CISA's Agentic AI Checklist: What SMBs Must Do Before Year-End • Answerbot

Brian Rozelle

Digital Transformation and AI Strategy Consultant

Founder, Rozelle.ai

Published November 2, 2026

In May 2026, six intelligence agencies published a 30-page warning about agentic AI. Six months later, adoption surged but governance lagged behind.

SMB agentic AI adoption hit 13% in 2026, an 86% jump from the year before. But 80% remain in experimentation, 43% of projects have been abandoned, and 71% launched with no formal governance framework.

Two compliance deadlines arrive January 1, 2027. Colorado’s Automated Decision-Making Technology requirements and California’s CCPA ADMT provisions both take effect. The EU AI Act’s transparency duties are already in force. Year-end is the last window to get ahead of requirements that will not wait for your business to catch up.

Why Year-End Is the Deadline That Matters#

Small business owners do not need another task in Q4. But the checklist below is built around time blocks: 30 minutes, one hour, two hours. Not weeks of implementation. The federal guidance is not asking you to build a new department. It is asking you to know what AI tools are running in your business, grant them only the access they need, and write down what they did so someone can explain it later. That is a 30-minute audit and five common-sense controls extending practices you already apply to any SaaS tool.

January 1 matters because compliance architecture built in November is far easier than compliance architecture built in January under enforcement pressure.

The 7-Item Checklist: What to Complete Before December 31#

1. Run a Shadow AI Audit (30 Minutes)#

You cannot govern what you do not know exists. Shadow AI is the norm: employees use personal ChatGPT accounts, browser extensions, and unapproved plugins for work tasks. The shadow AI audit is Step 1 because everything else depends on it. Walk the floor. Ask. Check browser extensions. Review SaaS spending for AI subscriptions. Document what you find.

Time: 30 minutes. Output: A list of every AI tool touching your business data.

2. Scope Every AI Integration (45 Minutes)#

Grant the minimum access required, and document what you granted. For each tool on your audit list, ask: what data does this tool actually need? If an AI assistant only needs to read help documentation, do not give it write access to your customer database. Review these permissions quarterly. This maps directly to CISA’s privilege risk category. Permissions tend to expand over time as you connect more tools. Stop the creep now.

Time: 45 minutes. Output: An access scope document for every AI tool.

3. Build a Human Override Into Every Workflow (30 Minutes)#

AI should never have the final word on actions that affect money, customer data, legal obligations, or security. Build an approval step into any workflow where an AI initiates a transaction, sends an external communication, or modifies a record. Verizon calls this the “crawl” phase: grant only needed permissions, and have a human review every action before expanding autonomy. Sixty percent of organizations lack a kill switch for misbehaving agents. Do not be one of them.

Time: 30 minutes. Output: An approval checklist for AI-initiated actions.

4. Sanitize Inputs and Validate Outputs (1 Hour)#

Prompt injection, hiding malicious instructions inside documents or web pages that your AI reads, is the most common attack vector. Configure your AI tools to reject requests from untrusted sources. Do not let agents browse arbitrary URLs without validation. Review AI-generated outputs before they reach customers, especially for high-stakes communications. The data protection controls that work for shadow AI also work here.

Time: 1 hour. Output: Input restrictions and output validation rules documented.

5. Start Logging What the AI Did (1 Hour)#

Maintain records of what your AI tools accessed, what actions they took, and who authorized them. You do not need enterprise software. A shared spreadsheet or documented approval chain can suffice at small scale. Thirty-three percent of organizations lack evidence-quality audit trails. Log enough to answer: what did the AI do, when, to what data, and with whose permission?

Time: 1 hour. Output: A simple logging template or documented approval chain.

6. Draft an AI Acceptable Use Policy (2 Hours)#

You do not need a 20-page legal document. You need a one-page policy that says: here is what AI tools we approve, here is what data they can touch, here is who needs to approve actions, and here is what happens if something goes wrong. The 71% of deployments with no governance framework are not missing legal expertise. They are missing a decision about what is allowed and what is not.

Time: 2 hours. Output: A one-page AI acceptable use policy, signed and stored.

7. Calendar Your First Quarterly Review (15 Minutes)#

Governance is not a year-end task. It is a recurring practice. Schedule your first quarterly AI governance review for March 2027, before the urgency fades. The CISA playbook covers the five risk categories in detail if you need reference material.

Time: 15 minutes. Output: A calendar invite for your first quarterly review.

”This Week” Implementation Order#

If you have two hours this week, do items 1 and 2: the shadow AI audit and the access scope document. If you have four hours, add items 3 and 4: the human override and input/output validation. If you have a full day, complete items 5, 6, and 7: logging, policy, and the quarterly review calendar. Total time to complete the full checklist: one focused work day, spread across a week if needed. The point is to start before the deadline forces you to.

What NOT to Do: Myths That Waste Your Time#

“We do not use AI agents, so this does not apply to us.” False. If anyone on your team uses ChatGPT, Copilot, Claude, or any AI assistant connected to work data, you have agentic-adjacent risk. Shadow AI is the norm.

“Our vendor handles security.” The SearchLeak vulnerability proved otherwise. In June 2026, researchers showed that a single crafted link could trick Microsoft 365 Copilot into searching a victim’s emails and files, then exfiltrating data through Microsoft’s own infrastructure. Vendor trust does not eliminate operator responsibility.

“We need a security team to implement this.” The CISA guidance explicitly states that you should integrate agentic AI into existing security governance, not create a new discipline. The checklist above extends what you should already do for any SaaS tool.

“Year-end is too busy for AI governance.” The compliance deadlines make year-end the ideal time. A 30-minute audit and five common-sense controls put you ahead of 71% of organizations with no governance framework.

“This is guidance, not law. We can wait.” While the CISA document is not currently enforceable, coalition guidance from five Western security agencies is a reliable leading indicator of binding requirements. The disclosure requirements already in effect under the EU AI Act show where this is heading.

What Changes in January 2027#

Three shifts take effect January 1:

Colorado ADMT requirements apply to businesses using automated decision-making tools that affect consumers. If your AI makes decisions about credit, pricing, or service eligibility for Colorado residents, you need transparency documentation and appeal mechanisms.

California CCPA ADMT provisions extend existing privacy requirements to automated decision-making technology.

EU AI Act transparency duties entered force in August 2026. If your business serves EU customers or processes EU resident data with AI tools, transparency obligations apply regardless of your location.

Sources#

“Ready to implement this?” Get the templates, checklists, and step-by-step guides at Rozelle.ai ↗ , everything you need to move from reading to doing.