The most expensive way for enterprises to import AI is to buy it and no one will use it: Moderna reversed the purchasing order

This article originally appeared in Issue 130 of Vista Newsletter (opens in a new tab).

The most common way for enterprises to introduce AI is not to choose the wrong expensive tools, but to buy them and no one uses them, and they end up shelving them.

The authorization was purchased, the account was opened, and the education and training was completed. Three months later, when I opened the backend, I found that less than 10% of the users were active weekly, and they were concentrated on those people who would have been able to explore on their own.

To be honest, this is not an isolated case. Market observation points out that about 88% of AI test projects never enter the formal production environment, and only about a quarter of them actually achieve the expected return on investment. In other words, those AI import success stories you’ve heard are statistically in the minority.

In this e-newsletter, I want to talk to you about a company that passed this hurdle and the sequence in which it got it right.

Everyone should be familiar with Moderna, the biotech company that makes mRNA vaccines. In early 2023, it used OpenAI’s API to build an internal chatbot called mChat.

The key to this is not technology, but timing: At that time, they had not bought ChatGPT Enterprise.

The adoption rate of mChat is over 80%.

The interesting thing about this case is that before they spend a lot of money to purchase something, 80% of the employees in this company have developed the habit of asking that thing first if they have questions. When the official enterprise version is introduced later, what they have to deal with is not how to get everyone to start using it, but how to get people who are already using it to switch to a better interface.

The difficulty of these two questions is exactly one order of magnitude different.

When most companies introduce AI, the order is exactly the opposite: purchase first, then announce, then conduct training, and then hope that the habit will grow on its own.

But the cruelest thing is yet to come, because habits will not grow by themselves. Because in the eyes of users, it is something ordered from top to bottom and has no clear relationship with their current work. People usually don’t change the way they work to comply with policies, they just change the way they work to get off work earlier today.

Moderna is equivalent to using a simple version first to develop needs and habits, and then letting formal tools take over an already hot market.

They didn’t rely on a briefing to decide whether to continue developing mChat or switch to ChatGPT Enterprise.

Brice Challamel, who is responsible for AI products and platforms, put it very bluntly: As a science-based company, they have to study everything. So the team conducted a large-scale user test to compare mChat, Microsoft’s Copilot and ChatGPT Enterprise.

The criterion they use is Net Promoter Score, not a functional comparison table.

This is worth stopping to think about.

The decision of most enterprises depends on the comparison list made by the purchasing department: are certain specific functions available, how much is the price, and is the security of the product acceptable? These are all questions, of course, but they don’t answer the only question that really determines success or failure: whether this group of people want to use it or not.

If a colleague gives up on a powerful tool after using it three times, its actual function is zero.

Moderna concluded that ChatGPT Enterprise’s Net Promoter Score was ridiculously high, and they went all-in on this one.

Challamel explained the entire logic in one sentence: they were never here to fill a bucket, but to light a fire.

Filling water means handing out the tools, completing the class, and handing in the usage report; lighting means asking someone to make something, and then the colleague next door sees it and says, “Let me use this.”

The former is a special case and will be closed. The latter is diffusion and will grow on its own.

This is what I find most interesting about the whole case.

The goal Moderna has set for itself is to achieve 100% adoption and proficiency among all employees with access to digital tools within six months.

Please pay attention to the two words at the end, proficiency. It’s not about opening an account or taking classes, it’s about being proficient.

The difference is very practical:

When the KPI is adoption, all your resources go to those who haven’t started using it yet.

When the KPI is imported, your resources will flow to the system itself.

If the acceptance criterion is that the system goes online, the project will end on the day it goes online. If the acceptance criterion is adoption, the project doesn’t start until the day it goes live.

To that end, their approach is three-tiered:

Personal level, in-depth research and listening plans, physical and online training, and exclusive AI learning partners. Their argument is that using AI to teach AI is the key to success.

Collective layer, this layer is the most interesting. They held a prompt word competition to identify the top 100 heavy users in the company, and formed this group of people into an internal generative AI champion team. Then set up office hours in every business line and region so anyone can come in with questions. There is also an internal AI forum with two thousand active participants every week.

Structural layer, the CEO and the operating committee are personally involved, establishing culture through executive meetings and all-hands meetings, and matching the reward system.

I want to point out one thing in particular: the one hundred champions were selected from the competition, not the designated ones.

This is a big difference. The designated seed person is usually the person sent by the heads of departments to hold meetings; but the winner selected through the competition is the person who is already using it crazily and cannot be stopped.

The former needs to be pushed, the latter is the push itself.

The numbers two months after importing ChatGPT Enterprise are as follows:

750 custom GPTs have been created across the company, with an average of 120 conversations per user per week, and a 100% adoption rate by the legal department. Later, this number continued to grow, with more than 3,000 custom GPTs accumulated internally.

But the most important thing in the whole set of numbers is actually another one:

40% of weekly active users have built GPT themselves.

Generally, companies measure adoption rate by looking at how many people have used it? To be honest, that number is easy to fake, just log in once. But what Moderna looks at is how many people have built something?

The transition from user to builder is a threshold that is difficult to fake. Because it requires you to have an idea about your work first, and know which section is repeated, which section is wasted, and which section can be handed over.

If an employee can create a customized GPT, it means that he has thought through his workflow. This thing is worth more than the tool itself.

What really sets this case apart from other companies is not the three thousand GPTs, but the changes to the organizational chart.

Moderna merged its human resources and information technology departments into one and created a new position called Chief People and Digital Technology Officer.

Moreover, the person who took over this position was the original head of human resources, not the technical director.

Some commentators pointed out that this choice itself is a message: technology serves people, not the other way around.

The logic of this step actually conveys a certain meaning very directly: when AI becomes part of the execution work, not just a support tool, then whether to find a person to do it or whether to use an agent becomes the same decision. In traditional organizations, these two issues are managed by two departments that have no contact with each other, separated by budget subjects, annual plans, and each other’s turf.

After the merger, Moderna also changed the way it thought about the organization: it no longer started from functions, but looked at the company as a series of tasks that needed to be arranged, and then asked one by one whether this should be done by humans, machines, or both together in some way.

This article is not a promotional advertisement for the manufacturer, so there are a few things that must be made clear to everyone.

First, some people are still being laid off. This organizational adjustment laid off about 10% of the technical manpower, about 50 people. It is worth noting that it is the technical department that has been laid off, not other departments. Some commentators believe that this shows one thing: if the main purpose of introducing AI is to reduce manpower, this group will not be laid off.

Second, the learning curve for entry-level positions is a real issue. GPT agent took over some of the work originally handled by junior staff, especially on the human resources side. The time saved is real, but there is no answer yet as to where young employees will learn from those trivial things.

Third, Moderna’s conditions are difficult to replicate. It has technology genes, the transformation muscle accumulated during the epidemic, a culture of experimentation, and a CEO who will take matters into his own hands. One commentator put it well: You can copy everything except the context that makes it possible.

I include these points because case studies that only talk about pretty numbers are useless. Everyone needs to know what can be copied and what cannot be copied, so that you know where to start?

This article is the third article in the series of corporate dismantlings I wrote at inno.tw (opens in a new tab). When viewed together, the three articles are exactly three points on a line.

Shopify and Duolingo (opens in a new tab) asks: Is there anything behind your policy? In April 2025, the two companies issued AI policy memos with almost identical contents three weeks apart. One was used as a model by other companies, and the other was so popular that social media accounts had to be cleared. The difference is not what was written in the policy, but the four years before the letter was issued.

Klarna’s article (opens in a new tab) asked: Are your indicators set correctly. The most widely circulated version is that the company laid off 700 people and replaced it with AI. After the quality collapsed, they had to bring people back. Upon verification, almost every paragraph in this sentence is inaccurate, and the true version is far more useful to businesses that are evaluating the introduction of AI.

Moderna’s article (opens in a new tab) asks: What does it look like after the indicators are set correctly?

These three things actually echo the same sentence: first there is implementation, then there is policy; first there are correct indicators, then there is correct design; first there are people who want to use it, then there is the question of whether the tool is good or not.

If the order is reversed, any step will become an empty performance.

The above is the skeleton I grabbed from the original article, but there are many details that cannot be put here: the actual practice of three-tier change management, the actual appearance of Dose ID and legal contract GPT, what to do with the prompt word competition, and a more complete comparison between the three articles. The full version is here, you are welcome to look over (opens in a new tab).

If your company happens to have an AI tool that no one uses, you can think about the following four questions:

First, are your KPIs about import completion, or adoption rate and proficiency? This question determines the flow of all subsequent resources.

Second, have you developed the demand before purchasing? If you haven’t bought it yet, consider making a very simple, even a little difficult to use version for everyone to play first. If you have already bought it, find a small unit to perform this sequence again.

Third, are your internal seeds assigned, or do they emerge on their own? Hold a competition, look at the background data, ask who is using it crazily, find those people who can’t be stopped, and then give them status and stage. The person assigned to the meeting couldn’t bring anything with him.

Fourth, among your success indicators, is there a number of people who have built something themselves? Used and built are two different things. The former can be pushed, the latter can only be stimulated.

In the past few years, I have used one person and a set of AI workflows to maintain and operate several product lines including writing, products, courses, learning and experiments. Several independent websites, list systems, cash flows, automated sequences, and dozens of self-built tools are all maintained by me.

Now, I’m bringing this practice into the enterprise (opens in a new tab).

However, let me first make it clear that there are some situations where it is not suitable: if what you want is a report that can be submitted, if you just want to listen to a motivational speech, or if you want to replace an entire batch of manpower with an AI solution, then this is not the original intention of this service. This approach requires at least one internal person who is willing to do it. Without that key man, no matter how powerful the consultant is, it may be useless.

If the conditions are right, the starting point is half-day AI innovative diagnosis. Invite key personnel to come in together, take stock of which processes can really be handed over to AI and which cannot, and produce a one-page introduction roadmap, including priorities and expected resistance. The fee for subsequent running appointments can be fully discounted for the first month.

Find out where the card is before deciding whether to spend more money? This sequence is about the same thing as the Moderna case. inno.tw is a new website I opened at the end of July. It currently has two axes:

Teardown, 1 to 2 in-depth case studies per week. In addition to the above three enterprise-oriented articles, there are also Differences between Two Innovation Firms in Japan (opens in a new tab), How Nendo uses more than 30 people to do 400 design projects at the same time (opens in a new tab), The real success point of a one-person company (opens in a new tab).

Signal is a batch of short messages every two weeks.

There are almost no hard-coded statistics on the site. The number of articles, tools, or update times are all re-crawled from the public endpoints of each site during each build. The reason is simple: anyone can write static copy, and numbers that change on their own are the real evidence. Everyone is welcome to visit: inno.tw (opens in a new tab)

See you in the next e-newsletter.

Sincerely, Vista