Your AI Governance Stops At The Screen: The Blind Spot In AI-Driven Document Workflows
Corey Ercanbrack is CTO at Vasion, leading the AI-native transformation of a cloud-native platform under FedRAMP High.
gettyA state agency mails benefits notices to citizens, with personal data on the page and legal rights tied to what’s printed. When I asked the administrator responsible for its 30,000 endpoints what governs that output process, he said it runs elsewhere, on infrastructure owned by a different team.
He wasn’t dodging me. That’s how these environments get built, and it produces a blind spot most leaders inherit rather than create. Information is governed where the data lives. Output is governed where the devices live. Those are two teams with different budgets, reporting lines and definitions of accountability, so the artifact carrying the heaviest legal obligation sits furthest from the controls protecting the information inside it.
The same split appears across other organizations I work with. Major American cities operating under state public records law must produce any official communication upon request, whatever the medium. Their digital channels are instrumented. Paper isn’t. The specialist responsible for its endpoint estate acknowledged that if he sent a printed memo, someone would still have to find it.
That gap has been survivable because a person stands in it. The employee who printed the memo knows he printed it, and if a request arrives two years later, someone can ask him. In audit terms, he’s a compensating control: what you document and test when the designed control can’t be implemented. Nobody documented this one. It’s been load-bearing for 30 years without appearing in a control matrix, which is why removing it looks costless.
Agents remove it. Deloitte’s 2026 “State of AI in the Enterprise” report surveyed 3,235 business and IT leaders across 24 countries. Nearly three in four companies plan to deploy agentic AI within two years, while one in five reports a mature governance model for autonomous agents. Among the capabilities Deloitte names as crucial are audit trails capturing the full chain of agent actions.
When an instruction comes from a system rather than a person, there may be no human intent to reconstruct later and no awareness that a records obligation was triggered. Unless the workflow was designed to capture that context, automation can generate unaccounted output far faster than the informal human process it replaced. Leaders may be automating away a control they never knew they depended on.
Most AI governance programs are scoped to the application: which model acted, what data it reached, what it was authorized to do and what it produced. That work is necessary, but it is not sufficient, because operational documents don’t stay inside one application.
I have watched a prior authorization request move from a medical record to paper, to a fax machine, to an insurer and return days later as an email. Each system recorded its portion and completed the step it was designed to handle, but none held the thread connecting them: the only view from which the organization could reconstruct the full event. The application finished its task. The business did not, and governance was scoped to the first while the obligation attaches to the second.
An IT leader at a global distributor gave me a paperless example. More countries where his company operates now require documents to be deposited with a government authority. The regulation specifies where the document must be sent without specifying how it gets there. Treating the output boundary as a printing problem misses the broader obligation to govern the document throughout its journey.
If an auditor picked up an AI-generated document tomorrow, what could you prove about it? Four questions help determine the answer.
1. Identity: Who or what produced it? Once a document is printed, exported or faxed, the chain of identity breaks. The artifact no longer shows which agent created it, who initiated the action or where the information came from.
2. Authorization: Who approved its release? Knowing who generated a document is not the same as knowing whether release was authorized. A finished document proves neither.
3. Audit: What happened between input and output? An agent extracts information, changes content and generates instructions. Record only the finished document, and you can’t show what it did. The record has to replay what went in, what changed, who acted and where it went.
4. Retention: Can you find and classify it? Without the first three, you accumulate documents you’re obligated to protect and unable to trace, turning a governance gap into a storage liability.
Miss one, and you don’t have end-to-end AI governance. Instead, you have a partial record of an action that continued without you. I recommend running these four against a live workflow before expanding agent authority. The gaps you find are the ones an agent will only widen.
The instinct is to connect every system, but these integrations are often fragile and require ongoing maintenance whenever one side changes.
Start with the work instead. Pick one operational document this week and follow it from creation to final disposition. Mark every point where somebody prints, exports, faxes, signs, scans or retypes information, and keep going after the data leaves the application. The handoffs are where governance stops without announcing it. Test the four questions at each one. Wherever context breaks, you’ve found the real boundary of your program before an auditor does.
The workflows that hold up often share one trait: The record was designed before the agent was deployed, rather than reconstructed after someone asked for it. That means someone decided where human review belonged and what needed to be recorded at every step, including the agent’s actions. You can’t deploy first and gather the evidence later. The record has to be created when the action occurs, or it may not exist at all. And the destination is rarely as modern as the agent. I’ve watched sophisticated automation end with a fax because that’s what the recipient accepts.
Governance has to follow the information through every handoff. The practical perimeter is everywhere the information goes.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
