AI adoption is not a business outcome

Giving a team access to AI is not the same as improving the business. Usage metrics can show whether a tool is being tried. The stronger question is whether a defined piece of work became faster, better, safer or less expensive.

The short answer

AI adoption is an implementation signal, not a business result.

A company can buy licences, train its team and see thousands of prompts without improving one customer outcome or removing one operational bottleneck. It can also create value through a quiet system that few employees ever open because the AI works inside an existing process.

This does not make adoption irrelevant. People need access, confidence and appropriate skills before many forms of change are possible. But access and activity sit at the beginning of the chain:

1. Access: people can use the system.
2. Capable use: they can apply it appropriately to real work.
3. Operational result: the work measurably improves without creating unacceptable cost or risk.

Only the third layer establishes business value. The practical goal is therefore not “get everyone using AI”. It is “improve this workflow, for these people, against these measures”. For a growing business, that often means evaluating AI automation inside one defined process rather than launching a company-wide adoption campaign.

↔ Scroll sideways inside the diagram on a narrow screen

Why adoption figures can mislead

AI use is already widespread. In McKinsey’s 2025 global survey, 88% of respondents said their organisations regularly used AI in at least one business function. Yet only 39% attributed any enterprise-level EBIT impact to AI, and most of that group estimated an impact below 5%.

Those figures do not prove that most AI programmes fail. They are self-reported survey results, and profit is not the only legitimate outcome. AI may improve customer satisfaction, employee experience, quality or innovation before its financial effect can be isolated.

They do show why “we use AI” is an incomplete management statement.

The same survey defined a small group of AI high performers—about 6% of respondents—using both reported value and an estimated EBIT contribution of at least 5%. These organisations were more likely to redesign workflows, track solution-level performance and define when human validation was required. The relationships are correlations, not a recipe. Still, they point toward a more useful unit of analysis: not the tool in isolation, but the way work has been reorganised around it.

Using AI and improving work are different questions

A login can tell you that someone opened a product. A prompt count can tell you that an interaction occurred. Neither reveals whether the person chose an appropriate task, supplied reliable context, checked the output or saved any time after correction.

The distinction matters because AI is unusually sensitive to context.

In a study of 5,179 customer-support agents, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that access to an AI assistant increased issues resolved per hour by 14% on average. The gains were much larger for novice and lower-skilled workers and minimal for the most experienced group. The system was not merely a blank chat box: it provided suggestions inside a specific support environment and had learned patterns from prior conversations.

A separate experiment with 758 Boston Consulting Group consultants found a different boundary. For tasks judged to be within the model’s capabilities, people using AI worked faster and received higher quality ratings. On a task outside that capability frontier, however, AI users were 19 percentage points less likely to reach the correct answer.

These studies examined different people, systems and tasks. They should not be combined into a universal productivity estimate. Their shared lesson is more modest and more useful: the effect of AI depends on the work, the worker, the design of the assistance and the way quality is checked.

That is why an adoption percentage alone cannot guide an investment decision.

Measure AI automation at workflow level

Start with one recurring business workflow: qualifying an enquiry, resolving a support request, preparing a weekly report, reviewing a document or moving information between systems.

Establish how that workflow performs before adding AI. Then compare the same measures after the change. M4TIC recommends a compact scorecard:

Not every workflow needs all six measures. Choose a small set before the pilot starts. If success is defined after the results arrive, almost any activity can be presented as progress.

Usage data still has a place. It can reveal whether access, relevance or training is blocking the experiment. Treat it as diagnostic evidence—not as the final score.

Two different jobs require two different approaches

Many organisations force every AI initiative into the same model: buy a general-purpose assistant, distribute accounts and ask people to find uses for it. That can be valuable, but it serves only one kind of need.

Support capable individual use

General-purpose tools are useful when employees regularly face open-ended research, drafting, analysis or problem-solving tasks and can judge the result. Some people will find high-value uses quickly; others will need examples, practice and clearer boundaries.

Training should therefore produce observable work, not attendance certificates. Ask participants to bring a recurring task, complete it with and without assistance, document the method and compare the result. Reusable patterns can then be shared with colleagues who do similar work.

The objective is not to turn every employee into a prompt specialist. It is to develop the judgement required to decide when AI helps, when it does not and how its output should be checked.

Embed AI into repeatable workflows

When the same operational problem affects many people, the better answer may be workflow automation or a designed AI system rather than a blank interface.

An embedded system can classify an incoming request, extract approved fields, retrieve controlled information, prepare a draft or route an exception inside tools the team already uses. The employee interacts with the result and the decision point, not necessarily with the model itself.

The architecture should match the variation in the work. Workflow automation can control a predictable sequence of triggers, rules and approvals. AI automation can add interpretation where the process contains variable language or documents. An AI agent may be relevant when the system must choose among approved actions as conditions change. Greater flexibility is useful only when the workflow genuinely requires it.

This approach does not remove the need for adoption. It changes what must be adopted: a revised process, new responsibilities and a clear review step. Its advantage is consistency. Context, permissions and evaluation can be designed once instead of reconstructed in every prompt.

The two approaches can coexist. Individual tools help people explore and handle variable work. Embedded systems make repeated workflows more dependable. The choice should follow the work, not a company-wide slogan.

A practical first experiment

Choose one workflow with enough volume to measure and consequences low enough to manage safely. Avoid starting with the most impressive demonstration or the most consequential decision.

Document seven things:

1. the event that starts the work;
2. the usable result it should produce;
3. the current median time and common failure points;
4. the information and systems involved;
5. the step AI will assist or perform;
6. the person who reviews exceptions and owns the outcome;
7. the measures and threshold that would justify continuing.

Run the old and new methods on comparable cases. Include review time, corrections, integration effort and operating cost. A draft produced in two minutes is not a time saving if it requires twenty minutes of repair.

↔ Scroll sideways inside the diagram on a narrow screen

At the end, make one of four decisions:

• Stop: the result is worse, riskier or not worth the total cost.
• Refine: the workflow is promising, but context, controls or handoffs need work.
• Scale: the improvement persists across enough cases and the exceptions are understood.
• Keep as individual assistance: the value depends on human judgement and does not justify a shared system.

This is a stronger basis for investment than a target number of active users.

What management should ask instead

Replace “What is our AI adoption rate?” with a short sequence of operational questions:

• Which workflows are we trying to improve?
• What baseline do we have for time, quality, rework and cost?
• Where does AI assist a person, and where does it act inside the process?
• Which outputs require human validation?
• What happens when information is missing or the system is uncertain?
• Who owns performance after the pilot ends?
• What result would make us stop, refine or scale?

An organisation may still track licences, active users and training completion. It simply should not confuse those numbers with value.

AI creates business value when a real piece of work improves and the improvement survives contact with cost, exceptions and accountability. Adoption may help that happen. It is not the destination.

If the work is predictable and follows stable rules, explore workflow automation. If selected steps require interpretation, see AI automation. When a system needs to choose among approved actions as conditions change, custom AI agents may be relevant.

Sources and further reading

Primary guidance behind this article.

1. AI Adoption is a MythVas Moza / Varick Agents

   The source article that prompted this M4TIC adaptation, particularly its distinction between tool rollout and operational value.

2. The State of AI: Global Survey 2025McKinsey & Company

   Survey evidence on reported AI use, scaling, EBIT impact, workflow redesign, human validation and KPI tracking.

3. Generative AI at WorkErik Brynjolfsson, Danielle Li and Lindsey R. Raymond

   Field evidence from 5,179 customer-support agents; NBER Working Paper 31161, later published in The Quarterly Journal of Economics.

4. Navigating the Jagged Technological FrontierFabrizio Dell’Acqua et al.

   Field experiment with 758 consultants on the task-dependent effects of generative AI.

5. The Adoption of Artificial Intelligence in FirmsOECD, BCG and INSEAD

   Wider context on organisational capabilities, skills and barriers affecting AI diffusion in firms.