Business case and benefits
- Commercial director
- A · Accountable
- Head of purchasing
- R · Responsible
- Solutions architect
- C · Consulted
- Technical PM
- C · Consulted
- Engineers
- I · Informed
- Finance
- C · Consulted
- Security
- I · Informed
Every order line goes to the best-priced supplier that can actually deliver for that account, at the negotiated price, and no agreed order slips through. Buyers decide every exception.
At a glance
An illustrative scenario. Linked to a real, anonymised case study below; the numbers here are still typical examples. The cloud services and UK rules were checked on 9 November 2026.
Orders arrive all day: EDI from the larger accounts, the trade web portal, emails, phone calls and orders that reps agree on site. Purchasing then has to buy each line from a supplier. The owner has negotiated prices, rebates and delivery terms with each supplier, sometimes per account, but at the point of ordering, staff tend to use the supplier they know, the one that answered the phone, or the one they used last time.
The result is money left on the table in a way nobody sees. Lines are bought above the agreed price. Small orders go to one supplier when they'd have hit a free-delivery threshold with another. Some products are only stocked by a handful of suppliers, so staff have to be allowed to source freely, and the rules get bent for everything else too. Worst of all, orders a rep agreed with a customer sometimes never get placed. The customer finds out when nothing arrives.
Then come the confirmations. Suppliers reply by EDI, email, PDF or their own portal, with short shipments, substitutions, price changes and new delivery dates. Someone has to read every one, compare it with what was ordered and decide what to do, and that's where most buyer time goes. Paying suppliers on time also matters more each year, with the government's late payment reforms now going through Parliament.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
Share of order lines placed with the right supplier at the contract price for that account.
ExampleAbove 98%
Orders a rep agreed that never reached a supplier. The problem nobody measured before.
ExampleClose to zero, every one investigated
Whether the routing picks suppliers that actually deliver, not just the cheapest.
ExampleUp, measured against last year
How close the predicted delivery date is to the actual one, by supplier.
ExampleWithin a day for most lines
Time moved from reading confirmations to handling real exceptions and suppliers.
ExampleDown by half
Whether faster matching helps the business pay its own suppliers on time.
ExampleIn line with your payment policy, reported half-yearly
Cross-sector rules are on the UK regulation map.
Late payers owe statutory interest of 8% over the Bank of England base rate, plus fixed compensation. A wholesaler is both a payer and a payee. Faster matching of confirmations, deliveries and invoices means disputes are raised early and suppliers are paid on time, so the design tracks invoice, dispute and payment dates for every line.
The government's March 2026 response confirmed plans for a 60-day cap on most business payment terms (no earlier than 2027), statutory interest that can't be contracted out of, a time limit for raising invoice disputes, and powers for the Small Business Commissioner to investigate and fine. It isn't law yet. Large companies already report payment performance twice a year under the 2017 reporting regulations, so the same data feeds that report.
Comparing supplier prices to buy well is normal. Passing one supplier's prices, terms or volumes to another is not, and nor is acting as a channel for suppliers to learn about each other. Supplier prices and terms are confidential to the buying team, and the agent can't include them in any message to a supplier. Get your own competition advice on how rebates and supplier data are used.
Since 24 February 2025, public contracts carry an implied 30-day payment term, and it flows down to subcontracts that are wholly or substantially for that public contract. Public-sector accounts often buy on framework prices too. The routing rules hold framework prices and terms per account, and supplier lines for those accounts are flagged for 30-day payment.
E-invoicing will run on a Peppol model, with a roadmap published on 28 October 2026. Purchase orders, confirmations and invoices are held as structured data from day one, so moving supplier invoices onto Peppol is a new channel, not a new process. Existing EDI keeps working alongside it.
There's little personal data here: buyer and rep names, account contacts, and sole-trader suppliers. Standard controls apply: collect only what's needed, a clear retention period, and a contract that stops providers training their models on your data. Routing and prediction are about suppliers and products, not people.
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: EDI, portal, emailKeep the EDI translator and trading-partner setup you have. UK suppliers still use EDIFACT, and some older ones TRADACOMS. | AWS Transfer Family (AS2), API Gateway, Amazon SES, S3 | Logic Apps integration account (EDIFACT, X12, AS2), API Management, Blob Storage | Your existing EDI translator and VAN; Kong; your mail server; Ceph or IONOS S3 Object Storage |
| Reading confirmations and supplier emailsTest on real supplier PDFs and email replies, including the ones typed on a phone. | Bedrock Data Automation | Azure Content Understanding (includes Document Intelligence) | Open-source parsing (e.g. Docling) plus an open-weight vision model |
| Routing rules and price bookRules, not a model, pick the supplier, so every choice can be explained line by line: price, rebate, MOQ, delivery threshold, predicted lead time. | AWS Lambda or ECS, Aurora PostgreSQL | Azure Functions or Container Apps, Azure Database for PostgreSQL | A rules service and PostgreSQL on Kubernetes |
| Availability and lead-time predictionTrained on your own history of orders, confirmations and deliveries per supplier and product. Amazon Forecast is closed to new customers, so use SageMaker AI. | Amazon SageMaker AI | Azure Machine Learning | MLflow and Kubeflow on Kubernetes |
| Reasoning and draftingUsed for reading and summarising confirmations and drafting chasers, not for choosing suppliers or prices. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| 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 |
| Workflow and order wavesWaves are timed to each supplier's cut-off, so a late order still makes today's delivery where it can. | AWS Step Functions, EventBridge Scheduler | Logic Apps or Durable Functions | Temporal, or n8n |
| Middleware and events | Amazon EventBridge, SQS | Event Grid, Service Bus | Apache Kafka or RabbitMQ |
| Link to the ERP and WMSReuse the integration layer you have. The ERP stays the record of every order, PO and invoice. | Through their published APIs, behind API Gateway | Through their published APIs, behind API Management | Through their published APIs, behind Kong or your existing gateway |
| Human reviewExceptions arrive as tasks with the reason and the options, where buyers already work. | In the buyers' existing ERP purchasing screens | In the buyers' existing ERP purchasing screens | In the buyers' existing ERP purchasing screens |
| GuardrailsBlocks any supplier message that contains another supplier's name, price or terms. | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protectionSupplier prices and terms readable only by purchasing and the routing service. | IAM Identity Center, KMS | Entra ID, Key Vault | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault |
| Monitoring and auditLog every routing decision with its reasons, every exception and who decided it. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| Data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency. Supplier terms are commercially sensitive, so check contracts before choosing. | 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 ERP, the WMS, the EDI translator and the price book stay where they are. Many wholesalers run these on their own servers, and nothing in this design needs them to move.
Uses cloud AI
The lead-time models, reading supplier emails and PDFs, and the confirmation agent can run in AWS or Azure, or on your own GPUs if supplier terms must not leave your estate. The routing rules can run on either side; keeping them next to the ERP keeps them fast.
How they connect
A private link (AWS Direct Connect or Azure ExpressRoute) and the API layer. Lead-time predictions are pushed to the routing service as a table refreshed through the day, so routing never waits on a cloud call.
If the link drops
Routing carries on with the last predictions and the rules on-prem. Confirmations queue and are read when the link returns, and buyers handle urgent ones by hand, as they do today.
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.
Supplier confirmation agent
Where a person signs offA buyer decides every exception: price changes, substitutions, re-routing and split deliveries. The agent's permissions are enforced by the platform's identity and policy controls and the ERP's roles, not by instructions in a prompt.
Order capture check
Where a person signs offSales ops confirm with the rep or customer and place the order. Nothing is ordered on a customer's behalf without a person checking it was really agreed.
The ERP stays the record of every customer order, purchase order and invoice, and the WMS keeps receiving goods as it does today. The routing service reads orders through the ERP's API, decides the supplier, and writes the purchase order back. It never writes straight to the database.
EDI stays where it is. Customer and supplier messages keep flowing through your existing translator and trading-partner setup, and the AI only reads the translated messages from a queue. Email and PDF confirmations join the same queue once they're read, so every confirmation is handled the same way.
Every order line carries an ID from the moment it arrives. If a supplier's system or the ERP is down, lines wait on the queue for the next wave, and a retry never creates a duplicate purchase order.
If the routing service is paused, buyers fall back to the price book in the ERP, as they did before. Orders keep moving.
Map today's ordering and buying, measure how often lines are bought off-contract and how many agreed orders go missing, collect a year of order and delivery history, and check the price book is complete.
Gate · must be true to passBaseline agreed, price book owner named, ERP and EDI integration route confirmed.
One supplier group in shadow mode: the routing service recommends, buyers order as normal, and the two are compared. The order capture check runs from day one, because it only raises tasks.
Gate · must be true to passRouting matches or beats buyers on price without hurting on-time delivery, prediction accuracy within agreed limits, and buyer sign-off.
Routing goes live for the pilot group, then the confirmation agent, with monitoring, support and a rollback to the ERP price book.
Gate · must be true to passMonitoring, support model, rollback plan and buyer training in place.
Add supplier groups in waves, single-source products last, then structured supplier invoices ready for Peppol.
Gate · must be true to passEach new supplier group passes the same pilot gate.
| Activity | Commercial director | Head of purchasing | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | Finance | SecurityI can fill this |
|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Informed | Consulted | Informed |
| Target architecture, ERP and EDI integration | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted |
| Price book and routing rules | Accountable | Responsible | Consulted | Informed | Consulted | Consulted | Informed |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Informed | Consulted |
| Model training and accuracy testing | Informed | Accountable | Responsible | Consulted | Responsible | Informed | Informed |
| 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: order intake, the waves, the routing rules, the ERP and EDI integration, and the gates. I've built this before. For a wholesaler, I designed and built a sourcing platform that took in high-volume orders, split them into waves and routed each product to the lowest-cost supplier for that account, and it captured orders that used to slip away. The prediction and agent layers here build on that same core. Platforms I've authored have moved billions of pounds of goods and services, so I know how much rests on the order flow. I'd report to your programme board.
Option B
If your partner or your own team already has an architect, I'd take on the routing engine, the lead-time models and the ERP and EDI integration under their lead, and give them a second senior opinion on the rest, especially where the agent is and isn't allowed to act with suppliers.
Goes wrong
The cheapest supplier always wins, and on-time delivery quietly gets worse.
How the design avoids it
Predicted availability and lead time are part of every routing choice, and on-time, in-full is a pilot gate alongside price.
Goes wrong
The price book is out of date, so the system enforces the wrong prices perfectly.
How the design avoids it
A named price book owner, change control, and price differences on confirmations fed back as price book fixes.
Goes wrong
A model picks suppliers, and nobody can explain why a line went where it did.
How the design avoids it
Rules make the choice and record the reasons. Machine learning only supplies the predictions the rules use.
Goes wrong
The agent mentions a competitor's price to push a supplier, and creates a competition law problem.
How the design avoids it
The agent can't negotiate, supplier terms are walled off, and guardrails block other suppliers' names and prices in any message.
Goes wrong
Buyers don't trust it and override everything, so nothing changes.
How the design avoids it
Shadow mode first, every routing reason visible, overrides allowed with a reason, and the reasons used to improve the rules.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Record invoice, dispute and payment dates against every supplier line now, so you can prove you pay on time and challenge customers who don't.
Hold purchase orders, confirmations and invoices as structured data, and ask your ERP and EDI providers for their Peppol plans.
Expect your EDI, hosting and IT providers to face new duties. Update supplier assurance and contracts, since a stopped order flow stops the business.
Record which encryption the EDI links, APIs and stored price data use, and keep it easy to change.
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.