83 Percent of Organizations Cannot See What Their AI Agents Are Doing
Eighty-three percent. That is the number from a 2026 Marber Security report that should keep every business owner awake tonight: 83 percent of organizations have zero visibility into what their AI agents are doing, and 86 percent lack visibility into their AI data flows.
Meanwhile, 1 in 3 enterprise employees now uses an AI assistant daily.
Put those two numbers together and you get the clearest picture of the AI risk landscape in 2026: AI adoption is outpacing visibility by a mile, and no one is checking.
What Zero Visibility Actually Means
It does not mean the agents are malicious. It means the organization simply does not know what those agents are doing.
When an AI agent is connected to your CRM, your email system, your document store, or your customer database, it can:
- Read sensitive information (client details, financial records, internal communications)
- Write or modify records (update client statuses, send emails, create documents)
- Call APIs on your behalf (query databases, trigger workflows, initiate payments)
- Route data to third-party providers (the AI model itself, connected tools, integrations)
None of this requires the agent to do anything wrong. It just requires the agent to exist with those permissions.
The Two Blind Spots
Blind Spot 1: Agent Activity
Who can see an AI agent decision log? Its tool call history? The exact prompts it received and the outputs it produced?
Most organizations: nobody. The agent runs, the user gets a result, and that is the end of the human interaction. But the agent may have called four different APIs, queried three databases, and sent two outputs to downstream tools before the user ever saw the final answer.
No audit trail. No log. No proof.
Blind Spot 2: Data Flow Visibility
Even if an organization could see what each agent did, 86 percent do not track where the data goes. An AI query might:
- Leave the company infrastructure and hit a model provider
- Get cached by a connected tool (ChatGPT plugin, Slack bot, Notion AI)
- Be logged by an analytics platform monitoring the workflow
- Get stored in a shared document or database that other agents can read
Data flowing through AI agents multiplies the attack surface without creating any new audit points. The data is not just in the cloud anymore. It is in a web of interconnected agent interactions that no single person can trace.
Why This Matters for SMBs (Not Just Enterprise)
There is a misconception that agent visibility is an enterprise concern. But SMBs face the same risk, and often more severely:
1. Fewer controls. An enterprise might have SIEM logs, SOC monitoring, and a dedicated security team. Most SMBs have a laptop, a Google Workspace account, and whoever understands it best. An AI agent connected via OAuth to GSuite bypasses all of that.
2. No IT team to catch problems. In a large org, unusual agent behavior might trigger an alert. In an SMB, the only person who notices is the one whose client data got changed, and by then it is already done.
3. Regulatory exposure is the same. GDPR, the Privacy Act, ACCC enforcement. These do not care whether you have 10 employees or 10,000. If you cannot prove what AI agent accessed whose data and when, you are exposed.
The December 2026 Deadline
Here is the concrete deadline every SMB needs to plan around:
December 10, 2026 under the Privacy Act amendments, if AI significantly influences decisions about individual customers, your privacy policy must disclose it. That is three months from today.
To comply, you need:
- An inventory of every AI system influencing customer decisions
- Proof of what those systems actually do (not just what you say they do)
- A mechanism to demonstrate compliance when asked
SMBs without agent visibility are three months from a compliance gap they cannot close.
The Attest Thesis
Visibility is not a governance problem. It is a verification problem.
Governance frameworks tell you what policies you should have. But when an AI agent acts, correctly or incorrectly, the organization needs independent, cryptographically-verifiable proof of what happened. Not a log maintained by the same system the agent controls. Not a dashboard the agent can manipulate. Independent evidence.
That is what Attest is built for:
- Capture -- Cryptographic recording of agent sessions, prompts, tool calls, and outputs
- Attest -- Independent verification of environment state before and after agent runs
- Prove -- Exportable audit bundles that survive the agent being compromised
The 83 percent stat is not a problem that governance frameworks solve. It is a problem that requires a completely new layer: verification infrastructure.
Governance sets the rules. Verification proves they were followed.
FAQ
- question: Can I get visibility into my AI agents right now without expensive tools?
answer: Yes, but it requires deliberate process. Start with an OAuth audit (Google Workspace Admin Console > Security > API controls), build an AI tool register (spreadsheet), and enforce a written AI use policy. These cover baseline governance but lack cryptographic verification.
- question: What is the difference between governance and verification?
answer: Governance tells you what should happen (policies, permissions, rules). Verification proves what actually happened (cryptographic capture of agent sessions, tool calls, and outputs). You need both.
- question: Does the Privacy Act apply to businesses using AI agents?
answer: Yes, if your turnover exceeds AUD $3 million or you handle health/credit data. From December 10, 2026 you must disclose when AI significantly influences decisions about individual customers.