Inference Public

Cloud foundations

A landing zone needs owners as well as accounts.

A new account is useful when a team can deploy, investigate a fault and get a decision. The landing zone has to make all three possible.

Two engineers reviewing a large unlabelled plan at a worktable.

The account arrives on Tuesday. By Thursday the application team has discovered that it cannot resolve an internal hostname, nobody knows who approves the route, and the logging dashboard belongs to an engineer on leave. Account provisioning took minutes. Joining the platform is taking rather longer.

That is an illustrative situation, not a client story. It captures a distinction worth making early: creating an AWS account and enabling a team to operate a service are different outcomes. A landing zone should establish both the technical boundaries and the working arrangements inside them.

Start with the team's first week

Ask what an incoming team must do before its first useful change reaches a non-production environment. It needs an account with a purpose, access through the organisation's identity arrangements, a deployment identity, a network route, a way to obtain secrets and somewhere to investigate failures.

Turn that into an onboarding contract. For each item, state what is supplied, what the team must configure, who supports it and how a change is requested. A contract that says only 'networking provided' leaves most of the difficult questions unanswered. Which destinations? Which protocols? Who controls DNS? Who investigates a failed connection that crosses two accounts?

Do the same for costs. A service owner should know which account costs are theirs, which costs are shared and how the shared part will be explained. A centrally paid Transit Gateway or log archive still needs an owner and a budget decision.

Separate the boundaries from the organisation chart

Account boundaries, organisational units and team names solve related but different problems. Accounts can separate ownership and workloads. OUs group accounts for policy. Team and service metadata explain who uses them. A reorganisation should not automatically require every technical boundary to move.

AWS describes service control policies as limits on permissions, not grants of access. A permissive SCP does not create a deployment role. An appropriate role still needs its own permissions and trust relationships. Conversely, the management account is outside SCP restrictions on member-account identities, so it needs deliberately controlled use. These details matter when an apparently simple onboarding request becomes a request for broader access.

See the AWS SCP guidance and our separate OU design article for those mechanics. The operating question here is who owns the decision when a boundary has to change.

Write down responsibilities at the point where work crosses teams

Consider this example allocation. It is a starting point to agree, not a universal operating model.

WorkFoundation teamApplication teamService or security owner
New accountProvision the agreed baselineSupply purpose and technical contactAccept ownership and budget
ConnectivityMaintain shared routes and DNS patternsDeclare required flowsDecide exceptions to approved patterns
Deployment accessSupply a bounded identity patternMaintain the service workflowApprove privileged departures
Logs and alertsOperate shared collectionDefine useful service signalsAgree retention and escalation
Account retirementExecute the closure procedureIdentify data and dependenciesAuthorise retirement and retention decisions

The important column is often the last one. Engineers can implement an exception; that does not mean they hold authority to accept its risk or commit another team's budget.

Make the exception a lifecycle

A supplier needs temporary access to investigate an inherited system. Record the purpose, boundary, approver, compensating measures, expiry and the person who will remove it. Decide what happens if the investigation is incomplete when the expiry arrives. Automatic expiry, an approved extension and an alert for manual action are different mechanisms; specify which one is actually implemented.

An exceptions register is useful only if someone reviews it. Reproduce one exception from request to removal during acceptance. Otherwise the register may document a promise rather than a functioning procedure.

Avoid treating an exceptions OU as a waiting room with no exit. Moving an account changes the policies that apply to it. Review the effective result, not just the destination's name.

Accept the foundation by using it

Choose a representative service and an incoming team who did not build the landing zone. Give them the supplied onboarding instructions. Watch them carry out a bounded exercise:

  1. Obtain approved access and demonstrate that an unrelated account remains inaccessible.
  2. Provision or request the agreed environment through the normal route.
  3. Deploy a small change using the deployment identity.
  4. Reach an approved dependency and show that an unapproved route is blocked.
  5. Find the resulting application and platform logs.
  6. Route a question or exception to the person authorised to decide it.

Record the revision, environment, identity and result. Where the exercise depends on a platform engineer quietly fixing something in the background, record that dependency. It is a useful finding, not an embarrassment to erase from the test.

The deliverable is a foundation the next team can use

Infrastructure as code is part of the handover. So are the onboarding contract, responsibility table, known exceptions and the evidence from an ordinary team using the platform. Together they explain how an account becomes a place in which work can happen.

If you are assessing an existing estate, start with one recent onboarding. Follow the waits, approvals and workarounds. That small exercise usually tells you more about the operating model than another tour of the account list.

Read about infrastructure engineering or discuss your AWS foundations.

Further reading

AWS: service control policies

AWS: recommended OUs and accounts

All insights

Start a conversation

Let’s talk about what you need to build.

Bring the service, the constraint or the engineering question. We’ll work through a sensible starting point.

Discuss a project