A zero trust implementation roadmap is a staged plan for replacing "trusted because it's on our network" with a decision on every request: who is asking, from which device, for which resource. For a company of 50 to 500 people, the order that removes the most risk first is to take an inventory, move admins to phishing-resistant MFA, swap vendors' VPN accounts for per-application access, and close admin ports to the internet. That fits in about 90 days. Single sign-on everywhere, device checks, segmentation and data controls follow over the next 12 to 15 months, and much of that later work falls on whoever owns each application, not only on IT.
The plan below assumes a typical mid-size setup: Microsoft 365 or Google Workspace, a few dozen SaaS tools, some workloads in AWS, Azure or Google Cloud, a VPN, and a small IT team without a dedicated security engineer.
What zero trust asks for
NIST's SP 800-207, published in August 2020, is still the reference definition. Its starting assumption is that no user account or device gets implicit trust from its network location or from who owns it. Every access request goes through a policy decision point, which weighs identity, device state and context, and a policy enforcement point in front of the resource, which opens the session or doesn't.
In a mid-size company the decision point is usually your identity provider (Microsoft Entra ID, Google Workspace, Okta, Keycloak) and its access policies. The enforcement points are an access proxy in front of internal apps, each SaaS app's single sign-on, and your cloud provider's IAM.
For planning, use CISA's Zero Trust Maturity Model. Version 2.0, from April 2023, splits the work into five pillars (identity, devices, networks, applications and workloads, data) and three capabilities that run across all of them (visibility and analytics, automation and orchestration, governance). It describes each pillar at four stages: traditional, initial, advanced and optimal. Score yourself pillar by pillar before you plan anything. If you use MFA but not yet phishing-resistant methods, expect "initial" at best on identity and "traditional" on most of the rest, which tells you where to begin.
For more depth, NIST's SP 1800-35 (June 2025) documents example builds with commercial products, and the NSA's Zero Trust Implementation Guidelines, first published in January 2026, break the work into phases for US defense organizations. Both are worth skimming, though neither is written for a company with a three-person IT team.
Why admin accounts and vendors come first
Verizon's 2026 Data Breach Investigations Report found that 31% of breaches started with an exploited vulnerability, overtaking stolen credentials as the most common way in, and that a third party was involved in 48% of breaches. VPN appliances sit where those two trends meet: they face the internet, and contractors use them. In January 2024 CISA told US federal agencies to disconnect their Ivanti Connect Secure and Policy Secure appliances within about two days, because attackers were still exploiting them after the first round of mitigations.
So the first stage goes after two things: what a stolen or phished admin login can do, and what an outsider can reach from the internet or from a vendor's laptop. Segmentation comes later, once you know who connects to what.
Stage 0: inventory (two to three weeks)
You can't put conditions on access you can't see. Build one sheet with four lists, and give every row an owner:
- Applications: every SaaS tool, internal web app, admin panel and database console, and how people sign in today (SSO, a local password, VPN plus password, a shared account).
- Privileged accounts: global admins, cloud root and owner accounts, domain admins, database superusers, and anyone who can change DNS, payroll or billing.
- Remote access: every VPN account, SSH key, RDP host and firewall rule open to the internet, with the person it belongs to. Expect to find former contractors here.
- Non-human identities: service accounts, API keys, CI/CD secrets, OAuth apps with access to mail or files, and integrations with other organizations' systems.
Pull the data straight from the systems where you can: your identity provider's sign-in logs, the VPN's account list and logs, each cloud provider's credential reports, and the list of third-party apps that users have allowed into their mailboxes.
The fourth list is the one teams skip, and it's often where sensitive data flows. Oplist, a hospital rostering platform in Türkiye and a client we build for, receives leave types from hospitals' HR systems through API integrations, because its scheduling rules depend on the kind of leave. Sick leave is one of those types. A rota tool doesn't look like a medical system, but a record that someone was on sick leave can count as health data, which both Türkiye's KVKK and the GDPR treat as a special category. A second case: the field-ticket workflow we built with an agency partner for an oilfield services company runs every evening, picks up tickets that arrived by Gmail and files them in Google Drive without anyone signing in. Each integration like these reaches data that matters, so it gets a row in the inventory and an owner, the same as a person would.
Stage 1, days 1 to 90: close the widest doors
Admins on phishing-resistant MFA
Move every privileged account to passkeys or FIDO2 security keys, or to certificate-based authentication if you already run a PKI. These credentials are bound to the real site's domain, so even a convincing fake login page gets nothing it can replay. SMS codes, authenticator-app codes and push approvals are far better than a password alone, but a phishing proxy can relay all three in real time.
Three details are easy to miss:
- Give admins separate accounts that are used only for admin work and have no mailbox.
- Create two emergency ("break glass") accounts with their own security keys, exclude them from any policy that could block sign-in, store the keys apart, and alert on every use. Microsoft's Entra ID guidance also asks for a test sign-in at least every 90 days, and the same logic applies to any identity provider.
- Turn off legacy sign-in protocols that can't do MFA.
Vendors and contractors: per-app access with an end date
Vendor VPN accounts are the cheapest big win. A contractor who needs one internal app and one database often has a tunnel into a whole network segment, created years ago and never reviewed. Replace these first with zero trust network access (ZTNA). A small connector inside your network or cloud account makes an outbound connection to the access service, the user signs in through your identity provider, and the service lets them reach only the applications their policy names. Nothing listens on an inbound port, and the rest of the network isn't reachable from their laptop.
Give every external person a named account in your identity provider, limited to what they work on, with an expiry date. In our clients' systems we are the vendor, so we see this from the other side too. On the Tribute Technology platform, the build phase ended in September 2023 when we moved the production code into Tribute's own repositories and AWS account, and maintenance ran into 2024. Changes like those are when a vendor's access should shrink or end, and they're easiest to enforce as dates entered when the account is created.
No admin ports open to the internet
Close SSH, RDP and database ports to the internet. The big clouds have identity-based alternatives. AWS Systems Manager Session Manager gives shell access controlled by IAM policies, with no inbound ports, bastion hosts or SSH keys, and it can log every session. Google Cloud's Identity-Aware Proxy and Azure Bastion cover the same need. Put anything else behind the same ZTNA connector.
Stage 2, months 3 to 9: everyone and their devices
- Put every app from the inventory behind SSO over SAML or OpenID Connect, starting with the ones that hold customer or personal data. Turn on SCIM provisioning where the app supports it, so that leaving the company removes access everywhere at once. Check plan tiers early: many SaaS vendors offer SSO only on their business or enterprise plans, and those upgrades can become the largest line in the budget.
- Move all staff to passkeys. The passkeys built into phones and laptops cost nothing extra. Go team by team with the help desk ready, because enrollment is where the hours go.
- Add device conditions for sensitive apps: a managed device, disk encryption, a supported and updated OS, endpoint protection running. Personal and unmanaged devices get browser-only access to a smaller set of apps, without downloads. Phones need a policy of their own; our post on enterprise mobile apps covers what a backend can check about a device.
- Move staff off the VPN one application at a time. Keep a VPN for the long tail (thick clients, unusual protocols) with a short list of named users, and stop treating it as the default way in.
If your remote access runs on a FortiGate, the next firmware upgrade may set the schedule for you. FortiOS 7.6.3 removed SSL VPN tunnel mode on all models and replaced it with IPsec VPN, and tunnel-mode settings aren't carried over in the upgrade. Since remote access has to be rebuilt anyway, rebuild it only for the people who still need a network tunnel and move everyone else to per-app access.
Stage 3, months 9 to 18: segmentation, workloads and data
Many roadmaps put micro-segmentation first. In a mid-size company it breaks far less when it comes last, after identity logs have shown who actually talks to what. Start with the systems whose compromise would hurt most: production databases, backups, admin interfaces, the CI/CD system.
- Separate production from everything else, including staging and the office network.
- Write allow rules for the flows you observe, run them in log-only mode for a few weeks, then switch to deny by default. In the cloud that means security groups that reference other security groups instead of IP ranges. In Kubernetes it means a default-deny NetworkPolicy per namespace, which only takes effect if your network plugin enforces policies.
- Give workloads their own identities. CI pipelines can get short-lived cloud credentials through OIDC federation instead of long-lived keys stored as secrets; our DevSecOps post covers the other security checks that belong in the same pipeline. Service-to-service calls should carry a credential the receiving service verifies, such as mTLS certificates from a service mesh or short-lived signed tokens.
- Give each integration with another organization, like the hospital feeds above, its own credential, limited to the endpoints it uses, rotated on a schedule and logged.
- Know which data stores hold personal or special-category data, log reads, and review who has access every quarter.
The application is the last enforcement point. A good access proxy doesn't help if the app behind it lets any signed-in user open any record, so authorization has to be checked on every request, inside the app, against the user's role and the record's owner. That's ordinary backend work. On Oplist it sits in the organizations and roles we added to the platform; on Tribute's migration platform, four roles (superadmin, admin, staff and client) control who can do what.
Across all stages, send sign-in and access logs somewhere a person will actually look, with alerts on a short list of high-signal events: a new admin, a changed MFA method, an emergency account used. Our incident response post lists good first alerts.
The roadmap at a glance
Stage | Main work | Done when |
|---|---|---|
0: inventory (weeks 1 to 3) | Apps, privileged accounts, remote access, non-human identities; CISA pillars scored | Every row has an owner |
1: days 1 to 90 | Admin passkeys or keys, break-glass accounts, ZTNA for vendors, admin ports closed | No admin signs in with a password and SMS; nobody outside the company has a VPN account |
2: months 3 to 9 | SSO and SCIM, passkeys for staff, device conditions, a smaller VPN | Apps with personal data require SSO and a healthy device; VPN users are named exceptions |
3: months 9 to 18 | Segmentation, workload identities, integration credentials, data access reviews | Production is isolated; no long-lived cloud keys in CI |
What it costs
Licences are the predictable part. On Microsoft's Entra pricing page, Entra ID P1, which includes Conditional Access, lists at $7 per user per month on an annual commitment. Entra Private Access, Microsoft's ZTNA product, is $5, and the Entra Suite bundle, which includes it, is $12 and requires P1. Microsoft 365 E3 and Business Premium already include P1. Cloudflare's Zero Trust plans are free for up to 50 users and $7 per user per month on pay-as-you-go beyond that. For 150 people, that's about $1,050 a month on Cloudflare, or $750 a month for Private Access on top of P1 licences you already have. Google, Okta, Zscaler and others sell comparable products, and on larger annual contracts the list price is where negotiation starts.
The costs that are harder to estimate:
- SSO upgrades on SaaS plans you already pay for.
- Two hardware security keys per admin, so that one lost key doesn't lock anyone out.
- Staff time: app owners testing each app behind the new access path, and help-desk hours during passkey enrollment. Put this in the plan explicitly, because it's the cost that gets underestimated.
How to tell it's working
Track coverage. A count of blocked attacks rises when attackers get busier and falls when they move on, so it says little about your own progress. These numbers say more:
- the share of applications behind SSO, and the share of those holding personal data that also check the device;
- the share of admin accounts on passkeys, security keys or certificates;
- the number of VPN accounts, and how many belong to people outside the company;
- external accounts without an expiry date;
- admin ports (SSH, RDP, databases) reachable from the internet, where the target is zero;
- cloud access keys and API tokens older than 90 days.
Pull them once a month into the same sheet as the inventory, so that every number can be traced back to the rows behind it.
Mistakes that stall the work
- Buying a "zero trust platform" before the inventory, then shaping the program around one product's boundaries.
- Exceptions that never expire. A "temporary" exclusion from an access policy needs an owner and an end date, just like a contractor account.
- Locking yourself out. Run new policies in report-only or audit mode first, and test the break-glass accounts.
- Treating MFA as finished while SMS codes are still accepted for admin accounts. Attackers go for the weakest method an account allows.
Frequently asked questions
Is zero trust a product you can buy?
No single product delivers it. You assemble it from an identity provider, an access proxy or ZTNA service, device management and logging, and run it as a set of policies. A vendor's "zero trust platform" covers one or more of those parts, and the inventory from stage 0 shows which ones you are missing.
Is zero trust worth it for a small company?
Stage 1 is, at almost any size. Phishing-resistant MFA for admins, SSO for the main apps and no vendor VPN accounts cost little and close the most common ways in. A 30-person company that runs on SaaS can do most of it in a few weeks. Segmentation and device posture can wait until you run internal systems worth segmenting.
Does this help with an ISO 27001 or SOC 2 audit?
For the access controls, yes. ISO/IEC 27001:2022 has Annex A controls for access control, identity management, authentication information, access rights and privileged access (5.15 to 5.18, 8.2 and 8.5), and the CC6 criteria in SOC 2 cover logical access. Auditors ask who has access, who approved it, how often it is reviewed and how it ends. The inventory with owners, SSO with SCIM, expiry dates on external accounts and the sign-in logs from stages 0 to 2 are the evidence. The rest of either standard, from risk assessment to business continuity, is separate work.
What's the difference between ZTNA and SASE?
ZTNA is one piece: per-application access to private apps. SASE (secure access service edge) is a cloud-delivered bundle that usually combines ZTNA with a secure web gateway, a cloud access security broker, firewall as a service and SD-WAN. You can adopt ZTNA on its own; a SASE contract isn't a prerequisite.
If the work lands in your own applications, such as adding SSO to an internal tool, checking authorization on every request or replacing shared API keys with per-service credentials, that's backend work we do.
Related Articles
DevSecOps Pipeline: What to Scan, Where, and What Should Block
Secret, code, dependency, image and DAST scans for a small team's CI/CD pipeline: where each runs, what should block a merge, and a working GitHub Actions file.
Read More: DevSecOps Pipeline: What to Scan, Where, and What Should BlockSecure SDLC: What Each Phase Should Produce
What a secure SDLC adds to each phase of development: requirements you can test, a worked threat model for a CSV import, and where a small team should start.
Read More: Secure SDLC: What Each Phase Should ProduceOWASP Top 10 2025: What Changed and How to Fix Each Risk
OWASP's 2025 list adds two categories and folds SSRF into access control. What each risk looks like in a real web app or API, and what to fix first.
Read More: OWASP Top 10 2025: What Changed and How to Fix Each Risk
