---
title: 88% of companies adopted AI. Only 6% got results
description: Why AI projects fail to move the profit needle - and what to do about it. No technical jargon.
date: 2026-08-09
tags: [agentic-ai, ai-adoption, enterprise]
cover: cover.webp
coverAlt: Abstract composition of glowing nodes on a dark background
draft: false
sources:
  - https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
  - https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
---

A McKinsey survey from November 2025 found that 88% of companies already use AI in at least one process. Yet only 39% of those companies see any impact on profit at all. And only about 6% have reached what McKinsey calls "high impact" - meaning AI contributes more than 5% to the bottom line.

The other 94% bought licenses, ran training sessions, filed "successful deployment" reports - and nothing changed.

This is not a technology problem. It is a problem with how companies think about adoption.

![Dashboard showing a large "88%" and a small "6%" side by side - the gap between companies that deployed AI and those that got results](fig-dashboard.webp)

## Why the reports look great but the work speed stays the same

When a company rolls out an AI assistant - say, a corporate chatbot connected to a large language model - leadership sees encouraging numbers on the dashboard. A dashboard is a screen that displays usage metrics, showing how many people are "using" the system.

The problem is in how "user" gets counted. Anyone who logs in even once during a month enters the statistics. Open, close, forget - already an "active user."

So the reports look optimistic. The reality is different.

The typical breakdown after an AI assistant rollout, at a company of any size, looks like this:

- **5-10% of employees** - power users. They work with AI every day, tune it to their specific tasks, and see a real speed increase.
- **Around 20%** - occasional users. A couple of requests a day: rewrite an email, draft a reply to a client. They have not quite figured it out, but they are not causing problems either.
- **Roughly 70%** - do not use it at all.

![Three employee groups: a small circle for "5-10%", a medium circle for "20%", and a large circle for "70%" - showing who actually works with AI](fig-usage-split.webp)

Seventy percent. That is not laziness or resistance to change. It is rational behavior from people nobody explained this to in the context of their specific job.

## Who pays and who works: the gap between licenses and real use

In large deployments - where a company buys thousands of licenses and pays tens of millions of dollars a year - the picture sharpens further.

Roughly 10% of employees consume roughly 90% of tokens. Tokens are the units you pay for AI work: the more text the AI processes, the more tokens get spent and the higher the bill.

The company pays for thousands of seats. A small group uses them at full capacity.

If every employee worked with AI as intensively as that top 10%, the bill would increase tenfold. That is not happening - and it will not happen on its own.

![Chart: 10% of employees account for 90% of tokens spent - nearly the entire budget flows to a small group](fig-seats-tokens.webp)

The 2025 MIT NANDA study "GenAI Divide" confirms the same pattern at the level of entire projects: about 5% of corporate AI pilots produced a fast revenue increase. The remaining 95% showed no measurable effect.

Five percent.

## A tale of two engineers: why the same tool produces opposite results

Two engineers receive the same task. Both use AI. The outcomes are opposite.

The first pastes the task into AI, gets edits across six files at once, glances at the green test results - everything passes - and accepts all the changes. Three weeks later the server breaks. The cause: AI silently changed one configuration setting that was never supposed to be touched. The first engineer missed it because he did not read every change.

The second engineer works differently. She tells the AI where the code lives and what must not be touched. She maintains a skill file - a plain text file the AI reads before every new task, like a standing instruction sheet. She reads every proposed change herself, spots the unnecessary edit, and rejects it. She accepts half as many changes as the first engineer. Each time AI does something wrong, she adds a new rule to the file - and week by week the AI performs more accurately for her.

The first engineer created a failure that someone else spent hours fixing. The second saved time and created no problems.

Same tool. Different skill.

![Two paths with AI: left side - accepted fast, then dealt with the fallout; right side - reviewed, configured, got a reliable result](fig-skill-gap.webp)

## Why working well with AI is a separate skill you cannot buy on a course

Working well with AI means doing several things at once. Keeping context clean - meaning not feeding the AI so much irrelevant information that it gets confused. Maintaining skill files. Understanding that AI makes mistakes and checking the output. Knowing which tasks AI helps with and which ones only add work.

Half the employees at any company will never do this. Not because they are poor workers, but because it is not their interest or their responsibility.

Consultants sell prompting courses - training on how to write better requests to AI. That is useful, but it covers a small part of the problem. The real question is different: which processes in this company are worth fully automating, and which ones should not be touched. Every company has its own answer.

Meanwhile, AI vendors keep releasing more powerful tools. The skill ceiling rises. Power users benefit from their colleagues falling behind - their value comes from that gap. They will not share knowledge voluntarily.

![The widening gap between employees who can work with AI and those who cannot - and how that gap grows over time](fig-chasm.webp)

This is the real "GenAI Divide" the MIT study describes. Not the divide between companies that have AI and those that do not. The divide inside every company - between those who can use it and those who cannot.

## What to do: two paths and three concrete steps

There are two ways to respond to this gap. The first is to train everyone and hope the majority catches up. The second is to accept that most people will not catch up and build a system around that fact.

A combination of both works.

![Two paths: broad training in hopes of results, and embedded agents that do the work instead of people - and how the two complement each other](fig-two-paths.webp)

**Step one: find the capable people - they hide in unexpected places.**

Running training is worthwhile not to turn everyone into a power user. It is worthwhile to find the people who will become power users. Some of them sit in accounting, in logistics, in the call center - places nobody thinks to look. They pick it up fast, start applying it, and within a month are doing in an hour what used to take a day.

**Step two: give power users status, rather than asking them to share for free.**

An internal library where advanced employees post their work - skill files, ready-made prompts, workflow templates - with visible ratings and author names attached. In exchange for losing their edge, they gain recognition. That works better than a corporate memo saying "please share knowledge."

**Step three: for the other 70%, build agents - not training.**

An AI agent is a program that completes a task on its own, without a person involved at every step. It plugs into tools people already use every day: a CRM (customer relationship management system, which tracks client interactions), accounting platforms, or a ticket queue (a list of incoming support requests). The employee does not learn to work with AI - they simply approve, reject, or edit a finished result.

A concrete example: a department that processes incoming invoices. Agents work through the invoices overnight - matching them against contracts, checking amounts, preparing a decision. In the morning, people open the queue and review: "approve," "reject," "clarify." Work that used to take half a day now takes an hour.

## How to report to the board so the numbers actually mean something

The standard metric - "adoption rate" or "number of active users" - says nothing about real impact.

A useful report looks different. Take any process and track what share of the work gets done three ways: manually, with AI assistance, and fully automatically. Then watch how that share shifts from quarter to quarter.

If the manual share is falling, the deployment is working. If it stays flat, it is not.

That is the one number that honestly answers the board's question: "Are we getting a return on AI or not?"

## Q&A

### We already bought licenses and ran training. What do we do next?

Look at real usage, not login statistics. Find the 5-10% of employees who work with AI every day - and start there. Ask them to describe exactly what they do. That gives you a map of the processes in your company that genuinely respond to automation.

### Does every employee need a prompting course?

Prompting courses are useful for finding capable people, but do not expect them to change the behavior of the majority. For the 70% who do not use AI, embedding agents into the tools they already use every day is more effective than any training program.

### How do we decide which processes to automate and which to leave alone?

There is no universal answer - every company has its own. A reliable guide: processes with repetitive, predictable steps and a clear output automate well. Processes that depend on judgment, relationships, or unusual situations do not. Start with the first kind.

### How do we convince the board that the AI project is working?

Do not show "adoption rate." Show what share of a specific process is done manually, with AI assistance, and fully automatically - and how that share changes each quarter. That is the only number that answers the question about return on investment.

*Majento builds AI agents and deploys them inside client companies. Majento engineers embed in your processes, find the bottleneck, build the system, and stay until results are real - they do not leave after the presentation. If you want to find out where AI can actually help in your company, reach out: [t.me/shimaoz](https://t.me/shimaoz) or [hello@majento.ai](mailto:hello@majento.ai).*