Practical guide · Adoption

You’ve automated the process. But if no one uses it, the output is zero.

It’s the link almost everyone forgets. You can pick the right process, standardize it perfectly, and build the best agent in the world: if the team doesn’t adopt the new way of working, the chain breaks right here. Adoption isn’t a final detail, it’s what produces the result.

Raffaele Zarrelli·Founder, Yempik·May 30, 2026·12 min read
In summary
  • Adoption is the last link: without it, a chosen, standardized, automated process produces nothing.
  • People don’t resist out of laziness: fear, trust, skill, and absent managers. It’s change management.
  • You don’t impose it, you cultivate it: champions, training in the flow, leaders who model it, a low-risk start.
  • You measure it with a spectrum, not a yes/no: what matters is how it shifts over time.
The last link

The chain breaks where you’re not looking

The path to integrating AI has four phases: you choose the process, standardize it, build the agent, and finally get it adopted. We covered the first three in the Standard-First method. The fourth is the one almost no one talks about, and it’s also the one where projects die.

The reason is simple: a system, however well built, produces value only when people actually use it, every day, instead of the old way of working. It’s the same logic as the weak link: you can have four perfect steps, but if the last one doesn’t hold, the chain’s output is zero.

The numbers confirm it. Only about a quarter of AI initiatives reach the expected return, and many licenses that were bought simply go unused.[4] Not because the technology doesn’t work, but because no one worked on adoption. About 80% of the challenge is people, process, and culture; technology is only 20%.[1]

An agent no one uses isn’t a half-finished project: it’s a full cost at zero return.

The diagnosis

Why people resist (it’s not laziness)

The first mistake is reading resistance as reluctance or a closed mind toward what’s new. Almost always it’s a rational reaction to one of four fears, and each one has a precise move to counter it. 75% of employees don’t feel confident using AI, and some even go as far as obstructing its rollout.[2] Ignoring these fears doesn’t make them disappear, it makes them silent.

Fear for their job

“They’re replacing me with a machine.”

The move

Explain what changes in the role, not just in the tool. The agent removes the repetitive work, not the person.

Lack of trust

“I don’t trust what it produces.”

The move

Start with low-risk tasks where a mistake is obvious right away and easy to fix. Trust is built on small wins.

Perceived skill

“I don’t know how to use it, it’s not my field.”

The move

Training in the flow of work, not one-day courses: micro-guides at the exact moment they’re needed.

Absent manager

“My boss doesn’t use it and never mentions it.”

The move

The leader has to be the first visible user. If whoever leads doesn’t adopt, no one adopts.

The principle

Adoption isn’t imposed, it’s cultivated

The instinctive reaction of whoever leads is to impose it: “starting today we use this, period.” That’s exactly the move that fails. Forcing usage without addressing the fears produces surface-level adoption, people open the tool when you’re watching and go back to the old method the moment they can, or work around it entirely.

It’s the same mechanism as shadow AI, but in reverse. Just as banning AI pushes it into the shadows, forcing it makes people reject it. There’s only one path that works: make the new way of working so useful, accessible, and supported that people choose it. Cultivate, don’t compel.

You can force someone to open a tool. You can’t force them to find it useful. Real adoption only comes from the second one.

The method

The four levers of adoption

Cultivating doesn’t mean hoping. There are four concrete levers, used together, that actually move people. Together, according to the data, they can raise adoption by up to 40%.[3]

Internal champions

The people who already tinker with the tools, earlier and better than everyone else.

How to do it

Pick one or two per team, give them 30-60 days and recognition. They spread usage peer to peer, more than any course.

Training in the flow

Learning while you work, not in a classroom you forget by Monday.

How to do it

Micro-guides and prompts at the moment of need, inside the tool. Skill is built through practice, not through slides.

The leader who models it

Whoever leads uses the tool first, visibly.

How to do it

Only 35% of employees have a manager who sets the example on AI. If the boss doesn’t use it, the team reads it as not really mattering.

Start with low risk

First uses on tasks where a mistake is harmless and fixed right away.

How to do it

Trust is built on small wins. Only afterward do you move to critical processes, once the habit has already set in.

The numbers

How do you measure AI adoption in your company?

“Are they using it?” is the wrong question, because it has only two answers and neither is useful. Adoption isn’t a switch, it’s a spectrum: people spread across five levels, and your job is to shift the distribution to the right over time.

01~35%

Non-users

They don’t use it, by choice or out of fear.

02~25%

Explorers

They try it now and then, with no method.

03~22%

Regular users

They use it for the tasks it’s meant for.

04~13%

Power users

They integrate it and find new uses.

05~5%

AI-native

They reach for it by default, every day.

Resistancethe goal is to move people to the right →Value
The percentages are a typical starting example. Adoption isn’t measured with a yes/no: you measure it by watching how this distribution shifts over time. If after three months the “non-users” shrink and the “regulars” grow, you’re winning.

Measure the signs of habit, not the licenses you bought: how many people actually use the system every day, on how many tasks, how often. A license that’s paid for and never opened counts as zero. The sign that you’re winning isn’t an initial spike, it’s a distribution that, month after month, slides toward the power users.

In practice: three kinds of metric

“Shifting the distribution” is the principle. In practice, you measure it by combining three kinds of data: what people do (usage), how they feel (perception), and whether the work improves (impact).[5] On their own, each one lies; together, they tell the truth. Token consumption tells you they use the tool, but not whether it’s any good; a survey says they’re happy, but not whether they produce more. You need all three.

Usage metrics

Usage

Are they using it, and how often?

  • Weekly and monthly active users (WAU / MAU)
  • Daily / monthly ratio: tells you whether usage is a habit
  • Paid licenses versus licenses actually used
  • Token consumption or API calls per person or team

How to collect it
From the tool’s analytics panels and the consumption logs. It’s the objective data, already there: you don’t have to ask anyone.

Perception metrics

Perception

How do people feel?

  • Short pulse surveys: 2-3 questions, every 2-4 weeks
  • Perceived trust in the tool (1-5 scale)
  • Time saved as reported by the people using it
  • Short interviews with non-users: what’s blocking them

How to collect it
With recurring micro-surveys and a few conversations. It explains the “why” behind the usage numbers, and surfaces the hidden resistance.

Impact metrics

Impact

Is the work getting better?

  • Output quality: how often it has to be corrected or redone
  • Process cycle time, against the baseline
  • Errors and rework, before and after
  • The process’s business KPI (e.g. response time)

How to collect it
By comparing the process numbers with the baseline measured before you started. This is where adoption ties back to the economic return.

A practical rule so you don’t drown in data: start from the baseline (the same one you set when you standardized the process), then measure on a fixed cadence, every month or every quarter. A few numbers, always the same ones, watched over time. A dashboard with one metric per kind, usage, perception, and impact, is worth more than twenty indicators no one reads.

Token consumption tells you they open it. The survey tells you they’re happy. Only output quality tells you whether it’s really working.

And once adoption is stable, the last question remains, the one the CFO asks: does all this translate into dollars? It’s a step of its own, with its own traps, and we cover it in how to measure AI ROI.

The link

Adoption and governance go together

There’s a point where adoption and governance meet. Giving people official, approved, supported tools, with a clear channel where they can ask “can I use this?”, is at once a lever for adoption and a defense against shadow AI. People stop improvising with personal accounts because they finally have a better, legitimate alternative.

That’s why adoption shouldn’t be thought of on its own: it rests on the rules and the inventory we described in AI governance for SMEs. Governance and adoption are two sides of the same goal: getting AI used well, by the right people, on the right data.

Have you automated something the team doesn’t use?

On a call we build the adoption plan: champions, training in the flow, and the metrics to see whether it’s really working.

Book a call
FAQs

The questions we get asked most

Why don’t employees use the AI tools the company introduces?

Rarely out of laziness. The real causes are fear for their job, a lack of trust in what the AI produces, low perceived skill, and the absence of a manager who sets the example. It’s a change-management problem: about 80% of the adoption challenge is people, process, and culture, not technology.

How do you actually get the team to adopt a new tool?

With four concrete levers: internal champions who spread usage peer to peer, training in the flow of work instead of courses people forget, leaders who use the tool first and visibly, and first uses on low-risk tasks to build trust. Banning or imposing it doesn’t work: it only pushes usage into the shadows.

How do you measure AI adoption, in practice?

By combining three kinds of data. Usage: weekly and monthly active users, frequency, token consumption, licenses used versus licenses paid for, pulled from the tool’s analytics panels. Perception: short, recurring pulse surveys, perceived trust, self-reported time saved, interviews with non-users. Impact: output quality (how often it has to be corrected), cycle time, and errors against the baseline. On their own they mislead; together they tell the truth. You start from the baseline and measure on a fixed cadence, monthly or quarterly.

Is adoption a separate phase or does it need to be planned from the start?

It has to be planned from the start, but it plays out at the end. It’s the last of the four phases in the method (Discovery, Standard, Agent, Adoption) and the only one that turns a working system into real value. Without adoption, all the upstream work stays potential: only about a quarter of AI initiatives reach the expected ROI, often precisely because the tools go unused.

Transparency note

I wrote this article myself. The method and the opinions come from my work and from Yempik’s real projects. For the writing I had Claude Opus 4.8 help me with editing, clarity, and layout. The substance is mine; the tool is disclosed.

Transparency

Sources

  1. [1]WRITER: in enterprise AI adoption, 80% of the challenge is people, process, and culture, not technology. writer.com
  2. [2]People Managing People: 75% of employees don’t feel confident using AI; 29% admit to obstructing the strategy. peoplemanagingpeople.com
  3. [3]Superhuman / Prosci: champions, training in the flow, and leaders who model the behavior raise adoption by up to 40%. blog.superhuman.com
  4. [4]WorkOS: only about 25% of AI initiatives reach the expected ROI; many licenses go unused. workos.com
  5. [5]Atlassian: measuring AI adoption by combining usage, perception, and impact metrics. www.atlassian.com