Your Company Is Adopting AI—Here’s a Practical 90-Day Security Plan

Executive summary: Many growing companies now use AI in employee tools, software products, customer service, development, and security operations, but few have a complete inventory or consistent security requirements. NIST’s preliminary Cyber AI Profile separates the problem into three manageable areas: securing AI systems, using AI for cyber defense, and preparing for AI-enabled attacks. The 90-day plan below translates those ideas into practical work for leadership, lean IT teams, and MSP-supported businesses.

Estimated reading time: 11 minutes

Many companies have reached an awkward stage in AI adoption: employees, vendors, and product teams are already using AI, but the organization does not yet have a coherent way to manage the cybersecurity consequences.

The result is predictable. Security teams ask for an inventory that does not exist. Business leaders are unsure which AI uses require formal review. Employees create accounts with public tools. Vendors introduce AI features through ordinary product updates. Automated agents receive access to applications or company data without going through the same review as a conventional software integration.

Meanwhile, leadership receives broad warnings about data leakage, prompt injection, model attacks, deepfakes, and AI-enabled adversaries without a practical way to prioritize them.

NIST’s preliminary Cybersecurity Framework Profile for Artificial Intelligence offers a useful way to reduce that confusion. It organizes the relationship between AI and cybersecurity into three focus areas:

Secure: securing AI system components;
Defend: conducting AI-enabled cyber defense; and
Thwart: thwarting AI-enabled cyberattacks.

The profile is built around NIST Cybersecurity Framework 2.0 outcomes. It complements rather than replaces the broader NIST AI Risk Management Framework, which addresses trustworthiness and AI risk across the system lifecycle.

For a growing business, the value is not claiming alignment with another framework. It is using the three focus areas to make better decisions over the next 90 days.

Who Needs This 90-Day Plan?

This plan is designed for companies that are adopting AI faster than they can govern it. That includes businesses using employee AI assistants, AI-enabled software products, coding tools, embedded vendor features, customer-facing chatbots, security automation, or locally operated models.

It is particularly relevant to organizations with 10–200 employees, lean IT teams, or MSP-supported environments. These businesses may not need a large AI governance department. They do need someone to identify what is being used, decide which uses create material risk, and establish minimum expectations.

You may need this plan if leadership cannot answer questions such as:

Which AI tools and embedded features are currently in use?
What customer, employee, or company data is sent to them?
Which systems can take actions instead of only providing recommendations?
Who approved each high-impact use?
How is AI used by the IT or security team?
What happens if an AI integration behaves unexpectedly?
How is the company preparing for AI-assisted fraud and cyberattacks?

An AI policy is a useful starting point, but it does not answer these operational questions by itself.

Start by Separating Three Different Security Problems

Organizations often discuss “AI security” as if it were one control domain. NIST’s three focus areas help show why it is not.

First, the company must secure AI systems it develops, integrates, buys, or operates. That includes data, models, prompts, identities, infrastructure, APIs, agents, connectors, and supporting software.

Second, the company may use AI inside its security program—for alert triage, code analysis, investigation, documentation, vulnerability review, or automation. Those uses may improve a lean team’s capabilities, but their output and actions still require oversight.

Third, the organization must prepare for adversaries using AI to improve phishing, impersonation, vulnerability research, malicious code development, and operational scale.

Combining all three into a general statement such as “AI creates cyber risk” produces vague activity. Separating them allows leadership to assign owners, controls, and measurable outcomes.

Days 1–30: Establish Visibility and Decision Authority

The first month should focus on identifying AI use and defining who can make decisions about it.

Build a usable AI inventory

Do not limit the inventory to models developed internally or applications with “AI” in the product name. Include:

AI features embedded in existing software services;
public generative-AI tools used by employees;
models integrated through APIs;
AI-enabled customer support or sales features;
internally hosted or open-weight models;
AI coding assistants;
security products that use AI;
automated agents with access to applications or data; and
vendors that process company or customer information through AI.

For each use, record the business owner, purpose, users, vendor or model, deployment method, data involved, connected systems, output recipients, and whether the system can take actions.

A spreadsheet is acceptable at first. Completeness, ownership, and regular review matter more than purchasing a specialized governance platform.

Define review triggers

Not every experiment needs executive approval. Risk-based review triggers keep governance proportionate.

Require a formal review when an AI system:

handles regulated, confidential, customer, or employee data;
connects to production systems;
can perform actions rather than only recommend them;
generates customer-facing content;
affects employment, healthcare, credit, security, or another consequential decision;
uses an open-weight model in a company-controlled environment;
creates source code intended for production; or
depends on a vendor whose failure would disrupt an important business process.

Publish these triggers in language that business, legal, procurement, IT, security, and product teams can understand.

Assign accountable owners

An AI committee can coordinate work, but committees do not replace ownership.

The business owner should be accountable for the use case. IT or engineering should operate technical controls. Security should evaluate identity, data flows, integrations, logging, and failure scenarios. Legal and privacy teams should address applicable obligations. Procurement should preserve vendor and contract evidence.

Leadership should define who can approve high-risk uses, who can accept remaining risk, and who can suspend an AI system when conditions change.

WCS practical guidance: If your company has an AI policy but no complete inventory, review triggers, or named owners, WCS can help convert broad principles into a working AI governance process. Explore WCS AI Security and Governance services.

Days 31–60: Secure AI Systems and AI-Enabled Defense

The second month should convert visibility into minimum safeguards.

Apply familiar controls before purchasing exotic ones

AI introduces new attack techniques, but many serious failures still begin with familiar weaknesses: excessive permissions, exposed credentials, poor logging, unreviewed vendors, insecure code, and sensitive data sent to the wrong service.

Prioritize:

strong authentication and narrowly scoped service identities;
approved secrets management for model and API credentials;
separation of development, testing, and production;
encryption and retention controls for prompts and outputs;
logging of administrative activity and important model interactions;
dependency and vulnerability management;
human approval before consequential agent actions;
testing for prompt injection and unsafe tool use;
output validation before generated content reaches another system; and
a documented method for disabling an AI integration.

An AI agent should not inherit broad access simply because the employee who configured it has broad access. Give machine identities their own credentials, permissions, owners, and lifecycle.

Classify AI actions by consequence

A model that summarizes an approved policy presents a different level of risk from an agent that disables accounts, changes firewall rules, modifies customer records, or deploys code.

Start with read-only access. If the business later enables a write action, add it individually with limited permissions, clear input validation, human approval, logging, and a rollback procedure.

Do not treat access to one application as permission to act across every connected system.

Govern AI used by IT and security teams

AI can help lean teams summarize alerts, analyze logs, draft detection logic, review code, explain vulnerabilities, or prepare incident timelines. That does not make every generated conclusion reliable.

For higher-impact security uses, require:

approved data sources;
documented human-review points;
representative test cases;
protection against manipulated inputs;
action limits and rollback procedures;
logging sufficient to reconstruct what occurred; and
periodic review after model or vendor changes.

Measure whether the tool improves outcomes, not merely whether it uses AI. Useful measures may include investigation time, analyst rework, false-positive rates, detection coverage, and the number of unsafe automated actions prevented.

Days 61–90: Prepare for AI-Enabled Adversaries

The third month should strengthen the company against attacks that may become faster, cheaper, or more convincing through AI.

The answer is not purchasing a separate “AI threat stack.” It is reinforcing existing controls where increased speed and persuasion matter most.

Strengthen identity and transaction verification

AI-generated messages, documents, images, and voices make it harder to rely on familiarity as proof of identity. High-risk business processes should use verification outside the original communication channel.

Review procedures for:

payment and bank-account changes;
password and multifactor-authentication resets;
payroll changes;
release of sensitive records;
changes to vendor contact information;
creation of privileged accounts; and
urgent requests attributed to executives.

Employees need a simple rule: slowing down a high-impact request to verify it is an approved security action—not poor customer service.

Reduce exposed attack surface

AI may help adversaries discover and exploit weaknesses more efficiently, but it cannot exploit a service that is not exposed or an account that no longer exists.

Use the 90-day program to remove unnecessary internet-facing systems, prioritize high-risk vulnerabilities, disable stale accounts, rotate exposed credentials, and validate cloud permissions.

Exercise an AI-related incident

Add an AI element to an existing incident exercise rather than creating a science-fiction scenario.

Practical scenarios include:

an employee entering customer data into an unapproved AI service;
an agent taking an unauthorized action after processing malicious instructions;
AI-generated code introducing a vulnerable dependency;
a vendor adding an AI feature that changes data handling;
a convincing executive impersonation triggering a fraudulent request; or
a model producing an inaccurate security recommendation during an incident.

The exercise should test escalation, decision authority, evidence preservation, communication, containment, and recovery—not participants’ knowledge of AI terminology.

The 90-Day Cyber AI Checklist

By day 30

☐ Inventory approved, experimental, embedded, and employee-selected AI uses.
☐ Record owners, users, data, vendors, integrations, and action capabilities.
☐ Publish risk-based review triggers.
☐ Define who can approve, reject, suspend, and accept risk for AI uses.
☐ Identify unapproved AI use that requires remediation or review.

By day 60

☐ Apply minimum identity, credential, logging, retention, and vulnerability requirements.
☐ Separate AI service identities from employee administrator accounts.
☐ Document human-approval requirements for consequential actions.
☐ Review AI used by IT and security teams.
☐ Test prompt injection, excessive permissions, unavailable dependencies, and inaccurate output.
☐ Establish a process for disabling an AI integration.

By day 90

☐ Strengthen out-of-band verification for high-risk business requests.
☐ Reduce unnecessary public exposure and resolve priority vulnerabilities.
☐ Run an AI-related incident or tabletop exercise.
☐ Assign owners and dates to unresolved findings.
☐ Present leadership with outcomes, exceptions, metrics, and next-quarter priorities.

What Leadership Should Expect at Day 90

A credible 90-day program should produce business artifacts, not only presentations.

Leadership should expect:

an owned AI system and use-case inventory;
documented review and approval triggers;
a practical risk-tiering method;
minimum security requirements for AI integrations;
a register of important AI identities and permissions;
rules for AI-enabled security automation;
an AI-related incident exercise and lessons learned;
prioritized remediation work with owners and dates; and
a small set of metrics for recurring leadership review.

The program will not solve every AI risk in one quarter. It will give the company a repeatable way to make decisions as technology, vendors, and threats change.

Use the Profile as a Decision Tool, Not a Badge

The NIST Cyber AI Profile is most valuable when it helps an organization distinguish among securing AI, using AI for defense, and preparing for AI-enabled attacks.

That distinction helps prevent two common mistakes: blocking useful adoption because every risk appears unfamiliar, or accelerating adoption without understanding identities, data, dependencies, and operational consequences.

The goal is not framework alignment for its own sake. It is a defensible operating model that lets the business adopt useful AI with clearer boundaries and fewer surprises.

Start With a Small Business Security Snapshot

AI governance depends on broader cybersecurity fundamentals. Weak identity controls, unmanaged vulnerabilities, incomplete asset records, poor vendor oversight, and an untested response plan will also weaken the company’s ability to adopt AI safely.

The WCS Small Business Security Snapshot provides a prioritized view of identity, cloud, vulnerability, incident-readiness, vendor, and governance gaps. It can help leadership determine what to address first and where AI-specific controls belong in the roadmap.

Start the WCS Small Business Security Snapshot or schedule a consultation.

References

This article provides general cybersecurity and governance guidance. NIST’s IR 8596 is a preliminary draft, and organizations should review current NIST publications and obtain appropriate legal, regulatory, and technical advice for their circumstances.