Cybersecurity Reach Foundation
AI Security & Governance

A Business's Guide to Agentic AI Adoption

Agentic AI is one of the fastest-moving technologies in business today, and the appeal is obvious: let the AI do the work for you. This is a beginner-friendly reference for adopting these tools effectively and responsibly — work through it and you may find yourself becoming an agentic AI power user.

SK

Seva Karonis

AI Security & Governance Researcher

The Basics

What Is Agentic AI?

Agentic AI refers to autonomous AI systems that can carry out tasks on their own. Unlike a traditional chatbot, which needs constant instructions, an agent independently decides what actions to take and then executes them.

Its capabilities range from sorting your inbox, sending emails, and scheduling meetings, to automating workflows and writing code.

Ground Rules

General Tips for Safe Usage

To keep your use of agentic AI safe and transparent, it's worth putting the following practices in place.

1

Don't share PII with an AI chatbot

This includes your place of work, client data, and financial records. Most importantly, do not share proprietary information unless the organization has explicitly permitted it — and only on an approved plan.

2

Require organizational accounts and devices only

This limits the data an AI can access and gives the organization more control and visibility over how the tools are used.

3

Enforce SSO and MFA

Require Single Sign-On and Multi-Factor Authentication when signing into company-approved AI tools.

4

Limit data retention

Employees should avoid storing sensitive prompts or outputs beyond what's necessary.

5

Keep a human in the loop at every step

Before allowing an AI to act on your behalf, establish (1) what the AI is allowed to do and (2) the escalation process if something goes wrong. An LLM cannot be held accountable — a named person must be responsible.

Bringing Tools In

The Onboarding Process

Before onboarding any tool, research and vet it thoroughly, and identify the specific role it will play in your organization. Start with one question: will your organization use it, or will your employees use it? In other words, do you want AI deeply integrated into daily operations, or simply automating individual workflows?

When assessing a tool, look at the permissions it requires, how it integrates with your existing technology, and its terms of service. Where possible, request a copy of the vendor's SOC 2 report — an independent audit that verifies how your data is protected — for added assurance.

Once a tool is vetted, scale the decision across the organization. Communication with staff is essential, and it's worth revisiting the decision annually, since these tools may not be needed permanently.

Ways to Communicate the Rollout

  • An onboarding meeting with a live demo
  • Policy documents outlining permissions
  • An organization-wide email
  • An annual review of whether the tool is still needed

The Most Important Step

Defining What Agents Can Do

Because agents act autonomously, they can take actions across your organization — which makes clear boundaries essential from the start. A useful way to scope agent permissions is by risk severity and reversibility.

The Three-Tier Model

1

Prepare only

Agents can draft actions, such as composing emails, but cannot execute them without human approval.

2

Execute low-risk, reversible tasks

Agents can independently handle things like scheduling meetings or sending internal updates on industry trends.

3

Never auto-execute high-risk, irreversible actions

Agents must not, on their own, email clients, move money, or share financial data.

Getting specific here prevents miscommunication and costly mistakes later. Decide in advance what happens when something goes wrong: Who is responsible for the AI's mistake? What is the escalation process? What can be done to limit the impact of an incorrect decision?

From this, create and share an organization-wide policy for employee guidance. Be specific about which tools are and are not permitted. For example, if your organization adopts Claude, can employees use it only for basic administrative tasks, or may they upload client data?

Reference Points & Cautionary Tales

Proactive policy

Apple

Restricted employee use of ChatGPT over concerns about data leakage.

Proactive policy

Amazon

A leaked internal email warned employees not to use unauthorized AI tools for work, citing privacy concerns.

Insufficient controls

Samsung

Initially allowed ChatGPT, then issued a company-wide ban after sensitive source code was leaked — three separate leaks within 20 days of access being granted.

Where to Start

Tool Recommendations

It's easy to be overwhelmed by the thousands of options. Here are a few popular ones — this list will change as AI tools advance rapidly, so re-evaluate your choices regularly.

Claude Cowork

An agentic AI tool built to execute complex tasks and simplify day-to-day operations. Commonly used for meeting preparation, document review, and calendar management.

Watch outSome users find token usage can run high, which can make it expensive at scale.

Microsoft 365 Copilot Cowork

An agentic AI tool built to handle complex tasks across Microsoft 365 apps (Word, Excel, Outlook, and Teams). A natural fit for organizations already in the Microsoft ecosystem.

Watch outFor teams without a Microsoft license, it's generally not worth purchasing on its own.

Manus

An agentic AI tool built to manage workflows end to end — generating content, running code, and conducting research. Approachable for non-technical users and well regarded for autonomous task execution and in-depth research.

Watch outUsers have reported stalled task loops, inaccurate browser actions, and incomplete deliverables — it may be better suited to prototyping and research than production work.

Adopt agentic AI on your terms.

The organizations that get this right aren't the fastest to adopt — they're the ones that set clear boundaries first. Vet the tool, define what agents can do, and keep a named human accountable at every step.