For Microsoft 365 administrators, the important distinction is between an approved tool and an approved use. The county’s reported authorization covers specifically permitted internal work through county-managed Copilot in its Microsoft 365 Government Community Cloud environment. It is not blanket permission to feed every county record into a chatbot.
Three risk levels determine what employees can submit
Decorah News reports that the policy divides information into three categories, with progressively tighter restrictions as sensitivity increases.
| Information category | Examples reported by Decorah News | Permitted approach |
|---|---|---|
| Low risk | Public information and general material without sensitive county records | AI may support research, brainstorming, summaries and general communications connected to county duties. |
| Moderate risk | Nonpublic routine emails, internal documents, meeting notes and reports | Only county-approved tools and specifically authorized uses are allowed. County-managed Copilot is the named option. |
| High risk | Social Security numbers, health records, confidential personnel information, sensitive financial information and law-enforcement records | A tool needs specific approval for the information involved, with appropriate security and privacy safeguards. |
The high-risk category explicitly addresses HIPAA-protected information and Criminal Justice Information Services requirements. Approval for one kind of protected information does not automatically cover another, according to the local report. If an employee cannot establish that a tool is approved, the instruction is to withhold the information.
That makes the moderate-risk category particularly important. “Internal” is not another word for “harmless,” and a routine document can contain information that changes its classification. As a practical interpretation of the reported framework, employees need to assess the actual contents—not simply the file’s title or the department that created it.
Why the county-managed Copilot account matters
Decorah News says employees must use county-managed accounts for the authorized Copilot work and contact IT for access. The policy covers both standalone AI applications and AI features embedded in other software.
Microsoft’s own documentation helps explain why account choice matters. Microsoft Support states that Copilot Chat provides enterprise data protection when users sign in with their work or school account. It also says prompts and responses are not used to train foundation models, while interactions are logged and can be available to administrators through search and audit tools. These are product protections, not evidence of the county’s particular configuration.
Microsoft Learn further explains that Copilot respects organizational permissions and administrative settings, with available controls varying by subscription. Its documentation also distinguishes web-search queries from other Copilot processing: the Bing search service has separate data-handling terms.
The practical conclusion is that “not used for training” does not mean “approved for every record.” Microsoft’s contractual and technical protections are one part of the assessment; the county’s authorization for a particular information category is another.
For administrators applying a similar framework, the useful question is therefore not merely, “Is Copilot enabled?” It is, “Which account, which information, which task and which approval?”
AI can assist, but cannot make consequential decisions
According to Decorah News, the county prohibits AI from independently making important decisions involving employment, discipline, benefits, legal matters, law enforcement or access to county services. Employees or officials must review the relevant information and make the final decision.
The reported policy also requires staff to verify AI-generated work, including facts, calculations, legal information and citations. Generated material may be inaccurate, incomplete, outdated, biased or fabricated.
Those requirements address different problems. Human decision-making preserves accountability; verification addresses whether the underlying material is trustworthy. A person clicking “approve” is not much protection if nobody checks the evidence first.
The local report also describes prohibitions on deliberate misinformation, unauthorized impersonation, bypassing security controls, exposing passwords or other credentials, improperly disclosing confidential information, and violating copyright, licensing requirements or applicable law. It says the policy bars Iowa-restricted tools, including DeepSeek, citing the governor’s February 19, 2025 directive.
Iowa’s statewide policy offers context—not extra county rules
Iowa’s Department of Management publishes a separate generative AI policy requiring human evaluation, accountability and restrictions on personal accounts for covered work. It also requires prior written authorization before covered entities submit sensitive or protected information to AI systems. The state document says it binds specified “Supported Entities” and serves as best-practice guidance for others; that wording does not establish that Winneshiek County falls within its binding scope.
The comparison is useful because both approaches emphasize controlled data use and human responsibility. Their approval processes should not, however, be combined into an imaginary statewide-and-county rulebook.
Cybersecurity access is a separate question
The board discussion also covered potential AI-assisted cybersecurity work, according to Decorah News. System Administrator Sherman Watts described access to advanced models for vulnerability testing and assessing county security, saying the county had been approved for an opportunity available to state and local governments.
The report does not identify those models. Nor does it establish that this access authorizes processing protected county information. That distinction matters: permission to evaluate a security tool is not automatically permission to give it confidential records.
Decorah News reports that the employee policy takes effect upon adoption, requires annual review and provides for required training. The county may restrict or discontinue unsuitable tools, while violations can lead to loss of AI access or disciplinary action. Written exception requests go to the board or its designee, and cannot authorize unlawful activity.
The policy’s strongest feature is its separation of three questions that are too easily blurred together: Can this tool be used, can this information be submitted, and can this output be trusted? For county employees—and enterprise IT teams considering similar rules—each needs its own answer.