Shadow AI in healthcare: closing the convenience gap

Tell us about your project

Please reach out to us about your project and see how we can help you accomplish your goals.

We value your privacy and will never share your contact information

Please note that submissions without a valid business email address or business phone number may not receive a timely response

Healthcare Solutions

If you're in healthcare, you owe it to yourself to learn how you can make your everyday business processes more efficient and save money at the same time. We can help in automating many of your routine and repetitive tasks, including Patient Engagement surveys.

Contact us to learn more

Transportation Solutions

If you're in the transportation business, you can automate many of your routine tasks like package notifications, surveys, collection calls and more. Improve your customer satisfaction by extending your service hours without extending your costs.

Connect with us to learn more

Why small, everyday tasks push staff to unapproved tools and what procurement teams should ask

Many data exposures begin not with a hack but with a person trying to solve a small, practical problem: summarizing a spreadsheet, cleaning up a list, or extracting a few fields for a report. The operational problem is simple. When approved systems are clumsy or unavailable for that kind of quick work, staff reach for open web tools. Recent reporting that Johns Hopkins Bayview notified about 3,182 patients after a worker uploaded a spreadsheet to an unapproved AI tool is a reminder that those convenience-driven choices create real risk.

Stopping shadow AI is not just about bans and retraining. It is about giving teams sanctioned, easy-to-use alternatives and aligning vendor answers with the legal and operational realities your organization faces. We see this pattern often in hospitals and clinics: the policy says one thing, everyday workflows push another way.

What small tasks actually drive people to shadow AI?

Procurement and compliance teams often assume shadow AI is used for big-picture automation projects. The quieter truth is different: people reach for unsanctioned tools for short, ad hoc jobs that approved systems make difficult. Examples include:

  • Quick summarization of a clinic’s weekly admissions spreadsheet so a manager can draft a memo.
  • Cleaning or de-duplicating a patient contact list before a focused outreach.
  • Translating a short set of free-text notes for a clinician who needs a rapid read.

Those tasks share two properties that matter: they are small and they have a tight time window. The fix is not just tighter rules. It is replacing the convenience gap with a sanctioned workflow that preserves privacy and auditability. Buying a bulk AI contract and dropping it in place rarely helps if the tool is not available at the desktop or within the team’s normal workflow.

What should a sanctioned alternative actually do, and what answers to expect from vendors?

When procurement asks vendors for a replacement to the shadow-AI problem, the questions should be concrete. A good answer focuses on how the tool fits daily work, not only on model performance. Reasonable vendor answers will describe how the tool:

  • keeps protected health information (PHI) inside systems that are approved for clinical data handling, and documents that protection;
  • lets a user accomplish the quick task from the same place they already work (a secure inbox, a secure messaging app, or the clinical records system) rather than forcing copy-paste into a web browser;
  • provides a simple, auditable record of who submitted what and why, so compliance teams can reconstruct the chain when auditors ask.

An evasive vendor answer focuses on generalities: “enterprise-grade model” or “we can train on your data” without explaining where the data lives, how staff access it, or what the audit trail looks like. Those gaps are the very openings people use to justify a browser search for a quick fix.

Legal and compliance teams will want references to the Health Insurance Portability and Accountability Act (HIPAA) and to policies that restrict moving PHI to third-party services. For practical context, current HHS guidance on HIPAA explains the covered entities’ obligations and is a useful reference when evaluating vendor claims about permitted data uses. Asking vendors to explain, in plain language, how data is kept in approved boundaries is a more telling question than asking about model accuracy alone.

What operational checks should your call center and notification program pass?

When a disclosure has patient-notification consequences, the communications program becomes the front line. Two operational checks are worth prioritizing.

First, the ability to stand up a staffed, auditable notification channel quickly. In the Johns Hopkins Bayview example, a toll-free line was set up to handle patient questions. That is the right move operationally, but the team answering the calls needs clear scripts, recording and retention controls, and a way to mark calls that require escalation. Ask whether potential vendors can describe how they manage call recording, access, and retention so that each interaction can be reconciled later if regulators inquire.

Second, your notification outreach should not create more risk. Sending messages, offering credit monitoring, and tracking opt-ins all have to be done without scattering PHI into unapproved systems. Inventory the small, repeatable steps your teams take during an incident and push for tools that keep the data inside approved, auditable stores. That work sits squarely in the same wheelhouse as broader security and compliance controls: messages, logs, and access all need protection.

Language access matters here too. If outreach materials, call scripts, or summaries are translated outside approved channels, that creates another leakage vector. Ensuring multilingual delivery and that open-ended translations are handled in an auditable way reduces the impulse to use the one-off browser tool that a staffer knows how to use. See why language coverage changes outcomes in automated communications.

None of this absolves training or governance. Clear limits and disciplinary expectations are necessary. But the durable fix is practical: give teams sanctioned, simple tools for the quick jobs they do every day, and make it easier to comply than to improvise.

That is the bottom line. Reducing shadow AI starts with mapping the small tasks staff perform, then matching those tasks to approved, easy-to-access workflows. Procurement that asks the right questions will spot vendors who understand the problem versus those who sell a vague capability. Policies and training remain essential, but they work only when the approved option is the convenient one.

Related coverage: Johns Hopkins alerts 3,000+ patients after data sent to AI tool – WMAR 2 News Baltimore