Skip to content
  1. Home
  2. Blog
  3. AI Agent Adoption Failure: What MIT’s 95% Figure Measures
AI

AI Agent Adoption Failure: What MIT’s 95% Figure Measures

Joris Beerda4 August 2026
Finance professional working through a printed report with a calculator while the company AI assistant panel on her screen sits unused

“Ninety-five percent of AI pilots fail” has become the standard opening line on enterprise AI. It comes from an MIT report published in July 2025, and it’s usually read as proof that AI agent adoption fails because employees won’t use the tools their company buys. An earlier version of this article read it that way too.

The report says something else. It found that employees use AI a great deal, mostly through tools they picked themselves, while the tools their company built or bought for them rarely get past the pilot. If you’re rolling out AI agents, that’s more useful than the headline. It tells you who you’re designing for: people who are already willing AI users and who have passed over your tool for their own.

At The Octalysis Group we design experiences that people take part in by choice. Below I go through what the MIT and Writer numbers do and don’t show, why the company agent loses to the personal tool, and what two of our own workforce projects suggest you can do about it.

What MIT’s 95% figure measures

The report is The GenAI Divide: State of AI in Business 2025, from MIT’s Project NANDA. Its opening claim is that “95% of organizations are getting zero return” on their generative AI investment. Fortune’s headline turned that into “95% of generative AI pilots at companies are failing”, and that’s the version that spread.

The number measures return. The researchers counted a deployment as successful when it got beyond the pilot phase with measurable KPIs, and they looked at ROI six months after the pilot. How many employees used a tool isn’t what they counted.

It also applies to one kind of tool. For custom and task-specific enterprise AI, 60% of organizations evaluated a tool, 20% reached a pilot and 5% reached production. General-purpose tools such as ChatGPT and Copilot did far better: more than 80% of organizations had explored or piloted them and nearly 40% reported deployment. The report covers generative AI as a whole, and it treats agents mostly as a possible remedy.

And it’s a rough number. The research rests on interviews at 52 organizations, a survey of 153 senior leaders and a review of more than 300 public AI initiatives. The authors describe their figures as “directionally accurate based on individual interviews rather than official company reporting”. I treat the 95% as a signal that most company-built AI isn’t paying off yet, and I wouldn’t put it in a board paper as a measured failure rate.

The report’s own explanation is about the tools: “The core barrier to scaling is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time.” That’s a product problem, and I want to be clear about it before I talk about motivation. If your agent is worse at the job than the tool your employee already uses, fix the agent first. Motivation design won’t rescue it.

Your employees already use AI

The finding I’d put in front of a steering committee is on page 8 of the same report. Only 40% of the companies said they had bought an official LLM subscription. At more than 90% of them, workers reported regularly using personal AI tools for their work. The authors call this the shadow AI economy and note that it often delivers better ROI than the formal initiatives.

Writer’s 2026 enterprise AI adoption survey points the same way for agents. Workplace Intelligence surveyed 1,200 executives and 1,200 employees for it between December 2025 and January 2026. Of the executives, 97% said their company had deployed AI agents in the past year. Of the employees, 52% said they had used one in that time, and every employee in the sample already used generative AI at work. Writer sells an enterprise agent platform, so read its numbers with that in mind.

The same survey produced a figure that gets quoted even more loosely: 29% of employees, and 44% of Gen Z, “sabotaging” their company’s AI strategy. The definition matters here. The report counts anyone who does at least one of a list of things, and the list includes using AI tools the company hasn’t approved, entering company information into a public tool and declining to use the company’s AI tools or training. It says only “a few” tamper with metrics or produce poor output on purpose. Most of what Writer calls sabotage is the behavior MIT calls shadow AI. The reason employees gave most often was fear for their job, at 30%.

Taken together, the two sources describe a workforce that wants AI and has reservations about the company’s version of it. I wrote about the same pattern in why your people ignore AI agents.

Why the company agent loses to the personal tool

MIT’s answer is blunt: “ChatGPT beats enterprise tools because it’s better, faster, and more familiar, even when both use similar models.” There’s a motivational difference as well, and that’s the part I can speak to.

Think about how someone uses a personal AI tool. They chose it. They decide what to hand it, they see the result a few seconds later and they adjust. In the Octalysis Framework this is Core Drive 3: Empowerment of Creativity & Feedback. A choice motivates when it has consequences you can see and respond to, and a personal AI tool gives you that all day long.

The company agent usually arrives with the choices already made. Procurement picked the tool, a project team picked the use cases and management decided what gets measured. What’s left for the employee is to comply.

We study games for a reason. Nobody has to play one, and the moment a game gets boring, people drop it. Employees treat AI tools the same way wherever they can, and with a free alternative one browser tab away, they can. A company agent can be logical and well integrated and still lose, because good functionality doesn’t make anyone want to use it.

Then there’s the fear. An agent that does part of your job touches Core Drive 8: Loss & Avoidance, and what you stand to lose is the thing you’re valued for. A launch email about efficiency doesn’t take that away, and it can easily read as confirmation.

Most rollout plans answer all this with change management: communication, training and an executive sponsor. You need those, and they’re a small part of the problem. They explain why the tool matters to the company. They don’t change what using it is like for the person doing the work, which is why I’ve argued before that AI adoption is a motivation design challenge.

What voluntary adoption looked like at CAIXA and Navo Orbico

We haven’t published a case study on an AI agent rollout, and I won’t present these two as one. They’re projects in which employees were asked to do work they had been skipping, and in which taking part was voluntary.

CAIXA Econômica Federal employs about 83,000 people in 4,500 branches across Brazil. The bank had been communicating its targets through KPIs, spreadsheets and directives from the top, and it had tried generic gamification platforms that fell short. About 10% of employees were participating.

The program we designed with CAIXA gave every employee individual goals tied to the actions the sales objectives depended on. Reaching a goal returned a small instant reward or a collectible piece. Pieces formed sets that unlocked larger rewards, and employees could trade pieces with colleagues to complete them. Participation wasn’t mandatory, and it reached 92%. During the program the bank reported net income of R$12.7 billion, up from R$8.6 billion.

Navo Orbico, a Procter & Gamble distributor operating in 19 countries, had a problem closer to software adoption. Its field reps did the selling. They often neglected the CRM entries and stock checks, and training completion was low. Cash bonuses were in place and no longer produced extra effort.

We decided against adding points and badges to the CRM. We built an experience over the existing CRM in which each rep captained a trading ship, upgraded it through training and chose a strategy for growth. The team shared a Port City that suffered when collective targets were missed. The rollout wasn’t mandatory and 99.5% of eligible employees joined. Revenue rose 28.6% against historical performance, and KPIs for sales input activities such as up-selling and accurate reporting improved by 60%.

In both companies the directive and the bonus were already there and had stopped working. What we changed was the experience of doing the task. Feedback came at once, choices had consequences and colleagues had a reason to help each other.

Designing AI agent adoption phase by phase

The same thinking applies to an agent, and it starts with what you count. A licence count tells you what was deployed, and logins tell you who opened the tool. Neither is the behavior you paid for. We begin every project with the business metric and the Desired Actions behind it. For an agent, that means naming the tasks you want people to hand over and what should improve when they do.

From there, motivation changes as people gain experience, so the design has to change with it. Octalysis describes this through four Experience Phases: Discovery, Onboarding, Scaffolding and the Endgame.

Onboarding: a first win on their own work

The first task someone gives the agent should be one they already do every day, like the vendor email they write each morning. They can judge the result immediately against their own version. The feeling we design for in Onboarding is: this is easy, I can do it, I came to the right place.

Early wins and visible progress work on Core Drive 2: Development & Accomplishment. That’s extrinsic motivation. It gets people started, and it doesn’t last if nothing follows it.

Scaffolding: their own way of working with it

This is where most rollouts stay flat. People don’t all work with an agent in the same way. Some like to talk to it. Some want a conversation partner to think with. Some want it to execute and nothing more. Teach people strategies that fit their style, and don’t hand everyone the same prompt guide.

I’d also avoid a design in which people pass their work to the agent and let it do everything. Let them find its faults and correct it. They should come away feeling they still matter, and that they’re better and more in charge with the agent than without it. That speaks directly to the fear, and it’s the subject of designing meaningful work when AI can do the task.

What one team works out should reach the next team as well. Shared strategies and visible examples from colleagues do more here than another training module. I looked at why leaders and individual contributors differ so much on this in The Agentic Gap.

Endgame: something for your veterans to do

If the agent looks the same on day 300 as on day one, your most experienced users have nothing to do that a new user doesn’t. Give them ways to contribute: teaching colleagues, sharing the strategies they’ve built and helping newcomers through their first weeks. That’s how veterans become ambassadors for the tool inside the company.

None of this means adding points and a leaderboard to the agent. That’s the mistake our guide to what gamification is and what it is not deals with.

The agent can’t design its own adoption

It’s tempting to assume the agent will personalize all of this by itself. AI can estimate what a typical user with a similar profile is likely to do next. It has very little information about why this particular person stopped using it, which phase they’re in or what motivates them. Someone has to design for that. I go deeper into this in how to design AI that people want to use.

Where to start

If your agent is live and usage is flat, check the tool honestly first. Is it at least as good at the task as what your people use on their own? If it is, the gap is in the experience around it, and that’s the work we do for digital transformation, employee engagement and learning and development programs.

Leaving it as it is has a cost you can already see in the MIT data. Your people keep using AI, on tools you don’t control, and the agent you paid for stays in the pilot. If that’s where you are, contact us and tell us which tasks you expected people to hand over.

Frequently Asked Questions

What does MIT’s 95% AI failure figure measure?

It measures financial return. It doesn’t measure usage. MIT Project NANDA’s 2025 report The GenAI Divide says 95% of organizations are getting zero return on generative AI, and that only 5% of custom or task-specific enterprise AI tools reach production. The research rests on interviews at 52 organizations, a survey of 153 senior leaders and a review of more than 300 public initiatives, and the authors call their figures directionally accurate.

Why do employees ignore the company’s AI agents?

Often because the tool they chose themselves is better, faster and more familiar, which is MIT’s own explanation. There’s a motivational reason too. With a personal tool the employee decides what to use it for and sees the result at once. The company agent usually arrives with the tool, the use cases and the measures already chosen by someone else. Fear of losing their job adds to the reluctance.

Is AI agent adoption a change management problem?

Only in part. Communication, training and executive sponsorship explain why the tool matters to the company. They don’t change what using the agent is like for the person doing the work. Adoption also depends on a first win on the employee’s own task, real choices in how they work with the agent, and something meaningful for experienced users to do.

What does Writer’s 29% AI sabotage figure mean?

In Writer’s 2026 survey of 1,200 executives and 1,200 employees, 29% of employees said they were working against their company’s AI strategy in at least one way. The list includes using unapproved AI tools, entering company information into public tools and declining to use the company’s AI tools. The report says only a few tamper with metrics or produce poor output on purpose. Writer sells an enterprise AI platform.

Does gamifying AI adoption mean adding points and badges to the agent?

No. At Navo Orbico, a Procter & Gamble distributor, The Octalysis Group decided against adding points and badges to the CRM and redesigned the experience of the daily sales work. The rollout wasn’t mandatory, 99.5% of eligible employees joined, and revenue rose 28.6% against historical performance. The design work is in the experience around the tool.

Already have a project in mind? Let’s work together!

Our team will listen to your engagement challenges and decide if we are the best match.

Get in touch

Case studies

Measured results from our client work

See all 14 case studies

Hospitality

Hospitality Conversion Gamification

The Octalysis Group created a new industry standard for hospitality and travel booking platforms.

33.4%Increase in gross bookings

Loyalty

Gamified Loyalty Program

Gamified loyalty experience with personalized missions and rewards, leading to record-breaking engagement and conversions.

+153%Increase in credit card acquisitions

Fast-moving Consumer Goods

Sales Force Engagement

Gamified sales platform transforming reps into virtual traders, boosting engagement and performance across 19 countries. Best Gamification Award.

+28.6%Increase in sales revenue