Home › AWS › AWS data services › Organizations and Control Tower
AWS data services · Security, identity and network · about 12 min read
AWS Organizations, AWS Control Tower and service control policies (SCPs)
The landing zone: many AWS accounts under one organization, with guardrail policies no account can override and a central log and security account.
In plain words
What it is
An AWS account is a sealed box: its own resources, its own bill, its own users. Big companies do not put everything in one account, because one mistake or one stolen password would then reach everything. They use many accounts, and they need a way to manage them together. That is what these services do.
AWS Organizations groups accounts under one management account. Accounts sit in organizational units (OUs), folders such as Security, Workloads and Sandbox, and the organization gets one combined bill. Organizations also holds policies that apply to whole OUs. The best known is the service control policy (SCP): a ceiling on what anyone in an account may do, even the account's own administrator. An SCP never grants anything; it only takes away.
AWS Control Tower builds and runs a ready-made landing zone on top of Organizations. A landing zone is the starting set-up every new account lands in: a log archive account that collects CloudTrail and AWS Config records, an audit or security account, single sign-on through IAM Identity Center, and a library of controls (preventive SCPs, detective AWS Config rules and proactive CloudFormation hooks). Its Account Factory creates new accounts that already follow the rules.
A concrete example: a retailer's data platform gets three accounts, data-dev, data-preprod and data-prod, under a Workloads OU. An SCP on that OU says no one may create resources outside Mumbai and Hyderabad, no one may stop CloudTrail, and no one may make an S3 bucket public. A developer with full administrator rights in data-dev still cannot break those rules, because the ceiling sits above the account.
Organizations is the company's building with separate locked floors, SCPs are the fire code no tenant can override, and Control Tower is the builder who fits every new floor to code before anyone moves in.
Big picture
Where it fits
The lit lane is where Organizations and Control Tower works. Boxes on the left feed it; boxes on the right use what it produces.
This is the bottom layer of the reference platform, under the network, storage and Databricks. The foundation's Terraform is applied in order, and the first folder, 1-org, sets up the accounts, the SCP guardrails and CloudTrail to the log account. Every later layer, from the VPC per environment to the buckets and KMS keys per catalog, lands in an account this layer created.
The reference design has three workload accounts, dev, pre-prod and prod, each with its own VPC and its own Databricks workspace, plus central security and log archive accounts and a network account that owns the shared Transit Gateway. Pre-prod reads prod read-only. The Databricks account itself is separate from AWS accounts: one Databricks account in Mumbai with one Unity Catalog serves the three workspaces, and each workspace is bound only to its own catalogs.
The log archive account is where the audit story ends. CloudTrail from every account, KMS key-use events and VPC flow logs land there, are kept in a bucket nobody in the workload accounts can delete, and are joined with Databricks audit system tables into the quarterly access report. The security account is where GuardDuty, Security Hub and Macie findings from the landing buckets are gathered.
If the company already runs Control Tower, the platform asks for three accounts from Account Factory and inherits the guardrails. If not, the platform adds a minimal landing zone first rather than starting in a single account and moving later, which is far harder than starting right.
How it works
The ideas you need
| Idea | What it means |
|---|---|
| Management account | The account that owns the organization and pays the bill. SCPs do not apply to it, so nothing runs there except organization administration. |
| Organizational unit (OU) | A folder of accounts, nestable up to five levels. Policies attached to an OU apply to every account inside it. |
| Service control policy (SCP) | A ceiling on what IAM users and roles in member accounts may do, including the root user. It removes permissions; it never grants them. |
| Resource control policy (RCP) | A ceiling on what may be done to resources such as S3 buckets, KMS keys and Secrets Manager secrets, whoever asks, which helps stop data leaving to outside accounts. |
| Declarative policy | A policy that sets a service's configuration across accounts, for example blocking public access settings or requiring IMDSv2 on EC2, and keeps it set as new features arrive. |
| Landing zone | The baseline every account starts from: account structure, central logging, single sign-on and guardrails. |
| Control | A Control Tower rule. Preventive controls are SCPs or RCPs, detective controls are AWS Config rules, and proactive controls are CloudFormation hooks that check resources before they are created. |
| Account Factory | Control Tower's way to create a new governed account from a template; Account Factory for Terraform (AFT) does the same from a Git repository. |
| Drift | When an account or OU no longer matches what Control Tower set up, for example after someone edits a managed SCP by hand. |
Choosing
When to use it, and when not
Use it when
- Use Organizations whenever you have more than one AWS account, which a data platform with separate environments always has.
- Use SCPs to enforce rules that must hold whatever IAM says: allowed regions, no disabling of CloudTrail, GuardDuty or Config, no public S3, no leaving the organization.
- Use Control Tower when the company wants a managed baseline with central logging and a control library, and the cloud team prefers a console-driven service over writing it all themselves.
- Use resource control policies when you need to say that buckets, keys and secrets in the organization may only be used by principals in the organization or by named partners such as Databricks.
- Use separate accounts per environment so that a dev mistake, a quota exhaustion or a cost spike cannot touch prod.
Avoid it when
- Avoid building a second landing zone just for the data platform when the company already has one; ask for accounts in it and add OU-level rules only where needed.
- Avoid attaching untested SCPs to the root of the organization; one bad deny can lock every account out of a service.
- Avoid running workloads, pipelines or Databricks in the management account; SCPs do not protect it.
- Avoid hand-editing the SCPs and Config rules that Control Tower manages; it reports drift and may overwrite or break the landing zone.
Production grade
What goes wrong in production
| Problem | What breaks | What to do |
|---|---|---|
| SCP blocks Databricks | A region or service deny SCP stops the Databricks cross-account role from creating EC2 instances, or blocks STS calls the workspace needs. | Test SCPs in a sandbox OU with a real workspace first; check the exact actions in Databricks' cross-account role policy; use CloudTrail errors that name the SCP to debug. |
| Global services and region deny | A deny-all-outside-Mumbai SCP also blocks IAM, CloudFront or Organizations calls, which run from us-east-1. | Exempt global services in the region-deny condition, as AWS's example SCP does, and test before attaching. |
| Service-linked roles not covered | Teams assume an SCP stops everything, but service-linked roles and the management account are not restricted. | Know the exemptions; protect the management account with few people, MFA and no workloads. |
| Policy size and count | SCPs are capped in characters and per target; teams pile on statements until a change fails to save. | Keep one broad allow and a few targeted denies; group rules by purpose; use declarative policies for configuration settings. |
| Landing zone upgrades | A Control Tower landing zone update changes buckets, controls or notifications, and log queries or alerts that read them break. | Read the release notes before each upgrade, test in a separate organization if you can, and update Athena tables and alert rules that read Control Tower buckets. |
| Account vending speed | Control Tower runs a limited number of account operations at once, so creating many accounts in a hurry queues. | Plan accounts ahead; vend them from Git with AFT or Terraform in batches. |
| Access sprawl across accounts | People get long-lived users in each account, and leavers keep access somewhere. | Use IAM Identity Center with groups from the company identity provider; no IAM users for people; review permission sets quarterly. |
Alternatives
What to compare it with
| Alternative | Pick the alternative when | Pick Organizations and Control Tower when |
|---|---|---|
| Landing Zone Accelerator on AWS | A regulated company wants a large, configurable baseline with many security services, deployed as code from AWS's open-source solution. | You only need a small, standard landing zone; Control Tower alone is simpler. |
| Organizations with your own Terraform, no Control Tower | The cloud team already runs everything from Terraform and wants full control of SCPs, CloudTrail and Config. | No one wants to own and test the baseline code; Control Tower gives a maintained one. |
| Single account with IAM boundaries | A proof of concept for a few weeks. | Anything that will hold real data or reach prod; separating later is painful. |
| Azure management groups and Azure Policy | The platform runs on Azure Databricks; the same ideas map to management groups, policy and landing zones. | The data platform is on AWS. |
| Google Cloud folders and Organization Policy | The platform runs on Google Cloud. | The data platform is on AWS. |
Cost
How you pay
AWS Organizations and SCPs have no charge. Control Tower has no charge of its own either; you pay for the services it turns on in your accounts, mainly AWS Config rule evaluations and configuration items, CloudTrail beyond the free copy of management events, S3 storage for logs, and any KMS keys and Lambda functions it uses. The bill grows with the number of accounts, regions governed and detective controls enabled.
What drives the bill
- AWS Config configuration items recorded per resource change, multiplied by accounts and governed regions.
- Number of detective controls (Config rules) enabled per OU.
- CloudTrail data events, if turned on for S3 or Lambda, and storage of logs in the log archive account.
- Security services added alongside, such as GuardDuty, Security Hub and Macie, priced per account and per volume.
Cost traps
- Config recording every change on busy resources, such as Databricks classic compute creating and deleting instances all day, makes configuration items the biggest line on the security bill.
- Governing every region by default records Config in regions you never use; govern only the regions you allow.
- S3 data events on the data buckets log every object read by pipelines; useful for audit of sensitive buckets, very costly on bronze and silver.
- Log retention without lifecycle rules keeps years of logs in a standard storage class.
Prices change; check the AWS pricing page for your region before deciding.
Security and governance
Keeping it safe
- SCPs set the ceiling for every member account: allowed regions only, no disabling CloudTrail, Config, GuardDuty or the S3 public access block, and no leaving the organization.
- Resource control policies keep data buckets, KMS keys and secrets usable only by the organization's principals and named outside parties such as the Databricks control plane account; test them against Databricks' documented account IDs before attaching.
- The management account runs nothing else, has very few administrators with hardware MFA, and its root user is locked away.
- People sign in through IAM Identity Center federated to the company identity provider; there are no IAM users for people, and break-glass access is logged and alarmed.
- The log archive account receives organization CloudTrail and Config history into buckets protected by Object Lock or deny-delete policies, readable by security and audit only.
- Every change to OUs, SCPs and accounts is made through Terraform pull requests and appears in CloudTrail in the management account.
Operating it
What to watch, and the limits
| Watch | Why |
|---|---|
| Control Tower drift notifications | In landing zone 4.0 these go to EventBridge in the management account; route them to the platform team's channel. |
| AWS Config compliance per control | Non-compliant resources per account for each detective control. |
| CloudTrail AccessDenied events naming an SCP | Shows when a guardrail blocks real work, often the first sign that a new Databricks feature needs an action allowed. |
| Organizations policy and account changes | CloudTrail events such as AttachPolicy, UpdatePolicy, CreateAccount and MoveAccount in the management account. |
| Root user sign-in | Any root user sign-in in any account; alarm every time. |
| Config cost per account | The Config line in the cost report per account, to catch busy resources driving configuration items. |
| Limit | Value and notes |
|---|---|
| SCP size | 10,240 characters per policy |
| SCPs attached | Up to 10 per root, OU or account (at least one always attached) |
| RCPs attached | Up to 5 per root, OU or account, including the default RCPFullAWSAccess |
| OU nesting | Five levels below the root |
| Accounts | 10 by default per organization, raised on request; Control Tower supports up to 10,000 |
| Control Tower concurrency | Up to 5 account operations and up to 100 control operations in progress at once |
Limits and quotas checked 2026-11-11; many can be raised through Service Quotas.
Decisions
Situations and verdicts
| Situation | Choice | Verdict |
|---|---|---|
| Company already has Control Tower with log and security accounts | Request dev, pre-prod and prod accounts under a Workloads OU from Account Factory | Pursue: inherit logging and guardrails; add only data-platform SCPs such as the region rule at the OU. |
| No landing zone at all | Minimal Organization, CloudTrail to a log account, a security account and a few SCPs, built with Terraform or Control Tower | Pursue: about a week, and far cheaper than splitting a single account later. |
| Restrict data buckets and keys to the organization and Databricks | Resource control policies at the Workloads OU | Pursue with care: powerful against data leaving, but test with a real workspace so Databricks access is not cut. |
| Vend accounts from Git | Account Factory for Terraform | Pursue with care: good for many accounts; check region support, as AFT cannot be newly deployed in some regions, including Hyderabad. |
| Separate AWS organization just for the data platform | A second landing zone | Drop: doubles security operations and splits audit; ask for an OU in the company's organization instead. |
FAQ
Questions people ask
What is the difference between AWS Organizations and Control Tower?
Organizations is the building block: accounts, OUs, consolidated billing and policies such as SCPs. Control Tower uses Organizations to build and maintain a ready-made landing zone with central logging, single sign-on and a library of controls. You can use Organizations without Control Tower, but not the other way round.
Do SCPs grant permissions?
No. An SCP only sets the maximum. A person or role still needs an IAM policy that allows the action, and the action must also be allowed by every SCP above the account. Since September 2025 SCPs accept the full IAM policy language, including conditions and resource ARNs in allow statements.
Do SCPs apply to the management account?
No. SCPs affect only member accounts and do not restrict service-linked roles. That is why the management account should hold no workloads and very few people.
Does Databricks need anything special in the landing zone?
It needs the cross-account role, storage roles and network in each workload account, and SCPs that do not block the actions those roles use. Region-deny and resource control policies are the usual culprits, so test them with a real workspace.
Is the Security OU still required in Control Tower?
Not since landing zone 4.0 in November 2025. Control Tower no longer creates or enforces it, and its integrations, such as Config and CloudTrail, can each be turned on or off. Existing landing zones keep working as they are.
Sources
Checked against
- AWS docs: Service control policies
- AWS docs: Organizations quotas
- AWS What's New: full IAM language for SCPs
- AWS docs: Control Tower landing zone 4.0 key changes
- AWS docs: Control Tower limitations and quotas
Checked 2026-11-11.
Related
Where it connects
AWS and Databricks reference platform · AWS IAM and STS · AWS CloudTrail · Amazon Macie · AWS on this site · All AWS data services
← Macie · PrivateLink →