PCI DSS scope is the set of systems an assessor will test, and it is decided by connectivity rather than intent. Every system that stores, processes or transmits account data is in scope, every system that can reach one of those is in scope, and every system that can affect the security of either is in scope. Segmentation is the only lever that shrinks the list.
The reason this is worth engineering rather than arguing about: scope is the multiplier on every other cost in the assessment. Each additional in-scope system means more evidence, more controls to test, more remediation and more disagreement. Reducing scope is the highest-return compliance work available, and the account boundary is where it starts.
What is actually in scope?
#Most teams think of scope as one ring. It is three.
| Category | What it means | Common examples | Can isolation remove it? |
|---|---|---|---|
| Cardholder data environment | Stores, processes or transmits account data | Payment service, checkout API, any datastore holding primary account numbers, log sinks that captured them by accident | No. This is the thing being protected |
| Connected-to systems | Can reach the environment over a network, or is reachable from it | Internal admin tools, shared databases, support consoles, anything on a permissive network path | Yes, this is exactly what segmentation removes |
| Security-impacting systems | Can affect the security of the environment without holding card data | Identity provider, key management, logging and monitoring, the deployment pipeline, the management account and its guardrails | No. Isolation does not help; hardening and evidence do |
The third row is the one that surprises people, and it is the row that quietly determines how long an assessment takes. Your single sign-on system is in scope because whoever controls it controls access. Your CI/CD pipeline is in scope because whatever it deploys into the environment is whatever an attacker could deploy. Your log archive is in scope because its integrity is the basis of every other claim you make.
Why the account boundary is the right unit
#An account boundary is enforced by the cloud provider, above the workload, and it holds even when something inside it is compromised. Cross-account access has to be granted explicitly, which means there is no inherited permission and no accidental network adjacency to discover during an assessment. It is also the natural unit of everything an assessor asks about: billing, quotas, logging destination, blast radius and evidence scope all follow it.
The AWS Enterprise Baseline deploys one organization with 15 accounts across 2 organizational units, which gives you the vocabulary for a scoping conversation that would otherwise be a survey of tags. If the cardholder data environment is an account, then “list every system in scope” is a query and not a project. The multi-account landing zone post covers why the structure is shaped that way.
But an account boundary alone is not segmentation. Two accounts whose networks are joined by a permissive route are one environment with two bills. The boundary has to exist at both layers.
How the network boundary is enforced
#In the AWS Enterprise Baseline, spoke networks connect through a Transit Gateway with two isolation domains, production and non-production, each holding a route table that blackholes the other’s address ranges. Production and non-production cannot reach each other, and that is a property of the routing table rather than a convention anyone has to remember. Egress is centralized in a hub network with a NAT gateway in each of three Availability Zones, so spokes have no direct internet path of their own. Addresses come from an IP address manager with one address map, and a CI guard script rejects any pull request containing a hardcoded network range — which matters here because the most common segmentation failure is a range typed by hand into a rule that nobody revisits.
For an assessor, that yields three artifacts that are hard to argue with: the route tables showing the blackholes, flow logs from all six networks showing what actually traversed them, and the pull request history showing when the routing last changed and who approved it.
What the PCI switches give you
#The AWS Enterprise Baseline ships PCI coverage in two shapes, both off by default because both cost money to run and only some organizations are in scope.
| Switch | What it is | Size | Default |
|---|---|---|---|
| Security Hub PCI DSS v4.0.1 standard | Control-level findings per resource, grouped by requirement | 144 controls | Off |
| PCI DSS v4.0 conformance pack, including global resource types | Config rules with a compliance score per account, evaluating global resources such as identity | 127 rules | Off |
| PCI DSS v4.0 conformance pack, excluding global resource types | The same, for additional regions, so global resources are scored once | 99 rules | Off |
| AWS Foundational Security Best Practices | Broad control findings that the PCI review leans on for data protection | More than 300 controls | Off |
The region rule is the part people get wrong: run the including-global pack in the home region and the excluding-global pack everywhere else. Run the including-global pack in every region and you will spend the first week of the assessment explaining why the same identity finding appears four times.
The underlying controls those checks evaluate are not optional extras. Organization-wide CloudTrail with log file validation, AWS Config recording in every account, GuardDuty with malware protection, Firewall Manager attaching a web application firewall to every workload load balancer, TLS-only storage policies, image scanning on push with immutable tags, and identity with no long-lived keys anywhere are on from the start. The switches add framework-shaped scoring on top of controls that already exist, which is the right order: if the pack is the first time a control appears, the pack will mostly tell you bad news.
Mapping requirements to mechanisms
#A condensed version of the mapping, enough to check your own environment against.
| Requirement | The mechanism that answers it |
|---|---|
| 1. Network security controls | Transit Gateway isolation domains with blackhole routes, centralized egress, a web application firewall on every workload load balancer, optional egress inspection behind one switch |
| 2. Secure configurations | Configuration recording in every account, a hardened account baseline applied identically, instance metadata service version 2 enforced |
| 3. Protect stored account data | Encryption defaults per account with customer-managed keys and annual rotation |
| 4. Protect data in transmission | TLS-only storage policies with a TLS 1.2 minimum, HTTPS-only listeners with managed certificates |
| 5. Protect against malicious software | Threat detection in every account, including malware scanning of attached volumes |
| 6. Secure systems and software | Image scanning on push, immutable image tags, infrastructure changes through reviewed pull requests with policy and vulnerability scanning in the pipeline |
| 7 and 8. Access control and authentication | Single sign-on as the only human path, no identity users and no long-lived keys, MFA enforced by policy above the account, permission sets scoped per environment, unused access flagged at 90 days |
| 10. Log and monitor all access | Organization trail into a separate archive account, 365-day retention with lifecycle to colder storage, seven CIS alarms, network flow logs |
| 11. Test security regularly | Continuous control findings, plus your own penetration testing, which no product performs for you |
| 12. Organizational policy | Yours. Infrastructure contributes nothing here and should not pretend to |
The honest limits of scope reduction
#Five things segmentation will not do, and it is better to know them before an assessor says them.
- It does not shrink requirement 12. Policies, roles, awareness training, incident response and vendor management stay exactly the same size whether your environment is one account or fourteen.
- It does not remove the shared services. Identity, logging, key management and the pipeline are in scope because they are security-impacting. Isolating them from the card environment does not help; the only available move is to control and evidence them well.
- It has to be tested, not asserted. Segmentation is validated by penetration testing at an interval set by your merchant or service provider level. A diagram is not evidence, and neither is a route table on its own.
- It decays. A peering connection added in a hurry, a security group opened for an incident, a shared database quietly reused: each of these silently re-expands scope. Drift detection against the committed configuration is what keeps the boundary honest, and drift is a fact, not a failure covers how to treat it as a signal rather than an embarrassment.
- Your assessor decides scope, not you. You can make the boundary defensible and the evidence easy. The determination is theirs, in exactly the way certification is theirs.
What to do next
#If card data is in your future rather than your present, the order is straightforward: decide whether you can avoid receiving account data at all, then put whatever remains in its own account, then make the network boundary say the same thing as the account boundary, then switch on the standard and the packs at least two weeks before anyone reads the score.
If card data is already in a shared account, the work is a migration rather than a flag, and it is worth starting before an assessment date exists. Readiness and certification are different jobs, and only the first one is yours to schedule.
The controls described here ship in the AWS Enterprise Baseline and are listed framework by framework on the compliance readiness page. If you want to know what your current environment would cost to bring inside a defensible boundary, that is what the BuiltForProd Assessment exists to tell you.