case and benefits
- Customer service director
- A · Accountable
- Product owner
- R · Responsible
- Solutions architect
- C · Consulted
- Technical PM
- C · Consulted
- Engineers
- I · Informed
- Legal & DPO
- C · Consulted
- Security & PCI
- I · Informed
Order, delivery and returns questions answered in seconds, on chat, email and phone, from your real systems. Refunds within set limits handled end to end, and people dealing with every exception, complaint and vulnerable customer.
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 multi-channel retailer takes orders in stores, on its own website and app, and through online marketplaces. Each order touches the order management system (OMS), the warehouse management system (WMS), a carrier's tracking, the payment service provider and, for click-and-collect and store returns, the EPOS. When a customer asks where their parcel is, an adviser opens most of those systems to answer one question.
In a typical retailer of this size, around half of all contacts are order status, delivery problems and returns. They are simple to answer once you can see the data, but they take an adviser four to six minutes each, and they arrive in waves: the Monday after a bank holiday, the week of a carrier delay, and the whole of peak. Queues grow, customers chase on a second channel, and the same order becomes three contacts.
The rules have also sharpened. Since April 2025 the CMA can fine a business directly for breaking consumer law, up to 10% of global turnover. In March 2026 it published guidance on using AI agents with customers: the business is responsible for what its agent says and does, as it would be for an employee. An agent that misstates a customer's right to a refund is not a small error at this volume.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
Contacts the agent closes where the customer doesn't come back about the same order within seven days. Better than 'containment', which can hide customers who simply gave up.
Example60–70% of order and returns contacts
How long it takes to reach an adviser after asking for one, or after the agent hands over.
ExampleUnder 2 minutes, at peak too
A weekly audit sample: was each refund due, the right amount, to the right payment, on time?
ExampleAbove 99%, with every error traced
A test set of consumer-law questions (cancellation, faulty goods, lost parcels), run before every change and monthly.
ExampleEvery test passes before release
Advisers rate whether the summary let them carry on without re-asking.
ExampleRated useful in 9 out of 10 handovers
Complaints that mention the bot, split out for customers flagged as vulnerable.
ExampleFalling, and reviewed weekly
Cross-sector rules are on the UK regulation map.
Since 6 April 2025 the CMA can decide breaches itself and fine up to 10% of global turnover. The agent must not mislead by what it says or leaves out, about prices, delivery or rights. Its answers come only from approved, versioned policies, and legal signs off the consumer-law test set.
The retailer is responsible for its agent as for an employee, including where a supplier built it. Tell customers they are talking to an AI, test before go-live, monitor regularly, keep a human in the loop, and make sure refund decisions reflect the right time limits and any extended returns policy.
Goods must be of satisfactory quality, fit for purpose and as described. A customer can reject faulty goods within 30 days for a full refund, and refunds are due within 14 days of agreeing them. Goods stay at the retailer's risk until delivered, so a lost parcel is never the customer's problem to chase.
Online, phone and marketplace orders carry a 14-day cancellation period from delivery. The refund, including standard delivery, is due within 14 days of getting the goods back or proof of return. Exceptions, such as personalised or unsealed hygiene items, are a person's call, never the agent's.
The agent never sees, hears or stores a card number. Refunds use the payment service provider's token for the original transaction. Phone payments use keypad (DTMF) masking, and transcripts and logs are scanned and redacted. This keeps the AI out of PCI scope rather than adding to it.
Chats, emails and call recordings are personal data, so a DPIA, a clear retention period and a contract that stops providers training on your data. A refused refund is a significant decision, so a person makes it. The agent sends service messages only, never marketing, and PECR fines now match UK GDPR levels.
Reasonable adjustments apply to service channels. Chat meets WCAG 2.2 AA, the voice agent can be interrupted and slowed down, and every channel offers a person, including a phone number, without the customer having to argue for one.
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 |
|---|---|---|---|
| Channels and routingKeep the platform your advisers already use if it can hand a conversation to an AI and back. Changing it is a separate project. | Amazon Connect (chat, voice, email), or your existing contact-centre platform | Azure Communication Services, or your existing contact-centre platform | Your existing contact-centre platform |
| SpeechTest on real calls: regional accents, mobile lines, background noise in a car or a shop. | Amazon Transcribe and Amazon Polly | Azure Speech in Foundry Tools | Open-weight speech models (e.g. Whisper) and an open-weight text-to-speech model |
| Reasoning and repliesBoth clouds offer a choice of models, including Claude. Pick for accuracy on your test set and speed on the phone, not benchmark scores. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Policies and help content (retrieval)Only approved, versioned content: returns policy, delivery options, statutory rights wording. Each answer records which version it used. | 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 |
| Returns and refund workflowLimits live here, not in the prompt: value, number of refunds per customer per month, and original payment method only. | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| Middleware and eventsOrder dispatched, delivery failed, return scanned in, refund issued: each one an event the agent can act on. | Amazon EventBridge, SQS | Event Grid, Service Bus | Apache Kafka or RabbitMQ |
| Link to OMS, WMS and EPOSRead-mostly. The only writes are booking a return, adding a note and asking the refund service to pay. | API Gateway, plus your existing integration platform if you have one | API Management, plus your existing integration platform if you have one | Kong or your existing API gateway, plus your existing integration platform if you have one |
| PaymentsNo card numbers anywhere in the AI. Phone payments use DTMF masking provided by the contact-centre platform or the payment provider. | Your payment service provider's refund API, by token | Your payment service provider's refund API, by token | Your payment service provider's refund API, by token |
| Human handoverThe adviser sees the conversation, what the agent checked and why it handed over. | In the advisers' existing contact-centre desktop | In the advisers' existing contact-centre desktop | In the advisers' existing contact-centre desktop |
| GuardrailsBlock card numbers, promises the systems can't back, and advice outside the agent's job. | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protectionCustomers are verified by the platform (signed-in account, or order number plus postcode) before the agent sees any order. | 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 answer, the policy version, every system call, every refund and every handover reason. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency. | 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 OMS, WMS, EPOS and merchandising systems stay where they are, whether in your own data centre, in stores or with their suppliers. Card data stays with the payment service provider. The agent never becomes a system of record.
Uses cloud AI
The agent, the policy knowledge, speech and the refund workflow run in AWS or Azure, or on your own GPUs with open-weight models if conversation data must stay in your estate. Only the order fields needed for an answer are sent.
How they connect
A private link (AWS Direct Connect or Azure ExpressRoute) to the API layer in front of the OMS and WMS. Order and return events flow out to the queue, and refund requests come back through the same API layer.
If the link drops
The agent tells the customer it can't see their order right now and offers an adviser or a callback. Contacts route straight to advisers as before the AI. Refund requests wait in the queue and are paid once, never twice, 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.
Customer service agent
Where a person signs offA person makes every refusal, every exception to policy and every refund over the limit, and handles every complaint. The refund limits and the agent's permissions are enforced by the platform's identity and policy controls and by the refund service, not by instructions in a prompt.
Adviser assistant
Where a person signs offThe adviser sends every reply and makes every decision. Drafts are marked as drafts until they do.
The OMS is the single view of the order. Marketplace orders usually reach it through a channel integration, and store purchases, click-and-collect and store returns through the EPOS. The agent reads the OMS, not each channel, so a store return and an online refund can't both be paid for the same item.
Every refund request carries the order line and return ID as an idempotency key. If the payment provider times out and the workflow retries, the customer is refunded once. The refund ledger in the OMS is checked before every payment.
Events decouple the agent from the systems of record. If the WMS or a carrier feed is slow, the agent says so honestly instead of guessing, and peak volumes queue rather than overload the OMS.
The agent can be switched off by channel without stopping the business. Paused, contacts go straight to advisers exactly as they did before.
Measure contact reasons and handling times, map the OMS, WMS, EPOS and payment APIs, and agree refund limits and the consumer-law test set.
Gate · must be true to passBaseline agreed, API access confirmed, DPIA and PCI scoping done, legal involved.
Order status on web chat first, in shadow mode and then live for a share of customers. Returns and refunds follow once the test set passes.
Gate · must be true to passRights test set passing, refund audit clean, time to a person within target, legal and customer service sign-off.
All chat traffic, then email, with monitoring, peak load testing and a rollback switch per channel. Live by the end of September.
Gate · must be true to passPeak load tested, rollback rehearsed, advisers trained, monitoring in place.
Phone, then marketplace messages, then other contact reasons one at a time.
Gate · must be true to passEach new channel and contact reason passes the same pilot gate.
| Activity | Customer service director | Product owner | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | Legal & DPO | Security & PCII can fill this |
|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Informed | Consulted | Informed |
| Target architecture | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted |
| Consumer-law test set and DPIA | Accountable | Consulted | Consulted | Informed | Informed | Responsible | Consulted |
| PCI scoping and payment integration | Informed | Consulted | Accountable | Consulted | Responsible | Consulted | Responsible |
| Build and integration with OMS, WMS, EPOS | Informed | Consulted | Accountable | Responsible | Responsible | Informed | Consulted |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted |
| Run, monitor and improve | Informed | Accountable | Consulted | Consulted | Responsible | 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: how the agent reads your OMS, WMS and EPOS, the refund service and its limits, the PCI boundary, and the gates. I'd report to your programme board and work closely with your customer service and legal leads.
Option B
If your partner or your own team already has an architect, I'd take on the integration with the order, warehouse, store and payment systems under their lead, and give them a second senior opinion on the agent's permissions, PCI scope and peak readiness.
Goes wrong
The agent tells a customer to chase the courier, or says they're out of time when they aren't.
How the design avoids it
Answers come only from approved policy content, and a consumer-law test set signed off by legal runs before every release.
Goes wrong
Customers can't reach a person, and complaints move from the queue to social media.
How the design avoids it
A person is offered up front and on request, time to a person is measured, and the agent hands over after two failed attempts.
Goes wrong
Refund limits are written in the prompt, so a determined customer talks the agent past them.
How the design avoids it
Limits are enforced by the refund service, and the agent can only ask it to pay.
Goes wrong
Card numbers turn up in chat transcripts and call recordings.
How the design avoids it
DTMF masking on the phone, redaction on every channel, and the AI kept out of PCI scope.
Goes wrong
It goes live in November and meets peak on day one.
How the design avoids it
Live by the end of September, load tested at peak volumes, with a change freeze through to January.
Changes already announced or under way that affect this design. More in the UK horizon scan.
It covers subscriptions taken out from that date, with new rules on pre-contract information, reminders and cancellation. If you sell a delivery pass or a subscribe-and-save option, the agent must handle cancellation simply and without obstruction. Draft regulations and guidance were published in October 2026.
Regulations in force since May 2026 require the ICO to produce a statutory code. Keep refund decisions, explanations and human review records in a form you could show the ICO.
The CMA, FCA, ICO and Ofcom published a joint statement on agentic AI through the DRCF in October 2026. Customers' own AI agents will start contacting your service. Decide now how they are verified, and apply the same limits as to a person.
Medium and large managed service providers come into regulation, with faster incident reporting. Expect your contact-centre, cloud and AI suppliers to face new duties, and your contracts and supplier assurance to need updating.
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.