Companies Have AI Tools. Do They Actually Know How to Use Them?

Over the last couple of years, AI adoption inside tech companies has happened surprisingly fast. Companies bought ChatGPT Enterprise accounts, rolled out Copilot or Cursor, gave teams access to Claude, increased API budgets, wrote a few security policies, and encouraged employees to experiment.

From the outside, this looks like AI transformation. But I’m not sure it really is.

What I keep noticing is that a lot of companies seem to be following the same playbook: give people access to powerful AI tools and hope that a good way of working with them emerges naturally. Some employees become very good at it. They develop their own workflows, learn how to work with coding agents, write better instructions, automate repetitive parts of their jobs, and share the occasional trick with their team. Others use AI as a slightly better Google search. Some use it occasionally, and some barely touch it.

Eventually management wants to know whether all of this is actually working.

And that is where things get complicated.

McKinsey’s 2026 State of AI survey captures this gap pretty well. Nearly nine in ten respondents said their organizations were regularly using AI in at least one business function. Eighty percent said AI had improved their own productivity. Yet only 37% said AI had contributed positively to their company’s EBIT, and the group McKinsey classifies as “AI high performers” was still only around 6% of respondents.

So employees are getting faster, but companies are not necessarily getting better at the same rate.

I think that is the real AI transformation problem.

Giving everyone AI does not create an AI-native company

For the first stage of adoption, the bottom-up approach actually makes a lot of sense. Nobody knew exactly what these tools would be good at, so companies let employees experiment.

This works especially well in engineering teams. Developers tend to discover tools on their own. One engineer starts using Claude Code heavily. Another builds a few Cursor rules. Someone creates an MCP integration. Another person figures out a surprisingly effective workflow for debugging production issues.

Slowly, some kind of AI culture begins to form.

The problem is that this creates huge differences between employees.

Imagine two engineers at the same company with access to exactly the same models. One has spent a year learning how to structure tasks, manage context, ask an agent to investigate before coding, separate planning from implementation, run multiple agents, review generated code and build reusable instructions around the repository.

The other mostly uses autocomplete.

Technically, both employees have “AI access.” In reality, they are using completely different tools.

This becomes an organizational knowledge problem. The first engineer may have learned dozens of useful things, but most of that knowledge probably lives in their head, their local setup, a CLAUDE.md file, or a few forgotten Slack messages. The company has paid for the learning, but it does not necessarily own the learning.

The obvious question is: how does something one employee discovered become a repeatable way of working for the rest of the company?

That part is much less developed than giving everyone an AI account.

Individual productivity is not the same thing as company productivity

There is another mistake that is easy to make: assuming that making individuals faster automatically makes the organization faster.

If a developer used to finish a task in six hours and can now finish it in three, that is obviously useful. But a company is a system, and systems have bottlenecks.

If developers suddenly produce twice as many changes, somebody still has to review those changes. More code can mean more pull requests, more QA, more coordination and potentially more bugs. Once implementation gets faster, code review might become the bottleneck. If review gets faster, product decisions might become the bottleneck. If everything gets faster, deployment or customer support might become the bottleneck.

Google’s 2025 DORA research found something very close to this. Higher AI adoption was associated with improved software delivery throughput and product performance, but it still had a negative relationship with software delivery stability. Their explanation is important: AI accelerates development, but that acceleration can expose weaknesses elsewhere in the system. Teams with strong automated testing, mature version control and fast feedback loops benefit more; teams with weak downstream processes can simply create instability faster.

In other words, AI does not automatically remove bottlenecks. Sometimes it just moves them.

This also makes AI ROI surprisingly difficult to measure. Tokens consumed, weekly active users, prompts sent and autocomplete acceptance rates are easy numbers to put on a dashboard, but they are not business outcomes. A company can have fantastic adoption metrics while creating more review work or shipping more low-quality output.

The more useful questions are harder: did cycle time improve? Did quality change? Did customer outcomes improve? Are employees spending more time on important work? Did the cost of producing an outcome go down?

Those are not AI metrics. They are company metrics.

Managers may have the hardest job in this transition

A lot of the public discussion around AI focuses on employees and executives. Employees want to know how AI affects their job. Executives want to know how much money AI can save or create.

Managers are stuck between the two.

A CEO can say that the company should become “AI-first,” but somebody has to translate that into actual day-to-day work. Should every employee be expected to use AI? Which tasks should they use it for? How much human review should remain? What does good AI-assisted work look like? If someone becomes dramatically faster with AI, should they simply receive more work? If someone uses almost no AI but continues to perform extremely well, is that a problem?

Gallup’s research suggests that companies themselves know they have a management problem. In an August 2026 survey of 102 CHROs, 99% said AI was important to their organization’s strategy, but half said they were not confident in their managers’ ability to guide employees on using AI at work. Fifty-seven percent were already providing AI training specifically for managers, and 62% were creating internal AI champions or centers of excellence.

The employee data makes the role of the manager even clearer. Gallup found that, among organizations implementing AI, employees who strongly agreed that their manager actively supported AI use were 8.7 times more likely to say AI had transformed how work gets done. They were also much more likely to use AI frequently.

That makes sense. Employees do not only need permission to use AI. They need context around where it is useful, where it is risky, and what the company actually expects from them.

Without that guidance, every employee is effectively running their own little AI transformation program.

Most companies are still putting AI inside old workflows

I think this is one of the most important distinctions.

Suppose the existing software development process looks like this: a product manager creates a ticket, an engineer investigates the problem, writes the code, adds tests, opens a pull request, and another engineer reviews it.

Now give the engineer Claude Code.

The engineer may investigate faster, write the implementation faster and generate the tests faster. That is valuable, but the company has not really redesigned anything. It has put a powerful new tool inside the same old process.

A more fundamental transformation starts when the workflow itself changes. Maybe an agent investigates the ticket and proposes an implementation plan. A human checks the plan. Another agent implements it and runs tests. Automated review catches simple problems, while the human reviewer spends more time on architecture, product behavior and unusual edge cases.

Whether that particular workflow is a good idea depends on the team. The important point is that the company is now asking a different question. Instead of “How can this employee use AI?”, it asks “How should this work be done now that AI exists?”

McKinsey’s 2026 data shows a striking difference here. Nearly three-quarters of the organizations it classified as AI high performers said they were fundamentally redesigning workflows because of AI. Among everyone else, only about one-quarter said the same. High performers were also more likely to have senior leadership commitment and defined processes for measuring the impact of AI initiatives.

That seems much closer to real transformation than simply buying more licenses.

Employees are excited about AI and worried about it at the same time

From an employee’s perspective, AI creates a strange situation because the benefits and the anxiety can come from exactly the same experience.

Imagine doing something in twenty minutes that used to take three hours. The first reaction is obvious: this is great.

The second reaction is a little less comfortable: if this can do in twenty minutes what took me three hours, what happens in another two years?

Gallup published an interesting result in September 2026. Employees who used AI frequently — daily or several times a week — were more than twice as likely to fear that their jobs could disappear within five years compared with people who only used AI a few times a month or year. Simply using AI more did not make people more comfortable with it. Good management practices, however, significantly reduced that anxiety.

This does not necessarily mean frequent AI use causes job anxiety. The people who use these systems every day may simply have a better understanding of what they can already do.

There is also a quieter concern around skills. EY’s Work Reimagined 2025 research, based on 15,000 employees and 1,500 employers across 29 countries, found that 37% of employees were worried that becoming too dependent on AI could weaken their own expertise. Sixty-four percent said their workload had increased because of growing performance pressure. At the same time, only 12% said they had received enough training to fully benefit from AI’s productivity potential.

That combination is worth paying attention to. AI can make people more productive while also raising expectations around how much they should produce.

Employees may save time and still end up feeling busier.

Training exists, but most of it is still catching up

There is no shortage of AI education now. OpenAI has expanded Academy into role-specific learning and adoption programs. Microsoft launched an AI Transformation Leader certification aimed at business leaders who need to identify AI opportunities, plan adoption and align investments with business goals. MIT Sloan offers a six-week AI Adoption program specifically focused on moving from AI awareness to organization-wide transformation.

The direction is clearly changing from “learn how to prompt ChatGPT” toward “learn how to redesign work.”

That is probably necessary because generic AI training has a ceiling. Engineers, salespeople, designers, finance teams and managers do not need the same AI education. An engineer may need to learn task decomposition, context management, agentic coding and how to review generated code. A salesperson may need workflows around account research, call preparation and CRM updates. A manager needs to learn how AI affects delegation, accountability, performance expectations and team design.

BCG’s 2025 global AI at Work survey found that regular AI usage was significantly higher among employees who had received at least five hours of training, especially when training included coaching or in-person support. Yet only about a third of employees felt adequately trained. The same research found that when employees did not have access to the AI tools they believed they needed, more than half said they would find alternatives and use them anyway.

That last part creates another problem: shadow AI.

If the company provides a bad or overly restricted experience, employees can simply use personal accounts and unapproved tools. EY found that, depending on the industry, between 23% and 58% of workers were using AI applications not officially approved by their employer.

So companies cannot solve AI governance by only writing rules. The approved way of working has to be useful enough that people actually want to follow it.

The missing piece might be organizational learning

The more I read about this, the more I think AI transformation is not mainly a tooling problem.

It is an organizational learning problem.

Companies already have access to very capable models. Most tech companies can buy roughly the same AI products their competitors can. What they do not automatically get is the knowledge of how to redesign their specific work around those tools.

OpenAI has started explicitly promoting the idea of internal “AI Champions”: people who help move an organization from individual experimentation toward repeatable ways of working. The interesting part is not the title. It is the idea behind it — that successful AI adoption requires turning isolated discoveries into shared practices.

I think there is a bigger opportunity hidden inside that idea.

Imagine if a company could actually collect the AI experiments happening across the organization. An employee finds a repetitive task, tries a new AI-assisted workflow and measures whether it helps. If it works, that workflow becomes reusable. Another team discovers it, adapts it and improves it. Over time, instead of having scattered prompts and tricks living in people’s heads, the company develops an internal library of proven ways of working.

The important word there is proven. A company does not need a giant database of prompts. It needs to know which workflows saved time, which improved quality, which created problems, where human review was still necessary, and under what conditions the workflow worked.

That is very different from an internal “50 ChatGPT prompts for productivity” document.

It is closer to building an operating system for how the company learns to work with AI.

Buying the tools was the easy part

I do not think companies are necessarily doing this badly. The technology is simply moving much faster than organizational practices can move.

A few years ago, AI at work mostly meant asking a chatbot questions. Then copilots became common. Now engineers can give agents tasks that involve reading large codebases, making implementation plans, editing multiple files, running tests and reviewing their own changes. Similar changes are happening in research, support, sales, finance and operations.

It would be strange if companies already knew the perfect way to organize themselves around technology that changes every few months.

What seems clear, though, is that access alone is not enough. The research keeps pointing in roughly the same direction: useful training matters, managers matter, workflow design matters, measurement matters, and companies that redesign how work happens are seeing different outcomes from companies that simply add AI to existing processes.

The next stage of AI adoption probably will not be about whether employees have access to ChatGPT, Claude or whatever model comes next. Those tools are becoming normal infrastructure.

The harder question is whether a company can take hundreds of small experiments happening across its workforce, figure out which ones actually work, and turn them into better ways of operating.

That is the difference between a company where people use AI and a company that has actually learned how to work with it.

References

1. McKinsey & Company — The State of AI in 2026: On the Road to ROIAugust 25, 2026Read the report
2. Gallup — AI’s Effect on Workplace CultureAugust 16, 2026Read the research
3. Gallup — Using AI More Does Not Reassure Workers, Managers DoSeptember 9, 2026Read the research
4. Gallup — Employee Engagement Remains Flat as AI Adoption AcceleratesJuly 21, 2026Read the research
5. Google Cloud / DORA — Announcing the 2025 DORA ReportRead the report overview
6. Boston Consulting Group — AI at Work: Momentum Builds, but Gaps RemainJune 26, 2025Read the report
7. EY — Work Reimagined 2025Read the research
8. Stack Overflow — 2025 Developer Survey: AIExplore the survey results
9. OpenAI Academy — The AI Champion RoleRead the guide
10. Microsoft Learn — Microsoft Certified: AI Transformation LeaderView the certification
11. MIT Sloan Executive Education — AI Adoption: Driving Business Value and ImpactView the program