The 99.95% promise and the 99.9% supplier

The arithmetic first. Availability of a serial chain is approximately the product of its parts:

0.9995 (checkout app) × 0.999 (ISP) × 0.9995 (IaaS) × 0.999 (payment gateway) ≈ 0.9970

That is about 99.70% in the best case, assuming independence and every supplier hitting its target exactly. The internal promise of 99.95% is therefore not achievable through good supplier performance at all; only redundancy or a change to what is measured and promised can deliver it.

Why nobody breached. Each contract is scoped to the supplier's own boundary, uses its own availability definition (different exclusions for planned maintenance, different sampling points), and is reported on its own cycle. A supplier can be fully compliant while contributing an outage to a downstream service. The 99.86% figure comes from a synthetic transaction that measures the composite, and no supplier owns the composite.

Design faults visible in the chain.

• The business commitment was set without a service level requirements exercise and without an availability model.
• Supplier contracts were not negotiated backwards from the platform commitment, so there is no underpinning contract chain.
• Measurement points and definitions were never harmonised, so the composite number cannot be decomposed into supplier contributions.
• Credits exist only on the internal SLA side; there is no cascade to suppliers, so IT absorbs the financial consequence of supplier shortfalls.
• The synthetic transaction covers checkout only, so the ISP dependency in the admin console may be invisible.

Conclusion. This is a defect in the design of the agreement chain, not a supplier performance failure, and Finance's question is answerable: the organisation has contractually accepted a risk from which it is contractually insulated upstream.

Both - but renegotiation first, because the number was never validated as a requirement.

1. Test whether 99.95% is the right requirement. An SLA should express a service level requirement derived from business need, not a number chosen for its appearance. The question to put to Sales is what an hour of unavailability actually costs - abandoned baskets, SLA credits, contact-centre load - against the cost of the resilience needed to reach 99.95%. If the business loss per hour is small, the current target buys expense for no benefit; if it is large, the target is too low for a single-region design.

2. Make the promise one the architecture can keep. A 99.95% end-to-end commitment on a serial chain requires either removing single points of failure (multi-region IaaS, redundant ISP paths, a secondary payment provider with automatic failover) or redefining scope - for example committing to checkout availability while treating payment as a separately described dependency with its own commitments.

3. Fix the definitions before the date is agreed. Availability needs a written measurement specification: measurement point, sampling method, treatment of planned maintenance, exclusions, the reporting period (calendar month versus rolling 28 days) and who validates it. Without this, next quarter's dispute is already scheduled.

4. Align the credits. The 5% per 0.1% credit is generous relative to most retail margins and is not mirrored upstream. Credits should reflect actual business impact and be cascaded into supplier contracts so the party causing a shortfall carries it.

5. Sequence. Renegotiate the internal SLA to a defensible target with clear exclusions for the renewal date, run the architecture options in parallel, and change supplier terms at their own renewal points, using the intervening period to instrument the composite properly.

Build the chain top-down, then verify it bottom-up.

Step 1 - Service level requirements (SLRs). Run workshops with Sales, fulfilment and customer service to capture what the order platform must do for the business: successful checkout rate, checkout latency, order visibility in the admin console. SLRs are the input; nothing downstream is agreed until they exist.

Step 2 - The SLA with Sales. One document, one availability definition, one measurement specification and one reporting period. It should state the service boundary explicitly - what is inside the commitment and which dependencies (payment gateway, ISP) are described but not warranted - plus credits and the review cycle.

Step 3 - OLAs between internal teams. The platform team depends on the network and database teams. Each OLA must use the same definitions, the same measurement tooling and the same reporting cadence as the SLA, otherwise internal performance cannot be added up to the customer-facing number.

Step 4 - Underpinning contracts (UCs). Renegotiate the payment, IaaS and ISP contracts so that each supplier reports against a compatible definition, provides raw availability data rather than a summary, and carries credit terms that can be passed through. Where a supplier will not accept an end-to-end target, record the residual risk explicitly.

Step 5 - A composite availability model. A documented calculation showing how each component contributes to the platform figure, with the measurement point for each, reviewed at every service review. This is what makes Finance's question answerable next quarter: which component consumed which part of the 0.14% shortfall.

Step 6 - Governance. A quarterly service review with Sales, monthly internal reviews, and a rule that any change to a component target is assessed against the composite before signature.