You’re Approving Work You Can No Longer Fully Verify
Last quarter I sat with a CEO I have worked with for years. Mid-sized professional services firm. Smart operator, twenty years building the business. The kind of leader who delegates well, knows his people, and runs a tight ship without micromanaging. He had just walked out of a board meeting that went exactly the way he wanted.
The deck was sharp. The strategy section in particular got real engagement from his outside directors. One of them asked, half-jokingly, who had written it.
He answered honestly. His head of strategy had built it. His operations lead had reviewed it. His marketing person had touched the language. He had signed off on the final version himself. Standard process. The same process he had used for years.
Then he paused. And he said something I have heard now from a dozen CEOs of similarly sized companies, almost word for word.
“I don’t actually know how much of that AI wrote. And I don’t think anyone on my team could tell me with confidence either.”
He wasn’t worried about the deck. The deck was good. He was worried about what that sentence meant.
You have always approved work after the fact. That is what running a company at scale looks like. The shift is not that you have stopped being the decision-maker. The shift is that the work coming up to you now sits behind a layer you cannot question the same way you used to.
When the person responsible for a recommendation hands it to you, you have always been able to ask them to walk you through how they got there. They could. That conversation was the actual quality control, not your signature at the bottom. The signature was the formality. The reasoning trail was the real review.
That trail is starting to break. Not in a dramatic way. In a thousand small ways, across every team, every week.
The number that probably will not surprise you
There is a stat being passed around right now that gets a lot of attention: 62.6% of Gen Z workers submit AI-completed work as their own (WalkMe, 2025). Most leaders read that and shrug. Their interns probably do it. Their junior associates probably do it. Their HR team will figure it out.
Fair. That stat alone is not where the story is.
The same survey found that more than half of senior leaders report not consistently disclosing how AI shaped the work they sign off on (WalkMe, 2025). The strategy memo. The investor update. The board paper. Something in there came out of a model. The person who reviewed it does not always know. The person who wrote it does not always say.
This is not a junior employee problem. It runs top to bottom.
When 78% of employees are using AI tools their employer never formally approved (WalkMe and SAP, 2025), and a meaningful percentage at every level are not disclosing it, the issue is not adoption. You have adoption. The issue is that nobody can fully reconstruct where the work came from anymore. You can still verify the output. You can still verify the outcome. What is harder to verify is the reasoning.
You probably already feel this. You may have explained it away. Most leaders do.
What a normal week looks like now
Let me describe what I see in companies your size. You can decide how much of it sounds familiar.
A senior associate has a client memo due. She is good at her job. She has been with the firm for six years. She is not careless. She has been to your AI training session.
She opens her personal ChatGPT account on her phone. She pastes in the meeting notes and a few paragraphs from the prior memo. Ninety seconds later she has a working draft. She rewrites it in her own voice, cleans up two sections that did not quite land, and sends it to her director. He reads it, makes one comment, and forwards it to the client.
Nothing about that workflow looked unusual to her. Nothing about it looked unusual to her director. The memo was good. The client was served. By any measure she has been trained to recognize, she did her job well.
In firms without tight controls on personal device AI use, three things tend to happen here that no one logs. First, internal information moves through an account the firm does not control. Cyberhaven’s research found that 73.8% of ChatGPT accounts used in workplace settings are non-corporate consumer accounts (Cyberhaven, 2025). Her account is one of them, and the data she pasted now sits inside a system governed by terms her firm never reviewed. Second, when she eventually leaves for another firm, her account leaves with her. So does the working memory of every client matter she has touched through that account. The firm cannot retrieve any of it because it never had access to begin with. Third, the next time someone audits how that client’s information has been handled, there is a gap in the chain of custody nobody knew was there.
You may be reading this thinking your team would not do that. Maybe your team is the exception. The honest answer is that even at well-run firms with strong cultures, the data says most teams are doing some version of this. Not because people are reckless. Because the tools are designed to be used the way they are being used, and the work pressure is real.
The cost most teams are quietly absorbing
Researchers at BetterUp Labs and the Stanford Social Media Lab gave one piece of this a name: workslop. AI output that looks finished but needs cleanup before it can actually be used. In their 2025 research, 41% of US desk workers received workslop in the prior month. Each instance took about an hour and fifty minutes to resolve, costing roughly $186 per employee per month on average.
That average will be lower at firms with strong process discipline and higher at firms where AI use is moving faster than the review structure can keep up. At any size, even small amounts of recurring rework absorb into normal time without ever surfacing as a number anyone is tracking. The bigger the company, the more places it can hide.
The reason that number is hard to feel is that it does not show up as a line item. It shows up as a director who feels busier than she used to without producing more. A team that hits its deadlines but is quietly burning more hours per deliverable. Workday’s 2026 research found that 37% of all time saved through AI is immediately consumed by rework (Workday, 2026). The savings show up in one column and disappear from another. Most companies are not tracking the net.
You are not imagining the feeling that your team is working harder than the productivity numbers suggest. That gap has a name now.
This is not the AI doom argument
Before I go further, it is worth being clear about something.
AI is delivering real value right now, in real companies, at real scale. Some firms are seeing genuine margin improvement, faster client turnaround, and better quality of analysis than they had two years ago. The Boston Consulting Group and Harvard study on consultants using AI found 12.2% more tasks completed, 25.1% faster, and 40% higher quality output on the right kind of work (BCG and Harvard, 2024). That gain is real and it is sticky.
The difference between the firms getting that gain and the firms quietly absorbing the cost is not which tools they bought. Both groups are usually running the same software. The difference is whether anyone in the company can tell you, in plain language, where AI is being used, by whom, on what kind of work, and what verification sits behind the output.
That is the gap this post is about. Not whether to use AI. How to see it clearly enough to make it compound instead of leak.
“AI did it” is not a defense, and your policy may not cover it
In 2024, the British Columbia Civil Resolution Tribunal heard a case against Air Canada. The airline’s chatbot had given a passenger incorrect information about bereavement fare refunds. The passenger booked a ticket based on what the chatbot said. Air Canada later refused the refund and argued that the chatbot was a separate legal entity responsible for its own statements.
The tribunal called this a “remarkable submission” and held Air Canada fully liable. The principle is straightforward and now well established: the company owns the output, regardless of which part of the company produced it.
That principle applies inside your business right now. A client memo built partly by AI, sent over your letterhead, with your partner’s sign-off, drawing on data the firm was contractually obligated to protect. If something in it is wrong, or if that data shows up somewhere it should not, the firm owns the consequences. The vendor does not. The employee does not. The business whose name is on the document does.
Your general counsel may have updated your AI policy in the last year. Most have. The challenge is that most of those updates were written before the pattern of personal-account use at this scale became visible. The policies generally cover approved tools and approved workflows. They do not always cover what is actually happening on personal phones, in personal accounts, with company information, in the seconds between a deadline and a deliverable.
That is worth a conversation with your GC this quarter. Not a panic. A conversation.
Why you may be the last to see it clearly
A reasonable question at this point is: if any of this is real, why hasn’t somebody on my team raised it?
The honest answer is that the people closest to it usually cannot raise it, because they cannot see it as a pattern. The associate using her personal account does not think of it as a policy issue. She thinks of it as getting her work done. Her director did not know AI was involved on that particular memo, because she did not flag it, because there has never been a clear conversation about what should be flagged. Your IT team is monitoring the tools you approved, which captures roughly a quarter of what is actually happening. The leaders who might raise it are sometimes the same ones the survey data says are quietly using AI themselves without disclosing it.
You are not the last to know because your team is hiding things from you. You are the last to see the full picture because the picture sits across every team, every tool, every account, and no single person in the company has the view from above. That view is your job. Nobody else is positioned to do it.
This is the AI Blind Side. Not the parts of AI you intentionally rolled out. The parts that filled in around them while you were focused on the rollout.
What it looks like when this is handled well
The firms that are actually pulling away from their peers on AI right now do not look dramatically different from the outside. Their offices are not full of dashboards. Their teams are not policing each other. The difference is quieter and harder to copy.
Three things tend to be true inside those companies. First, every client-facing AI output has a named human owner before it goes out the door, and that owner can answer the question “walk me through how you got here” without flinching. Second, leadership can tell you, on a Tuesday morning, which AI tools the company is actually using, on which categories of data, with which level of approval behind them. Not a perfect list. A real one. Third, when a workflow needs to be shut down or changed, somebody knows exactly which workflow it is, who built it, and how long it takes to unwind. Usually under an hour.
None of those three are about the technology. They are about ownership, visibility, and the ability to act quickly when something needs to change. The technology was the easy part. Most of these firms had it adopted within a year. The operating system around it is what took the second year. That is the part most companies have not built yet, and it is the part that decides who compounds and who leaks.
What to do this week
You do not need a new platform. You do not need to ban anything. You do not need to call an all-hands meeting. What you need first is a clear picture of where AI is actually being used inside your business, by whom, on what kind of work, with what level of verification behind it. That picture takes about four minutes to start drawing.
The AI Blind Side Assessment is 12 questions, written for leaders of 15 to 250 person companies who have already adopted AI and are starting to feel the gap between what was deployed and what is happening on the ground. No email required. No sales call. The questions are built to surface where your visibility is strong, where it is thin, and where the exposure is sitting that nobody has named yet. You get your result immediately.
Most leaders are surprised by at least two of the answers. Some are surprised by more.
The companies that win the next phase of this will not be the ones using the most AI. They will be the ones who can tell you, in plain language, what AI is doing inside their business and what it is not. That work starts with seeing it clearly.
Before spending money on tools, policies, or consultants, most leaders need one thing first: a clearer picture.
Take the AI Blind Side Assessment →
David Stanbridge is the author of The AI Blind Side and founder of WinAt.ai. Over the last three years, he has worked with and studied companies adopting AI early, focusing on the gap between visible gains and what is actually happening underneath.
