Skip to content
Back to the blog
  • Inteligência Artificial
  • Transformação Digital
  • Liderança Técnica
  • Produtividade
  • Futuro do Trabalho
  • Governança de IA

Why AI Projects Fail

Selling Productivity, Not Replacement

The number shows up in every talk about enterprise AI: 95% of projects fail. On the first day of AI Brasil Experience 2026, on October 28, I heard a reading of that number that fits what I live as head of AI Enablement at InBot. The cause is rarely the model. It is what the company promises people when it shows up with AI. Promise cuts and you get resistance. Promise time back and you get allies. In this post I share what I heard on stage and in a conversation afterwards, explain what the 95% actually measures, and show why adopting AI in-house is harder for a company that existed before it, even when that company sells AI. At the end, I leave five practices for your team.

What I Heard at AI Brasil: Replacement Sold Better Than Reality

The talk was called "Quando a máquina trabalha pra gente?" ("When does the machine work for us?"), given by Ricardo Heidorn, CEO and co-founder of Leads Per Hour, a company that sells AI agents for sales, and CRO at Lincros. Keep that detail in mind: this is someone who makes money from AI saying that the most common way of selling AI is wrong.

He opened with a provocation: the promise that AI would replace humans was more successful than the results it delivered. Every launch brought a new scare, and the scare reached the headlines long before it reached the balance sheets. Later came the line that gives this post its title: selling productivity works; selling replacement does not.

The retreats of those who bet on cutting people are already documented, and I listed some of them in AI Governance in Practice. A new data point comes from the very report that made the 95% famous: among the myths it debunks is the idea that AI will take most jobs within a few years. The researchers found few layoffs caused by generative AI.

If the cuts did not come, why do so many AI projects stall? Before answering, it is worth understanding what that 95% counts.

Do 95% of AI Projects Fail? What the Number Measures

The statistic comes from The GenAI Divide: State of AI in Business 2025, a preliminary report by MIT NANDA published in July 2025. The version that circulates is "95% of AI projects fail". The report's version is narrower: among custom-built enterprise AI tools, only 5% reached production with a clear and sustained impact on results, assessed about six months after the pilot. The funnel narrows fast: 60% of companies evaluated a tool of this kind, 20% ran a pilot and 5% put it into production.

The study has real limits. It is based on 52 interviews and 153 surveys, with self-reported answers and a demanding, loosely defined bar for success. Rob Wiblin, from 80,000 Hours, ran the math from the other side: among the companies that reached a pilot, roughly one in four succeeded. That is why I treat the 95% as a sense of direction, not as a scoreboard.

The problem is not the model, it is the approach

The part worth reading is the authors' diagnosis:

"This divide does not seem to be driven by model quality or regulation, but seems to be determined by approach."

What separates those who get results from those who do not is neither model quality nor regulation, but approach. The report's list of barriers includes tools that do not fit the workflow, people unwilling to adopt new tools, poor change management and a lack of executive sponsorship. As Fortune put it, the technology was not the problem; the way companies were using it was.

Heidorn called this culture, and that is the word I would use too. Microsoft and LinkedIn's 2026 Work Trend Index surveyed 20,000 AI users in 10 countries, Brazil among them, and found a telling ratio: in the report's statistical model, organizational factors such as culture and manager support account for 67% of the impact people report from AI, and individual factors for 32%. It is an association, not a cause, as the report itself warns. Only 26% see their leadership aligned on the subject. Since Microsoft has a direct interest in adoption, I read the number with that caveat, but it points to the same place as MIT.

Tools became a commodity, culture did not

Another remark from the talk: whatever is good ends up becoming a commodity, and AI will be no different. Prices already show it. Running a GPT-3.5-level model has become hundreds of times cheaper since 2022, according to Stanford's AI Index. Any competitor of yours can subscribe to the same model tomorrow. What they cannot subscribe to is a team that knows how to get value out of it.

Heidorn also pointed out something that often goes unnoticed: most people do not know what AI is, even inside tech companies. Without that basic explanation, the empty space gets filled by whatever is in the headlines. And what is in the headlines, almost always, is fear.

Fear Breeds Sabotage: Resistance to AI Inside the Team

Heidorn's line was short: some people sabotage AI because they are afraid of losing their jobs. It is a predictable reaction: if the tool arrived announced as your replacement, helping it succeed feels like digging your own grave.

Perception data confirms how big that fear is. Pew Research Center found that 52% of U.S. workers are worried about AI's impact on work. In BCG's AI at Work 2025, 41% think their own job may disappear within ten years. And sabotage is not a figure of speech: in a Writer survey, 41% of millennial and Gen Z employees admitted to sabotaging their company's AI strategy, for example by refusing to use the tools. Writer sells AI and the sample covers only two generations, but the behavior is the one Heidorn described.

Gallup adds an important nuance: only 15% of workers think it is likely they will lose their job to AI within the next five years. General concern is much higher than the real expectation of being laid off. In my reading, this means that what fuels resistance is the story being told about AI, more than the risk each person sees for themselves.

Duolingo learned this in public. In 2025, a memo from the CEO declared the company "AI-first" and said contractors would gradually be replaced by AI wherever it could do the work. The backlash was so bad that the CEO himself softened the tone and admitted: "I did not give enough context". No full-time employee was laid off, but the message that stuck was the memo's.

Companies born with AI vs. companies that already existed

After the talk, I went to speak with Heidorn in person. I wanted to know what he did with AI internally, and the answer came in two parts, one for each company.

At Leads Per Hour, internal adoption was natural. The company was born AI-first, so nobody had to change the way they worked: that already was the way they worked. At Lincros, it was much harder. It is a company that existed before generative AI, with a team that had its own processes and more resistance to change. There, he said, adoption took a lot of top-down pressure.

That conversation helped me see where InBot stands. At first glance, we look like the first case: InBot is an AI company. It has built conversational AI since long before the term became fashionable. Robô Ed, created by InBot and launched in 2004 for a Petrobras campaign about energy consumption, talked in natural language with children and teenagers and stayed online until 2016. But that is the product InBot delivers to its clients. Using AI in our own work is something else, and on that front InBot looks more like Lincros: a team with years of established processes that already worked well the old way.

Here, the beginning came from the bottom, with developers adopting AI before there was a dedicated area for it (I told that part in the governance post). The challenge now is different: how to take that culture to the whole company without the top-down pressure that Lincros needed. In other words, moving forward in the least aggressive way possible.

Culture starts at the base, not at the top

In the same conversation, Heidorn gave a hint for that path that was not on his slides: find people at the base of each team. Those who work in operations know the bottlenecks much better than leadership does. I asked whether he meant AI champions, people on each team who drive AI use among their peers, and he confirmed that was exactly it.

A champion solves two problems at once. They know where the team's week gets stuck, so they point AI at the right place. And they talk to colleagues as a colleague, which takes away the top-down, order-from-above tone of the change. It is the opposite of top-down pressure, and it is what makes it possible to move forward without running anyone over.

Having AI champions at InBot is an idea I already had, and it was put on hold at the time. I left that conversation wanting to put it back on the table.

Resistance is not only fear of losing your job

There is a kind of resistance that the fear explanation does not cover, and I live with it. To this day, some people on the development team avoid using the corporate Skills Hub because it downloads things onto their computers.

I understand the distrust, and the data backs it up. In Stack Overflow's 2025 survey, 81% of developers cite security and privacy as a concern when using AI agents. In August 2025, the attack known as s1ngularity showed why: tampered versions of a popular npm package triggered the AI command-line tools installed on victims' machines to hunt for credentials, and more than 2,000 GitHub accounts were exposed. Someone who asks what is being installed on their own computer is doing exactly what a good developer should do.

This kind of resistance is not solved by an order, and it should not be. The path I see is to open the box: show what the Hub installs, where it comes from, who signed the release and who reviewed it before publishing. And accept that convincing people takes time. Culture is not a stage you complete. It is an ongoing conversation.

Fear and distrust explain what gets stuck. What gets things moving, from what I have seen so far, is changing the promise.

Selling Productivity Means Giving Time Back

On stage, Heidorn presented Leads Per Hour's mission like this: give back to humans what belongs to humans. For a person, the scarcest resource is time, and an AI agent is only worth it if it gives some of that time back.

This is where the two narratives part ways. The replacement narrative speaks to finance and scares everyone else: "we will need fewer people" reaches each person's ears as "maybe they won't need you". The productivity narrative speaks to the people who do the work: "that tedious part will take minutes, and you decide what to do with the rest of the afternoon". Only the second gives people a reason to root for AI.

The MIT report has a finding that reinforces this. Most of the users interviewed welcomed automation, especially for manual and repetitive tasks, on two conditions: secure data and measurable results. Those are the same two things the developers who distrust the Hub are asking for.

The promise needs proof

Saying "nobody will be replaced" is easy. The hard part is for that sentence to still hold six months later. What builds credibility is proof: at InBot, automations have already turned hours-long tasks into minutes, and nobody left because of them (the numbers are in the governance post). When the colleague next to you gains a free afternoon and stays on the team, the message stops being just talk. To see how this plays out beyond software development, I wrote about AI skills by area.

The other half of the argument is still missing: what happens to results when the human stays in the process.

Human + AI Delivers More Than AI Alone

During the Q&A, Fernando Sapata, who had spoken earlier on the same stage, asked Heidorn what lead generation will look like going forward. The answer came with numbers from Leads Per Hour, which I wrote down on the spot and share as an account of the event, not as research:

  • in lead capture, the conversion rate was 1.6% with AI working alone and 3% with a human alongside it;
  • in conversations with customers, AI alone succeeded in 1 out of every 100 attempts; with a person in the middle of the process, receiving leads that were already warm, it was 20 out of every 100.

In his view, the market still wastes hours on forms, and that is the part AI will change. The machine handles the repetitive triage and hands the person the conversation that is worth having.

Public data points the same way. Fang and co-authors tested a pre-sales assistant on a large online retail platform: compared with no assistance, the AI-only version increased sales by 16.3%; when the AI could hand the conversation over to a human, the increase was 25%. It is a preprint, not yet peer-reviewed. The NBER study on customer support, which I have cited on this blog before, measured AI working alongside human agents, and the gains were largest among the least experienced.

The point is not to put a human at the end of the line to check the machine's work. It is to design the process with the person at the center: AI takes the volume and the repetition, and the human keeps the judgment, the relationship and the decision.

How to Change the AI Conversation in Your Team

None of this requires a director title. It requires choosing carefully what you say and what you show. I left the event with five practices in mind:

  • Look for champions at the base. Almost every team already has someone who uses AI on their own and knows where operations get stuck. Give that person room and recognition to bring their colleagues along.
  • Talk about time, not cuts. Replace "what AI will replace" with "how much time it gives back, and for what".
  • Measure together with the person who does the task. Time the task with the person who does it, before AI and after it. That way the gain belongs to them, not to a management spreadsheet.
  • Open the box for those who distrust it. Show what the tool accesses, what it installs and who reviewed it. A security question deserves an answer, not pressure.
  • Show the promise being kept. When someone gains hours in their week, tell the team, and also tell them that this person is still there.

Communicating well is still rare, which makes it an advantage. According to Gallup, only 22% of workers say their company has shared a clear AI plan, and those who received one are 3 times more likely to feel very prepared to use the technology.

Conclusion

The 95% is often used as proof that enterprise AI does not work. Read carefully, the report says something else: the technology is available to everyone, and the difference lies in how each company presents it and fits it into people's work.

In a company born with AI, that culture comes for free. In a company that already existed, even one that has been selling AI for more than twenty years, like InBot, culture has to be built, and you can build it with top-down pressure or with people at the base bringing their colleagues along. I am betting on the second, with transparency for those who distrust it and with time given back as the promise.

If you want to start this week, pick a repetitive task, measure it with the person who does it and show the team how much time came back to that person. And notice who on your team is already doing this without anyone asking. That is probably where your first champion is.

References