Treat shadow AI as an operating decision that needs visibility and ownership. Find out which tasks employees are trying to complete, what information tools receive and which systems they can access. Pause uses that cross an unresolved boundary, while giving useful low-risk work a supported route. Approval should specify the task, account, data and permissions, not just the name of a tool.
Key takeaways
- Discover the job employees are doing as well as the software they use.
- Approve a defined use with specific data and permissions, not every capability of a product.
- Make review and support fast enough that employees can use the approved route in practice.
Ivan Nosov, Co-founder & COO, ACW Soft Solutions. Prepared as an editorial Q&A draft for founder review, not as a transcript of an interview.
What would make you investigate informal AI use in the first place?
Investigate when business work starts depending on a tool whose account, inputs or access nobody owns. Shadow AI means work-related AI use outside the company's agreed approval or oversight. The problem can begin with a chatbot, a browser extension or a connection between an assistant and shared business systems. A paid subscription does not establish that the use is approved.
Consider a hypothetical account team using personal assistants to turn meeting notes into client updates. The output may be useful, but the founder needs to know whether those notes contain confidential decisions, whether the assistant connects to a shared drive and who can remove access when someone leaves. These questions concern the work, not the employee's enthusiasm for AI.
Start with a simple inventory: task, tool, account owner, information supplied, connected systems, output destination and person who checks the result. Record uncertain answers rather than treating them as permission. NIST's voluntary AI RMF Playbook provides suggested actions for managing AI risks; it is not a certification or a complete checklist that makes an unknown workflow acceptable.
Sources[1]
How would you ask about personal accounts without losing the team's trust?
Explain what information the business needs and how it will use it. Ask employees to describe the task, the categories of information involved and any connections they authorized. Do not ask them to send passwords, API keys or private chat histories into a shared spreadsheet. Where business content needs investigation, use the appropriate internal process and limit access to the people handling it.
Make the initial discussion concrete. 'How did you prepare last Friday's update, and which step took the most time?' is more informative than asking whether someone is compliant with an abstract AI policy. Show the team an example of an acceptable input and an example that requires review.
Discovery and accountability can coexist. Do not promise that every disclosure will have no consequences, and do not imply that reporting a problem makes the employee responsible for designing the solution. Give the issue an owner and explain what happens next. People need a reliable way to raise uncertainty before more company information is copied or another integration is connected.
Which workarounds reveal a gap worth solving?
A workaround is useful evidence when several people need the same result and the supported process makes it unnecessarily difficult. Repeatedly rewriting client updates may indicate scattered project information. Manually translating product descriptions may indicate a content workflow gap. The tool employees choose is one possible response to that gap, not proof of the right long-term solution.
Follow one task from source to final use. In the account-team example, the assistant may save drafting time while the PM still searches several systems to confirm every commitment. Buying more licenses will not necessarily improve that process. A single current project record and a clear approval step may matter more.
Compare three responses: improve the existing process, enable an approved capability in software already in use, or introduce a new tool for a defined task. Include the people who actually do the work in that comparison. The useful measure is an accepted result with less total effort and fewer unresolved errors. A tool employees enjoy using still needs to fit the company's information and access boundaries.
Which information boundary should pause an experiment?
Pause the experiment when the team cannot establish that the proposed use is permitted for the information or access involved. Examples include uploading customer records to an unreviewed personal account, sharing confidential source code without an agreed basis, or granting an assistant write access to a production system. Resolve the specific uncertainty before continuing that part of the workflow.
OWASP identifies sensitive information disclosure as an LLM application risk and recommends measures including least-privilege access and restricted data sources. Its guidance also notes that instructions inside a prompt are not a dependable boundary by themselves. A reminder telling an assistant to keep information private cannot substitute for account configuration, permissions and a suitable data-handling arrangement.
Approval should cover the actual account and feature: what is sent, what is retained, who can access it, which connectors are enabled and how access is revoked. Check the provider's current terms and settings for that use rather than assuming all plans behave alike. This article proposes an operating review; contractual or regulated-data questions may require the company's relevant specialist.
Sources[2]
Which low-risk use would you support first?
Start with a useful task that can be bounded without sensitive inputs or automatic external actions. One example is drafting an internal summary of an already-public product announcement. The employee checks it against the original source before colleagues rely on it. This is a proposed starting point, not a claim that public inputs make every tool or output harmless.
Create a short approval record covering the tool and company-controlled account, allowed input types, the task, the reviewer and any prohibited actions. For this example, connecting the entire company drive is unnecessary. A later request for that connection is a new decision because it changes the information the tool can reach.
Give users a clear request route when the approved scope is insufficient. Choose a response target the team can actually meet and name a backup reviewer. Revisit the approval when a new connector, automatic-send capability or material provider change affects the use. The aim is to make an appropriate route easy to follow while preserving a deliberate decision about more consequential work.
What would you measure after a month?
Measure whether the company understands the uses that matter and whether supported workflows produce better results. A month is a practical review interval for a small pilot, not evidence that every risk has been discovered. Compare the tasks identified, uses with assigned owners, unresolved access questions and time taken to handle a new request.
Track quality alongside adoption. In the client-update example, sample whether dates, approvals and commitments match the project record. Count corrections and the time spent checking the draft. A high number of AI-generated updates is not a useful result if the PM still rewrites them or customers receive conflicting information.
Interpret discovery numbers carefully. More reported tools may mean employees are being more open rather than that behavior has suddenly worsened. Fewer detected tools may reflect incomplete visibility. Decide which uses to continue, change or stop based on the evidence for each workflow. Expand access only where ownership, information handling and the benefit remain clear; unresolved high-impact access should stay a separate priority.
What should a founder do next?
- Inventory one team's AI-assisted tasks without collecting credentials or unnecessary personal content.
- Review the highest-impact data access and external actions first.
- Publish a short approved-use record and a named route for new requests.
- Review output quality, correction effort and unresolved permissions after the pilot.
Reader questions
Does buying business accounts solve shadow AI?
It can improve account ownership and controls, depending on the product and configuration. It does not define which information belongs in a workflow, who checks the output or whether an employee can connect another system.
Should every AI experiment need the same approval process?
No. Use a proportionate review based on the data, permissions and consequences. A draft made from public information deserves a different assessment from an assistant that can alter customer records or send messages automatically.
Sources and further reading
- NIST AI RMF Playbook (opens in a new tab). NIST · Accessed 6 Nov 2026
- LLM02:2025 Sensitive Information Disclosure (opens in a new tab). OWASP · Accessed 6 Nov 2026
About the author. Ivan Nosov, Co-founder & COO, ACW Soft Solutions. Ivan Nosov is the COO and co-founder of ACW Soft Solutions, where he helps companies build, launch and scale digital products. His focus is the business challenge behind a product: whether a company needs an MVP, a modernized platform, a better user experience, AI or automated processes, the right first step is sometimes defining what to build. He works with founders and CEOs, product and operations leaders, SaaS companies and technology partners. View profile

