~5m36:40
David Shapiro

Why everyone lies about AI

Jul 29, 2026

Read: ~5m · You save: 32 min

Why everyone lies about AI

Shadow AI is spreading through every level of your company. Learn why bans fail and how to turn hidden AI use into a safe, productive, well-governed advantage.

Overview

Employees across industries are using generative AI tools, whether or not their organizations have approved them. This practice, referred to as shadow AI, creates risks and opportunities. This article summarizes the expert discussion on the causes of shadow AI, the postures organizations adopt toward it, and a recommended approach for bringing it under governance.

Defining Shadow AI

Shadow AI borrows its concept from shadow IT, which refers to technology used within an organization without formal approval. Shadow AI extends this pattern to generative AI. Chatbots and, more recently, AI agents form the newest layer of unapproved technology in the enterprise.

The discussion distinguishes two linked but separate activities that organizations often confuse under the single label "AI transformation":

  • Intelligence deployment: Providing employees with access to frontier models to make them more productive, creative, and capable.
  • Automation: Using technology to automate processes, remove bottlenecks, and eliminate manual work. Automation can but does not need to involve generative AI. Automation experts generally prefer predictable results over generative outputs.

Scale and Incentives

Survey figures on AI use vary widely. Depending on the survey, between 40% and 80% of employees use AI, whether or not it is approved. The speakers noted that some surveys lack robustness and rigor. Their own observations from conversations with CTOs, CEOs, vice presidents, directors, and other staff point to use at every organizational level.

Several patterns emerged:

  • Senior executives are among the most frequent users of unapproved AI, according to some studies.
  • Middle managers use AI to keep pace, as some organizations reduce middle management roles.
  • Individual contributors use generative AI to outperform peers or to protect their positions.
  • Concealment occurs at every level. Surveys cited figures of 26% of individual contributors who would use unapproved AI even when it is explicitly banned, and 46% under a different measure.

The speakers identified the incentive structure as the main driver. Employees want to keep up with peers but fear punishment for the associated risks.

Organizational Postures Toward Shadow AI

The speakers described four postures. They identified the most common posture as the most problematic.

PostureDescriptionAssessment
Don't ask, don't tellLeaders know employees use AI and take no action until an incident occursMost common; not optimal
Blanket banProhibition enforced through restrictionsCreates willful blindness; use continues
Mandatory quotasAI use required for all employeesIncreases workload; rewards low-quality use
Champions and evangelistsAdvocates who lead adoptionRecommended approach

Risks of Unmanaged Use

The speakers identified two primary risks:

  • Data breaches. The speakers cited an average cost of $670,000 per data breach.
  • Hallucinations. Inaccurate AI-generated content has caused significant harm. A major consulting firm delivered a report containing hallucinations to a large client, refunded the client's entire fee, and corrected the report at its own cost. Law firms and airlines have also faced lawsuits over hallucinated content.

Recommended Approach: Balanced, Approved Use

The speakers recommended approved use that is safely constructed and tied to defined business purposes. Two points shaped this recommendation:

  • Providing enterprise-wide access to costly capabilities such as Microsoft Copilot can drive prices sharply higher without a proportional gain in value. Some large organizations have experienced this outcome.
  • Leaders who benefit most from frontier models should have safe access first, without the cost of extending the same capability to the whole enterprise.

Leading from the Top

The speakers stated that AI is too important to delegate from the chief executive to the IT function. They recommended the following sequence:

  1. Build executive fluency. Leaders, including board chairs, directors, and CEOs, should first develop personal understanding of frontier models, including their power, potential, and limitations. The speakers noted that this step requires a mindset shift rather than formal training and can happen quickly.
  2. Decide who should be empowered. Leaders should determine which roles require access and cascade fluency downward.
  3. Make the architectural choice. Organizations must decide whether to build or subscribe to AI capabilities.
  4. Establish policies. Organizations should define approved-use policies internally and guiding principles that they can publish externally.

The speakers described a pattern in which an executive who uses AI personally and sees its benefits champions it to the next leadership layer. They described this as a domino effect that occurred over a period of more than one year across organizations they worked with.

The speakers also noted that choices about data residency and sovereignty differ for companies inside and outside the United States. They stated that these factors can be addressed.

Five-Step Framework for Bringing Shadow AI into the Open

The speakers proposed five practical steps:

  1. Amnesty. Employees who disclose AI use should not face punishment unless their use causes significant financial or reputational harm. Amnesty encourages disclosure.
  2. Survey. Organizations should run genuinely anonymous surveys, not pseudo-anonymous ones, to identify which tools employees use and why. Amnesty can be incorporated into the survey design.
  3. Governance policy. A clear policy should specify approved uses, prohibited uses, and the process for approving new use cases. Middle managers can serve as lead champions only when they have such a policy to reference.
  4. On-ramps. Organizations should distinguish three requests: a tool, a process, or a use case. Approving a tool is a software-as-a-service decision, but it also affects people, processes, data, and other organizational elements.
  5. Onboarding. Organizations should distribute the survey, train employees on the governance policy, and clarify rules. One approach is a "bring your own AI" policy that lists the data that employees may never enter into personal AI tools and the data they may use.

The speakers advised against excessive policy length. They cited internal use policies of 15 to 50 pages and recommended a one-page document for a medium-sized business that states what is allowed and what is not. They noted that understanding how employees use AI can generate ideas for improving the business.

Starting with Personal Devices

The speakers described starting executive engagement on a personal computer and phone. Organizations typically have not approved frontier models for enterprise use at that stage. Some have enabled Microsoft Copilot or are piloting tools without enterprise versions of ChatGPT or Claude. The speakers described configuring a personal AI project that accesses only publicly available documents and functions as a chief of staff. They described this as the start of an experience that shows what additional intelligence can do when properly configured.

Emotional Progression of Leaders

The speakers described a recurring sequence of emotional responses among executives and business leaders:

  1. Skepticism
  2. Excitement after using the tools regularly
  3. Fear or anxiety about the scale of change
  4. Paranoia upon realizing that many employees already use AI and that leaders lack full knowledge of that use
  5. Renewed excitement once leaders engage with advisory support

Conclusion

The speakers presented shadow AI as a widespread condition that organizations must address rather than suppress. Their recommended path combines executive fluency, a clear governance policy, anonymous discovery, defined on-ramps, and approved use tied to business purposes.

Introduction and Mission

Introduction of the podcast, hosts, and mission to help business leaders navigate the age of intelligence and automation.

  • Podcast is titled Critical Path and focuses on what executives and leaders can do across industries in the face of generative AI.
  • David Shapiro has about 20 years in technology: first 15 in IT infrastructure and cloud automation, last five in generative AI.
  • Dalabore Petravvic is a retired senior partner at Deloitte with more than 30 years of industry experience.
  • The mission is to provide clarity and confidence to business leaders, aiming to make organizations more intelligent, more automated, and still autonomous.
  • Shapiro's personal mission is building an automated future, with a focus on post-labor economics.
  • The first topic is shadow AI, borrowed from shadow IT.

The Scope of Shadow AI

Shadow AI is the next layer of shadow IT, driven by employees using unapproved generative AI, especially chatbots and agents.

  • Shadow AI is defined as generative AI in the form of chatbots and agents used outside approved channels, extending shadow IT.
  • Survey estimates range from 40% to 80% of employees using AI, whether approved or not; the hosts caution that some surveys lack rigor.
  • Usage occurs at every level, including C-suite, middle management, and individual contributors; some studies suggest C-suite is the biggest transgressor.
  • Some surveys cited: 26% of individual contributors would use unapproved AI even if banned, and 46% or similar figures.
  • Employees fear falling behind but also fear punishment for risks involved.
  • Dalabore notes large public sector organizations (50,000+ people) and small private companies both show high AI use.

Inviting AI Properly

Organizations should invite and properly channel AI use by separating intelligence deployment from automation.

  • The hosts argue: You cannot fight this. Invite it. Invite it properly.
  • Two linked but separate activities: deploying additional intelligence to people, and automation of processes.
  • Deploying frontier models enhances productivity and creativity; automation removes bottlenecks and manual work and should be predictable, often not generative.
  • Shadow IT occurs because people circumvent lack of rules or overly strict rules to access intelligence.
  • Goal: give people access to intelligence, surfacing usage without penalty and deciding who should access safe, protected, orchestrated frontier models.

Postures Toward Shadow AI

Common corporate postures toward shadow AI are flawed; the recommended posture is balanced, approved, and champion-led.

  • Most common posture is don't ask, don't tell: executives ignore usage until something breaks, such as a data breach or hallucinations.
  • Data breaches are cited as costing about $670,000 on average.
  • Deloitte had a widely reported incident where a report with hallucinations led to refunding a major client's entire fee; law firms and airlines have also been sued over hallucinations.
  • Blanket bans are ineffective and amount to willful blindness.
  • Quotas increase workload and incentivize bad behavior; one example mentions a $40 per month Copilot seat.
  • Recommended posture: champions or evangelists who advocate AI after becoming fluent.
  • Executives must personally become fluent in frontier models; AI should not be delegated by the CEO to IT.
  • Balanced response: give approved, safe access to leaders who genuinely benefit, avoiding costly enterprise-wide licenses where value is low.
  • Examples exist of large organizations extending Copilot enterprise-wide with prices rising and value not matching.
  • Sponsor segment for a hat company was excluded as unrelated promotional content.

Five Steps to Invite and Corral Shadow AI

Five steps to bring shadow AI into the open: amnesty, anonymous surveys, governance, on-ramps, and onboarding.

  • Step 1, amnesty: employees who speak up about AI use should not be punished, though serious cost or reputational harm is not forgiven.
  • Step 2, anonymous survey: must be genuinely anonymous, not pseudo-anonymous, to reveal which tools people use and why.
  • Step 3, governance policy: a clear policy stating approved and disapproved uses, and how to get new use cases approved; middle management acts as lead champions.
  • Step 4, on-ramps: distinguish requests for a tool, a process, or a use case; approving a tool like ChatGPT also affects processes, tools, and data.
  • Step 5, onboarding: send surveys, train on governance, and consider bring-your-own-AI policies with clear restrictions on what may never be entered into personal AI.
  • Policies should be short; a medium business needs a one-pager, not a 15 to 50 page document.
  • Data residency and sovereignty matter, especially for organizations outside the United States.

Executive Starting Point and Emotional Arc

Executives should start with personal AI use on home devices, using a personal chief-of-staff project with public documents, then cascade adoption downward.

  • Executive engagement typically begins with personal home computers and phones, because enterprise frontier model access is often not yet approved.
  • Suggested setup: configure personal projects as a chief of staff using only publicly available documents.
  • Dalabore states there is no time to wait and that the domino effect occurs when the owner adopts AI and champions it to the next layer.
  • Emotional arc often seen: skepticism, excitement, fear or anxiety, paranoia, then renewed excitement once working with advisors.
  • Goal: illuminate shadow AI into sunlit, well-balanced AI that makes financial sense; not every employee may need deep access yet.

Closing and Contact Information

Hosts close with contact information, resources, and the Summit Leadership Alliance.

  • Dalabore Petravvic can be found on LinkedIn under his name; he runs the Summit Leadership Alliance, an expert network of senior leaders.
  • David Shapiro's main YouTube channel is Dave Shapiro; Dalabore has the YouTube channel and Substack Tectonic Conversations.
  • A white paper titled the definitive shadow AI July 2026 edition is available to Patreon and Substack subscribers.
  • Email is provided for consulting inquiries.
  • Shapiro's L0 book is referenced as addressing societal and governance interventions.