Gleb Tsipursky on the psychology of AI adoption, work identity, beyond social stigma, and win-win dynamics (HAI Ep63)

Transcript

Ross Dawson: Gleb Tsipursky, welcome to the show.

Gleb Tsipursky: Thank you, Ross. I appreciate you inviting me.

Ross: You are looking deeply at AI adoption at work, particularly the psychology of it.

What are you seeing happening with AI adoption in the workplace?

Gleb: We have some very concerning results coming out of recent research. A recent MIT study showed that 95% of AI pilots fail to demonstrate a return on investment when scaled up.

Some might look good on paper and when initially launched, but they do not scale to deliver profit-and-loss benefits. And it is not simply this MIT study.

A recent Boston Consulting Group study found that only 6% of enterprises have shown value at scale from AI initiatives. Sixty percent show little or no material impact at all from AI initiatives.

McKinsey reports from its own study that 80% of organizations say individual employees are achieving productivity gains from AI tools. However, only 6% are high AI performers.

Ross: Actually, the larger figure is 37% getting results in terms of EBIT impact.

Gleb: Right. Eighty percent are reporting individual productivity, 37% are reporting some impact on EBIT, and only 6% report real, material EBIT impact.

When we are talking about enterprise value at scale, 37% might have some impact, but not much. Of course, AI costs money as well.

Only 6% are high performers, only 6% have value at scale, and only 5% of AI pilots scale to demonstrate profit and loss. That is very concerning because both you and I, Ross, know the technology is great.

It is not an issue with the technology. The technology is excellent.

The issue is adoption. My book, The Psychology of AI Adoption at Work: From Resistance to Results, addresses this topic.

It is published by Georgetown University Press and is peer-reviewed. It examines the totality of the research and is based on my own research.

That includes more than 100 consulting projects, over 600 focus-group respondents, more than 10,000 survey responses, and over 50 interviews. Those interviews included leaders from Microsoft, Uber, and Experian.

It shows that this is really about human psychology and that leaders are not approaching AI adoption in the most effective way. That is what we are seeing, and that is the problem we are encountering.

We can discuss why leaders are not approaching it effectively, but that is what I am seeing in the marketplace.

Ross: The MIT study has been debunked in many ways and is not particularly robust.

But it is absolutely true that not many organizations are getting value commensurate with their investment. I suppose there is a whole array of layers explaining why that might be.

A very large part is certainly psychological, concerning whether people want to adopt AI and are able to adopt it. But I think the gap is particularly interesting in the McKinsey study.

It found that 80% are getting individual productivity gains, while only 37% are getting EBIT impact. There is a lag involved, but there is also the issue of work redesign.

That was really the primary finding of the McKinsey report: the need to redesign the structure of work. That complements individual uptake and how we get people to respond effectively.

But a lot of it is also about organizational structure.

Gleb: Absolutely, and that is an important insight, Ross. Organizational redesign is only enabled by effective individual uptake.

It needs to be based on effective individual uptake and usage. Organizational redesign is not AI 101; it is AI 201 or 301, so we need to focus on AI 101.

That is human psychology. Let us discuss what is happening in many organizations.

Leaders are thinking about AI adoption as a typical IT project, and they are letting IT take the lead. When I ask organizations who is leading AI adoption, they say, “Well, our IT department is leading AI adoption.”

But what happens with typical IT projects? Let us say you are adopting a new CRM.

You roll out a top-down process in which IT, working with marketing, for example, develops standardized operating procedures. Everyone is trained on the same standardized operating procedures, and you gain value from everyone working in the same way.

If people enter data differently into a CRM, you cannot segment your customers effectively. The resistance to change and change-management issues come from people having to learn a complex new system and change their habits.

With AI, it is very different. We both know that AI is not like a CRM at all, but IT is approaching it like one.

They might get Claude by Anthropic accounts for everyone and provide generic training on using AI and writing prompts. Then they let people figure things out themselves.

They say, “Now we have given you training and an account. We have done our jobs.”

Maybe they also provide some guidelines about what not to do. But then people are left to figure things out themselves because AI is unlike a CRM.

You cannot have standardized operating procedures for AI. Each individual employee needs to determine their own use cases.

They essentially need to build their own CRM and develop their own use cases. It is an incredibly flexible and adaptable tool.

The right mindset is more like hiring an executive assistant for everyone. It is each person’s role to train that executive assistant and essentially build a digital twin for their own workflows.

They must discover how AI can most effectively complement those workflows and how they can work effectively with it. Each individual employee needs to do that separately.

That is a very different mindset. You cannot simply get people to comply with it as you can with a CRM.

You need people to be motivated, curious, engaged, and committed to AI adoption. They need to take initiative and innovate to determine how AI can be effectively adapted to their individual workflows.

That is a very different mindset that many companies do not understand. It is also made much more difficult by the psychological obstacles to AI adoption.

Those obstacles arise because AI is the first technology that challenges our identity as human beings. Think about intelligence: what is it?

Our brain is essentially a biological computer. It is trained on all our past experiences and has a particular model of the world.

It engages with new information to make decisions, take actions, and generate content. The same applies to artificial intelligence.

It is trained on past data, including publicly accessible data on the internet, and has a particular model of the world. Claude 3.5, or whatever it might be, engages with new information, makes decisions, takes actions, and generates content.

It therefore challenges our identity as human beings and what we bring to the table. This is the big challenge for many people.

A recent Pew Research survey showed that approximately 52% of respondents were more anxious than excited about AI tools. Only 9% were more excited than anxious.

Those numbers have been getting worse over time. A similar Pew Research survey in 2024 showed that 50% were more anxious than excited, while approximately 13% were more excited than anxious.

Unlike with other technologies, anxiety has not decreased over time with exposure; it has increased. That is certainly an important insight.

Ross: One very interesting thing is that this differs across countries. The figures you are quoting are for the United States.

There are very different dynamics in other countries, which means there are also different dynamics around AI adoption. The point you are making about this being extraordinarily different from previous technologies is extremely valid.

You cannot simply say, “This is a tool. Here is the manual; use it.”

It requires active engagement. As you say, everyone now has technology equivalent to assistants or minions that can work for them.

But the larger frame, as I see it, concerns leadership attitudes and two issues. Do leaders genuinely want to amplify and augment their workforce to build greater productivity?

Or do they want to become more efficient and cut costs, potentially including reducing headcount? One issue is the reality of the leadership’s attitude.

The other is what people in the organization believe. That is often correlated with what leaders are actually thinking and doing.

Sometimes leaders can have better intentions, but people are still very worried. Concern about the future of their jobs is central to that anxiety.

Gleb: You are absolutely right, Ross. From my research, the demographic profile you have identified is the largest group in corporations.

For this group, fear of job loss is the primary economic source of resistance. I call these people the AI alarmists, and it is not as though they have no reason to worry.

A recent Stanford study called Canaries in the Coal Mine examined the most AI-exposed industries relative to the least AI-exposed industries. In January 2025, it found a 15% employment gap among young professionals aged 22 to 25.

By June 2026, the same figure was a 19% employment gap. Things are therefore getting worse in terms of employment.

A recent letter was also signed by 16 Nobel laureates—

Ross: Because younger people could be affected differently. In fact, there are positive trends for more senior people.

Gleb: Yes, but I am talking about people aged 22 to 25. Those are entry-level workers, and they are currently very endangered.

A recent letter signed by 16 Nobel laureates said that we will face massive job displacement from AI tools over the next several years. Of course, we are already seeing that displacement happen, especially in technology companies.

Microsoft has just offered buyouts to a quarter of its workforce. That is something coming down the pipeline for other industries because Microsoft and other technology companies are at the forefront of AI adoption.

This is very concerning, and people have reasons to be concerned. However, there are ways to address that concern.

What you discussed earlier is one aspect of how I work with leaders to determine how to address it. Another Stanford study examined the five-year impacts of AI adoption at companies.

It found that the most AI-forward companies had approximately 9% faster revenue growth than the least AI-forward companies in their industries. They also had 6% faster headcount growth on average over a five-year period.

That tells us something important. We are not only talking about the technology sector; this applies across broader industries.

Overall, companies adopting AI and focusing on growth are doing more with less, but they are also taking market share from less effective companies. They consequently grow and need employees.

A company can decide and commit to focusing on growth and taking market share instead of cutting jobs. The leadership team can then communicate that message effectively and widely to its workforce.

It can say, “This is what we are doing. These are our policies, this is our strategic plan, and this is how we will use AI going forward.”

That definitely helps create a win-win dynamic between employees and leaders.

Ross: What are some examples of companies that have done that well?

I mean companies that have adopted a growth stance and communicated it effectively.

Gleb: A great example is IKEA. IKEA has chosen to grow rather than lay off its customer-support staff.

Using AI tools, it has enabled faster and more effective customer support. As a result, it no longer needed 8,000 people in their previous customer-support roles, but it did not lay anyone off.

Instead, it provided those employees with other positions and retrained them to offer customers a higher level of service. It is focusing on growth and taking market share from less effective companies.

That is one brand-name company. I work all the time with companies that are not brand names and do exactly this sort of thing.

For example, an insurance company I worked with chose to improve its customer service and service quality instead of laying people off. Rather than cutting claims staff because of more effective AI usage, it focused on decreasing claims-processing cycle time.

That means reducing the time between receiving and processing a claim and sending a response to the insured. People receive their money faster, become more satisfied, and spread the message by word of mouth.

As a result, the company is getting more insurance business because it is improving its reputation. I cannot name the client, of course.

But this is an example of what can be done, and companies are doing it. The broader Stanford study shows that these two companies are not the only examples.

Ross: Let us talk about what happens on the ground inside an organization. We have discussed the larger frame of leadership attitudes.

We have also discussed introducing AI as something people must learn to use for themselves. What else have you seen in effective adoption programs?

Suppose an organization says, “We are transforming our organization, and we are bringing in AI.” What other factors drive value for the people working there as well as for the organization itself?

Gleb: We have discussed fear of job loss, which is very important. There are two more sources of resistance, so let us discuss the next one.

This source of resistance comes from a threat to professional identity. Let me clarify what I mean.

AI is unlike a CRM. People do not feel that their professional identity is threatened by entering and retrieving information from a CRM instead of using spreadsheets.

AI tools, however, can write great sales outreach and marketing emails. They can produce excellent marketing briefs, images, wireframes, PowerPoint presentations, and so on.

These are skills that many people have spent years developing. AI can now perform those tasks as quickly as they can, or much more quickly.

Sometimes it can also produce the same quality, especially when prompted well. People consequently feel that their professional identity and expertise are threatened.

That expertise is often a source of meaning and purpose. I will return to the insurance company I worked with.

Before I started working with it, the company bought an AI solution from a vendor that performed claims interpretation. When the company received claims from insured customers, the solution interpreted them using the underlying policy documents.

There were approximately 80 policy types. It interpreted each claim and recommended to the claims adjuster what was and was not covered.

That is really at the core of claims adjusters’ professional expertise. People felt threatened by it.

They expressed this threat by saying they did not trust the tool. Even after it was rolled out and they were effectively required to use it, they still returned to the underlying documents.

They said, “I cannot simply trust it. I have to look at the underlying documents and read them.”

They consequently still spent approximately 95% of the time they had previously spent on the task. Given the investment and the tool’s additional cost, it did not deliver a worthwhile return.

That illustrates how people fight back when their professional expertise is threatened. This resistance is a major problem.

Ross: There is another point in the example you just gave. I think it is also about system design.

If you simply give someone an answer to approve without explaining the underlying reasoning or structure, they become a yes-or-no button-pusher. I obviously do not know the details of the software.

But if it presents the underlying logic, people do not need to return to the original documents. Instead, the system can support them in making the decision themselves.

I think that is often a problem. This returns to the redesign point.

You should create a human role that is not simply pressing a yes-or-no button as the point of accountability. Instead, present the underlying data as efficiently as possible.

The system can say, “This is my recommendation, this is why, and this is the supporting logic.” I agree with your point, but I think there is also a redesign issue.

How do you provide the most efficient interface between the AI and the human? People should feel they are applying their professional experience and judgment, but they should also genuinely be doing so.

That requires efficient access to the right information.

Gleb: The tool was not making worse decisions than the claims adjusters. When the vendor showed side-by-side comparisons, it was making perfectly reasonable decisions.

But the people simply did not trust it. In this highly regulated industry, humans also had to make the final decision under Department of Insurance guidelines.

If a human said, “I do not trust it enough to make the final decision; I need to read the underlying document,” that became the issue. I understand what you are saying about redesign.

But my preferred way of approaching this problem is to work with people and see what they want to automate. When I started working with the company, we set this tool aside and conducted a survey.

We asked, “What creates obstacles for you? What do you find annoying?”

We also asked, “What manual work do you dislike doing? What prevents you from doing the things you want to do?”

In other words, what did they want to do more of, and what did they want to do less of? That is what we focused on.

For example, we found that the claims adjusters did not like writing claims letters after interpreting claims. Those letters had to cite legal language.

They had to copy and paste text from the relevant section of the document. There were approximately 80 policies, each in a PDF of around 20 pages.

They had to find the right section, copy and paste it, and get the formatting right. Then they had to interpret it in layperson-friendly language.

That was annoying manual labor that they really disliked. So that is the part we automated.

We developed prompts and then agents that the different specialists involved in each part of the claims process could adapt to their needs. That is what we helped them automate.

We also automated tasks such as classifying information sent by the claimant and placing it in the appropriate file. These are examples of the tasks they disliked, although the claims letter particularly sticks in my mind.

They were happy to use the resulting tools. Once we rolled them out, they achieved widespread adoption.

People were glad to use them because they fit their use cases and addressed what annoyed them. They were incentivizing and motivating, and they worked with people’s psychology rather than against it.

The company might find claims interpretation most financially beneficial because it requires significant time and expertise. But that is also the task claims adjusters find most threatening, so do not start there.

Eventually, once there is widespread adoption of lower-hanging fruit and AI becomes a regular part of the enterprise, you can address more challenging tasks. You can tackle emotionally demanding work that people find threatening after they become comfortable with the tool.

But you need to make them comfortable with the tool first. That happens by addressing the tasks they find most annoying and are inherently motivated to automate.

That is how you work alongside human psychology and create a win-win situation. Just as we discussed creating a win-win situation around jobs, you also need to create one around tasks.

That is how you actually achieve adoption.

Ross: There is also a great framework from Stanford with a useful matrix.

It maps where AI can have the greatest impact against where people actually want or do not want to use AI. That is obviously a very good starting point.

You introduce it in areas where people want to eliminate work, rather than where they want to retain it. I think there is one final point of resistance you were going to address.

Gleb: That is right. The final point of resistance comes from shame and social stigma.

Many people experience shame and social stigma from colleagues when discussing their use of AI tools. That includes using AI to draft emails, create PowerPoint presentations, and perform similar tasks.

I will give you an example. The first two categories tend to be rank-and-file employees.

The AI alarmists fear for their jobs, while pragmatic resisters experience threats to their professional identity. Both are discussed in Chapter Six of my book, The Psychology of AI Adoption at Work: From Resistance to Results.

This third category, which is also discussed in that chapter, concerns shame and social stigma. I call these people reluctant adopters.

They use AI tools, but they are reluctant to talk about them. This issue reaches the top of the organization.

I gave a presentation in Pittsburgh, a city of approximately two million people in Pennsylvania. I was presenting to a group of approximately 10 executives from middle-market firms.

One CEO led a financial-management company with approximately 300 or 400 employees. He told me he knew exactly what I was talking about.

Approximately two weeks before my presentation, his vice president of sales had presented to all the company’s executives. The presentation concerned the largest client they had acquired the previous year.

The presentation went very well, and all the other executives loved it. The vice president of sales then approached the CEO privately.

He said, “You ask me to give this type of presentation only once a year because it takes me three days to assemble. This time, I gave it to my executive assistant.”

“He used an AI tool to put it together in an hour and a half. I checked it for 15 minutes, it looked great, and the presentation went very well.”

“But I do not want you to tell the other executives. They will think I am cheating and lazy.”

That attitude pervades the organizational culture from the top. How can you achieve widespread AI adoption under those conditions?

There is a reason the McKinsey study shows 80% achieving individual productivity gains but only 37% reporting any EBIT impact. Only 6% are achieving real, material EBIT impact.

People might finish their individual tasks early, but they do not tell their colleagues or supervisors. They then spend the additional time on Facebook or listening to the Humans and AI podcast.

That is the third major obstacle I encounter.

Ross: How do you address that?

How do you get people to feel proud, share, or acknowledge that they are using AI in useful ways?

Gleb: The first step is leadership modeling. I told that CEO to return to his vice president of sales and ask him to email the other executives.

He should brag that the presentation used to take three days. This time, his executive assistant used AI and completed it in an hour and a half.

The vice president spent only 15 minutes checking it before presenting it. He used the additional time to find the next major client for that year.

He should then send the same email to all his sales staff and generally model that behavior. All executives should model how they use AI.

They should tell their subordinates how they use AI to draft emails and PowerPoint presentations and examine profit-and-loss statements. They should also share how they use it to conduct performance evaluations.

Leadership modeling is one crucial tool. Another is creating a shared asynchronous channel where people can discuss their AI accomplishments.

That could be a Slack or Microsoft Teams channel or a Trello card. You will have a small group of employees who are more excited than anxious.

That is the 9% we mentioned, at least in the United States. The figures are different in other countries, as you rightly pointed out.

Some people will be excited to discuss their AI tools and will not experience shame and stigma as obstacles. Use and leverage those AI champions.

Get them to share all their interesting AI use cases. Leaders should reinforce that sharing and positively praise those people.

People should also share AI failures. You need to bring failures into the open so that everyone does not make the same mistakes.

People need to feel psychologically safe when discussing these problems as well. Leaders should positively reinforce this behavior and even provide rewards and incentives.

For example, they might provide a $25 Starbucks card once a week for the best AI use case. Depending on the enterprise, it could be a reward from a local business.

Providing rewards and incentives for people to share publicly is very important.

Ross: Fantastic. Thank you for sharing your insights and your work.

Where can people learn more about your work, Gleb?

Gleb: If people want to learn more about me personally or connect with me, I am very available on LinkedIn. Just tell me you heard me on the Humans and AI podcast.

I do not accept random LinkedIn connection requests. My book is published by Georgetown University Press, an academic press.

It is written in a Harvard Business Review style. In fact, Harvard Business Review recently published an article of mine based on the book.

You can find it by searching for my name and Harvard Business Review. The book itself is peer-reviewed.

It is available from Amazon, Barnes & Noble, and wherever you buy books. If you want to read a free sample, visit disasteravoidanceexperts.com/aibook.

Again, that is disasteravoidanceexperts.com forward slash AI book.

Ross: Great. Thanks so much, Gleb.

Gleb: Thank you for inviting me, Ross.