Generative AI business applications are moving from experiments to everyday work, but widespread use does not mean that every use is producing business value.

The research presents a mixed picture. McKinsey’s 2026 The State of AI survey found that nearly nine in ten respondents reported regular AI use in at least one business function. At the same time, only a minority reported a measurable effect on earnings before interest and taxes, or EBIT, and a much smaller group met McKinsey’s definition of a high performer.

These findings do not mean generative AI has failed. They show why business owners should look beyond demonstrations and headlines and ask a more practical question: where does this work earn its place in an ordinary business, and where does it quietly waste time?

The practical goal is not to introduce AI everywhere. It is to identify specific work where AI can assist, put a person in charge of checking the result, and keep the work that proves useful.

1. The Paradox: Widespread AI Use, Thin Business Impact

Generative AI has become familiar in workplaces, but its adoption and its business results tell different stories.

McKinsey’s The State of AI in 2026, published on August 25, 2026, reported that nearly nine in ten respondents said their organizations regularly used AI in at least one business function. The survey was based on 1,719 participants across 97 nations, with fieldwork from May 4 to June 8, 2026.

The same survey reported three findings that deserve attention:

These figures measure different outcomes. The productivity finding reflects respondents’ assessments of their own work. The EBIT finding asks whether AI contributed to the organization’s earnings. The high-performer figure applies a stricter test, combining a meaningful EBIT contribution with broader enterprise impact.

None of these figures means that a business should expect a particular financial result from adopting AI. They are survey findings, reported by McKinsey, and they describe what respondents reported in that survey. A small business should treat them as evidence about the gap between use and business impact, not as a promise.

The distinction matters to a small-business owner. An employee may use AI to draft proposals, summarize information, or prepare notes more quickly. That can feel useful. It does not automatically mean the business is earning more, serving customers better, or completing work with fewer errors. Individual usefulness and business impact are related, but they are not the same thing.

Adoption figures also need to be read in context. McKinsey’s 2025 survey reported 88% regular AI use in at least one business function. That 88% figure belongs to the 2025 survey. For 2026, McKinsey used the wording “nearly nine in ten.”

Generative AI has its own adoption history. In its March 2025 analysis of survey results, McKinsey reported regular generative-AI use in at least one business function at 71% in July 2024, up from 65% in early 2024.

The distinction between AI and generative AI is important because the terms do not describe precisely the same thing. AI is the broader field. Generative AI refers to systems that can produce new content, such as text, summaries, images, or code, based on the information they are given. A business can use AI without using generative AI, and it can use generative AI in one task while leaving other work unchanged.

What the MIT research adds

A July 2025 preliminary report from MIT Project NANDA, The GenAI Divide: State of AI in Business, highlighted a different part of the picture. Its executive summary said that 95% of organizations were getting zero return, while only 5% of integrated pilots were extracting millions in value.

This is a significant warning about the difficulty of turning AI experiments into measurable business results. The report was based on research from January to June 2025, covering more than 300 publicly disclosed AI initiatives, interviews with representatives from 52 organizations, and a survey of 153 senior leaders at four major industry conferences. The authors described the findings as preliminary and warned that the figures were directionally accurate based on individual interviews.

That context matters. The MIT finding should not be read as a universal failure rate for every AI project, tool, company, or use case. It is a warning about how often pilots fail to reach production and produce measurable profit-and-loss impact, particularly for enterprise-grade, task-specific or embedded AI systems.

The central lesson is straightforward: using AI is not the same as putting it to productive use.

For a small business, the right question is not whether competitors have adopted generative AI. It is whether the business has a specific task, a clear standard, and a person responsible for checking the result.

2. Why AI Pilots Die Before They Become Useful

An AI pilot often begins with an appealing demonstration. A tool produces a convincing marketing draft, a summary, or a set of ideas in seconds. The result looks ready to use, and the team is enthusiastic.

The difficulty comes when the experiment meets ordinary business work.

A demonstration shows what a tool can produce under particular conditions. A working business process needs the same result repeatedly, with the right information, at the right time, and with someone accountable for checking it.

Several practical problems can interrupt that transition.

Nobody owns the outcome

A manager may approve an AI experiment without assigning responsibility for its results. Employees try the tool when they remember it, but no one is responsible for deciding whether the output is good enough, whether the process is being followed, or whether the experiment should continue.

When something goes wrong, the responsibility becomes unclear. When the experiment stops being used, no one notices for weeks. The pilot does not fail loudly. It fades away.

A workable pilot needs one named owner. That person should understand the task, review the results, and decide whether the process is working. The owner should also be able to say when the tool is not helping and the experiment should stop.

The owner does not need to be a technical specialist. A person who understands the business task may be better placed to judge the output than someone who only knows how to operate the tool.

The tool sits outside the real workflow

An AI tool may work well when someone opens it separately and experiments with a prompt. It may be less useful when the work requires information from a document, a system, or a colleague, and the result must be delivered in a particular format.

The more disconnected the experiment is from the actual task, the easier it is for employees to abandon it when they are busy. A separate step that adds work is unlikely to survive, even if the output is good.

Before choosing a tool, map the work as it currently happens. Identify what information enters the process, who does the work, what the output should look like, and who checks it.

Then decide exactly where AI should assist. Sometimes it can prepare a first draft. In other cases, it can summarize information, organize notes, or prepare questions for a meeting. The point is to define the step, not to add a new activity around the work.

The goal is to fit AI into a defined step rather than create a separate activity that employees must remember to do.

Nobody checks the output properly

Generative AI can produce fluent, confident text that contains mistakes, missing information, or unsupported claims. A draft may look complete while leaving out an important condition, using the wrong detail or date, or describing a service the business does not offer.

If employees are expected to trust the output without reviewing it, the business has replaced one task with a different risk. The effort saved by generating a draft can be lost by correcting errors, or worse, by sending an error to a customer.

The review process should be appropriate to the work. A brainstorming draft may need a light editorial check. A customer-facing document, a proposal, or a summary used for a decision needs a careful review against approved information.

The person reviewing the output should know what to verify, where the supporting information comes from, and what to do when the draft is wrong.

The experiment quietly stops being used

A tool may receive enthusiastic feedback during its first few days and then gradually disappear from daily work.

That can happen because the output needs too much editing, employees do not trust it, the process takes longer than the old method, or no one is responsible for keeping it going.

A pilot should therefore examine actual use, not just initial enthusiasm.

Is the team still using the process after the novelty has passed? Are employees following the review step? Are they using the approved information, or are they entering whatever seems convenient?

If the answers are disappointing, the right response may be to change the workflow, narrow the use case, improve the instructions, or stop the pilot. A pilot that ends with a clear decision is not a failure. It is evidence.

An AI pilot is useful even when the answer is no. It becomes a problem when the business continues spending time on a tool that no one uses, or when it expands a process that has not been tested.

3. The Denominator: Why AI Results Can Look So Different

Research about AI performance can appear contradictory. One study may report widespread use and productivity gains. Another may report that most pilots produce no measurable return. Both can be describing real evidence, because they may be measuring different things.

The apparent contradiction becomes easier to understand when we examine what each study measures.

Surveys of organizations already using AI

A survey of organizations that have adopted AI can provide useful information about how people use it, which functions they use it in, and whether they feel more productive.

But the organizations and employees responding to such a survey may not represent every unsuccessful pilot, every abandoned tool, or every business that tried AI and stopped. A survey of users can tell us about the experience of users. It does not tell us about every experiment that never became part of everyday work.

McKinsey’s 2026 survey, for example, asked participants about AI adoption, individual productivity, EBIT impact, and high performance. Those are valuable measures, but they do not establish that every AI pilot succeeded or that every organization using AI is gaining business value.

Research examining pilots and implementation difficulties

MIT Project NANDA approached the subject from a different direction. Its The GenAI Divide report focused on the gap between pilots and production, and on whether AI experiments become embedded in everyday operations and produce measurable profit-and-loss impact.

Its headline finding about zero return emphasizes the difficulty of turning pilots into productive, integrated systems. That is a different lens from asking respondents how often they use AI or whether they feel more productive.

Why the distinction matters

Imagine two hypothetical studies.

The first asks employees at organizations using AI whether the tools help them complete tasks. Many may say yes, because the tools help with drafting, summarizing, or organizing information.

The second examines whether AI experiments become embedded in everyday operations and produce measurable business results. Many of those experiments may not reach that stage, even if individual employees find the tools useful.

Both findings could be true because they measure different things.

This is the production gap: a tool can work in a demonstration, help someone with an individual task, and still fail to become part of a reliable business process.

The distinction also explains why business owners should be cautious about universal claims regarding AI results. A figure that describes individual productivity is not the same as a figure that describes EBIT impact. A figure that describes organizations already using AI is not the same as a figure that describes every pilot that was started.

There is no single survey figure that settles the question of AI’s business value. Adoption statistics, productivity reports, and pilot-to-production research each answer a different question.

For a small business, its own evidence should guide the final decision. The business needs to establish what good output looks like, who checks it, and whether the process works in its own workflow.

4. Where Generative AI Business Use Actually Earns Its Place

The most useful starting point is to examine the types of work where organizations already report using generative AI.

McKinsey’s published chart, Gen AI’s broad reach, released on April 17, 2025, draws on its July 2024 survey. It identified marketing and sales as the most commonly reported function for generative-AI use, at 42% of respondents reporting use in that function.

The survey also identified product and service development, service operations, software engineering and IT as important areas of use.

These findings describe reported use in a survey. They do not prove that every application in these functions is successful. Nevertheless, they offer a practical way to think about possible applications.

Marketing and sales: preparing useful first drafts

Marketing and sales work often involves creating, adapting, and organizing information for different audiences.

Generative AI can assist with drafting campaign ideas, preparing initial versions of website copy, or summarizing research for a sales team.

For example, a small business could provide a tool with verified information about a service and ask it to prepare a draft for a specific audience. The draft can then be reviewed, corrected, and adapted.

The tool should not decide what the business promises customers. It should work from information the business has approved.

The owner or responsible employee still needs to check factual claims, product or service descriptions, and any promises made to customers.

A useful result is not simply a document that looks professional. It is a draft that meets the brief, uses accurate information, and can be sent without substantial correction.

Product and service development: organizing ideas and feedback

Generative AI can help teams organize information during the development of a new service, feature, or offer.

Possible tasks include grouping feedback into themes, summarizing research notes, preparing questions for customer conversations, or organizing ideas from a meeting.

These activities can help people work through information in a more structured way. They do not make decisions for the business.

The business still needs to determine whether the information is representative, whether important views are missing, and whether an idea is worth pursuing.

AI-generated suggestions should be treated as options to evaluate, not evidence that customers want a particular feature.

Service operations: supporting routine knowledge work

McKinsey’s July 2024 survey identified service operations as an important area of generative-AI use.

For a small business, relevant tasks may include preparing response drafts, summarizing an issue for a colleague, or helping staff find information in approved documents.

The value depends on whether the tool can work with reliable information and produce output that fits the business’s process.

A draft response based on an outdated policy may be worse than no draft at all. A summary that omits an important detail may mislead the person who reads it.

The business should therefore define which information the tool can use, what the output must include, and who is responsible for checking it before it is sent.

Software engineering and IT: assisting technical work

McKinsey’s survey also identified software engineering and IT as important areas of use. In its industry-specific reporting, software engineering was a leading area in some industries, and IT was identified as an important function in later McKinsey reporting.

For businesses working with software, AI may assist with drafting code, explaining unfamiliar technical material, or preparing documentation.

These applications still require competent review. Generated code can contain errors, explanations can be incomplete, and documentation can describe the wrong behaviour.

A business should ensure that someone with the appropriate skills checks the output, tests changes, and confirms that the result works as intended.

The broader lesson is that AI can assist with technical work without becoming the final authority on whether that work is correct.

Knowledge management: making existing information easier to use

Knowledge management is another important area in McKinsey’s reporting on AI use. In professional services, McKinsey reported knowledge management use at 34% in its July 2024 survey chart.

Businesses accumulate information in policies, procedures, proposals, training materials, research documents, and past work. Employees often spend time searching for the right document or the current version of a procedure.

Generative AI can help summarize approved material, create structured outlines, or assist employees in finding relevant information.

The business must still control the information being used and verify that the response reflects the approved source. An AI-generated summary should not silently replace the original policy. Employees need a way to check the source when the answer matters.

HR: assisting with routine documentation

McKinsey’s function-level chart identified human resources as an area of reported generative-AI use, with a 22% industry-specific figure in media and telecommunications.

Appropriate assistance may include drafting general training materials, organizing non-sensitive process notes, or preparing first drafts of routine internal documents.

People decisions require particular care. AI output should not become an unchecked basis for hiring, performance, pay, or disciplinary decisions.

The business remains responsible for the fairness, accuracy, confidentiality, and suitability of its people processes.

Risk, legal, and compliance work: organizing information, not replacing judgment

McKinsey’s survey also reported use in risk, legal, and compliance functions, including a 21% industry-specific figure in financial services.

Generative AI may help organize documents, summarize source material, or prepare a first draft of questions for a professional adviser.

However, an answer that sounds legally or technically convincing may still be incomplete or wrong. The appropriate level of review depends on the consequences of an error. Important conclusions need professional judgment, not an AI-generated summary.

Strategy, supply chain, and manufacturing: supporting analysis

McKinsey’s chart also documented generative-AI use in strategy and corporate finance, supply chain and inventory management, and manufacturing. The reported industry-specific figures were 21% for strategy and corporate finance in advanced industries, 15% for supply chain and inventory management in advanced industries, and 13% for manufacturing in advanced industries.

These are industry-specific adoption figures, not evidence that AI automatically improves results in these areas.

Possible applications include summarizing research, organizing operational notes, preparing scenario descriptions, or structuring information for discussion.

Across all these functions, the same principle applies: AI is most useful when it assists with a defined task, uses approved information, and is checked by a person who understands the work.

5. A Small-Business Playbook: Test One Task Before Expanding

A small business does not need a company-wide AI program to learn whether a tool is useful. It needs a small, controlled test with a clear standard and a named reviewer.

A short, controlled experiment makes it easier to identify what works, what needs correction, and what should not be expanded.

Step 1: Choose one repeated piece of work

Start with a task that happens regularly and has a clear output.

Examples include preparing first drafts of standard marketing content, summarizing meeting notes, or organizing customer feedback into themes.

Avoid starting with a task that involves many unclear decisions, sensitive information, or consequences that are difficult to reverse.

The first objective is to understand whether AI can assist with a manageable part of the work.

Step 2: Define what good output looks like

Before testing the tool, write down the requirements for an acceptable result.

For a marketing draft, those requirements might include accurate service information, a defined audience, a clear call to action, and a tone that matches the business.

For a document summary, the requirements might include coverage of the main points, preservation of important conditions, and no added claims.

A vague instruction such as “make this better” makes it difficult to judge the result. Specific requirements make the review process fair and repeatable.

Step 3: Assign a person to check the work

Choose someone who understands the task and can recognize an incorrect or incomplete result.

That person should review the output against the agreed requirements, identify recurring problems, and decide whether the process needs to change.

Do not assume that an employee who knows how to prompt a tool is automatically qualified to approve its output. The ability to operate a tool is not the same as the ability to judge the work.

Responsibility for the final work should remain clear.

Step 4: Run the pilot for a fortnight

Use the same defined task over a two-week trial period. Keep the scope narrow enough that the team can follow the process consistently.

Record whether the tool produces acceptable output, what needs correction, whether employees follow the review step, and whether the process fits into the normal working day.

The point is to gather practical evidence rather than rely on first impressions.

A fortnight is a suggested testing period, not a research-established rule. Some tasks may need a longer trial, particularly if they happen infrequently.

Step 5: Decide based on evidence

At the end of the trial, compare the results with the requirements established at the start.

Ask:

The answer does not have to be yes. A tool that produces unreliable results or does not fit the task should be narrowed, changed, or stopped.

Step 6: Expand only after the first task works

If the pilot meets its requirements, document the process and identify the next task worth examining.

Keep the approved instructions, review criteria, and responsibility for checking the output clear. The next task should be tested in the same way, with the same discipline.

This approach follows the broader lesson from the McKinsey and MIT research: moving from experimentation to productive use requires a defined workflow, a responsible reviewer, and evidence from actual use.

6. Build vs. Buy in Plain Terms

A ready-made AI tool is often the simplest place to begin when the task is common and the available tools already match the work. It can be tried quickly, and if it does not fit, it can be dropped without much disruption. The trade-off is that a ready-made tool works the way its maker designed it, not necessarily the way your business works, so you adapt your process to the tool.

Having something built around your own workflow is the other route. It follows your process, uses your approved information, and fits the checks your team already does — but it commits the business to defining that process clearly, keeping it maintained, and having someone responsible for it. That commitment is worth making only when the task is central to how the business runs, not for a one-off experiment.

The choice overlaps with the wider automation question. A generative AI tool that drafts or summarises is one piece; the surrounding workflow — where the information comes from, who reviews it, and where the result goes — is the automation around it. For a closer look at how small businesses can think about AI automation in general, see https://websphereinnovations.org/ai-automation-small-business/ . Where a business needs help assessing a more tailored approach, WebSphere Innovations also describes https://websphereinnovations.org/ai-automation-services/ and https://websphereinnovations.org/enterprise-software-development/ .

7. Risks and Guardrails: Keep People Accountable

Generative AI can assist with real work, but businesses need safeguards that reflect how the tool is actually used.

The main risk is not simply that an AI system will produce an obviously bad answer. It may produce a polished, confident answer that contains a mistake, an unsupported claim, or a missing condition.

Wrong or invented information reaches customers

Generative AI can produce statements that are inaccurate, unsupported, or inconsistent with the information the business has approved.

A draft might describe a service incorrectly, present an assumption as a fact, or include a detail that does not apply to the customer.

Practical guardrail: Check factual statements against approved information before publication or delivery. If a claim cannot be verified, remove it or investigate it before the work proceeds.

Customer or business information enters an unsuitable tool

Employees may paste customer details, internal documents, commercial information, or confidential correspondence into a tool without considering whether that tool is appropriate for the information.

The fact that a tool accepts information does not establish that the business is authorized to share it, or that the tool will handle it appropriately.

Practical guardrail: Establish which tools employees may use and what categories of information they may enter. Where the business cannot establish that a tool is appropriate for the information involved, do not use it for that information.

Employees trust drafts they have not read

A polished response can create a false sense of confidence. Employees may assume that generated text is correct because it reads well.

That assumption can allow errors to move from a draft into a final document or decision.

Practical guardrail: Make review an explicit step in the process. Tell employees what they must check, and make it clear that sending or approving AI-generated work without reading it is not acceptable.

The reviewer should understand the task well enough to identify a problem rather than simply scan the text for obvious mistakes.

The business loses track of who made the final decision

If AI-generated recommendations are treated as decisions in their own right, accountability becomes unclear.

Practical guardrail: Keep a named person responsible for approving important outputs. Document the requirements, the review step, and the person who approved the final version.

The team uses AI without knowing whether it works

An experiment can become routine simply because it has been introduced, even when its output is unreliable or its use is inconsistent.

Practical guardrail: Review the pilot against its original requirements. Continue only when there is evidence that the process works, and stop or change it when the evidence does not support it.

The appropriate safeguard depends on the application. A preliminary brainstorming exercise does not need the same level of review as a customer-facing document or a decision that affects a person’s employment.

8. Frequently Asked Questions

What is generative AI in business?

Generative AI in business means using AI systems that can produce new content, such as text, summaries, images, or code, to assist with a defined business task. It is most useful when it works from approved information, follows a clear brief, and is reviewed by a person who understands the work.

Does widespread AI adoption mean businesses are getting strong results?

No. McKinsey’s 2026 The State of AI survey found that nearly nine in ten respondents reported regular AI use in at least one business function, but only 37% attributed at least some EBIT impact to AI use, and only 6% met McKinsey’s definition of AI high performers. Adoption and business impact are different measures.

Why do so many AI pilots struggle to produce measurable results?

A pilot may fail to become useful when nobody owns the outcome, the tool is disconnected from the real workflow, no one checks the output properly, or the experiment quietly stops being used. MIT Project NANDA’s preliminary 2025 research also highlighted the difficulty of moving pilots into production and producing measurable profit-and-loss impact.

Which business functions are most suitable for generative AI?

McKinsey’s July 2024 survey, published in April 2025, identified marketing and sales as the most commonly reported function, at 42%. Product and service development, service operations, software engineering and IT were also important areas. The right function for a particular business depends on its repeated work, the information available, and the review process it can maintain.

How should a small business start using generative AI?

Choose one repeated task, define the standard for acceptable output, and assign a person to review the work. Run the process for a fortnight, record the results, and decide based on evidence. Expand only after the first task works.

9. Turn AI Experimentation Into a Practical Business Plan

Generative AI deserves attention, but a business does not need to adopt every new capability to make progress.

The research from McKinsey and MIT Project NANDA points to the same practical challenge from different directions. AI is widely used, and individual employees often find it helpful, but turning that use into reliable business value requires a defined task, approved information, a responsible reviewer, and evidence from actual work.

Start with one task. Set a clear standard. Review the output. Learn from the evidence before expanding.

If your business is exploring a more structured approach to AI automation, WebSphere Innovations can help you assess what fits your workflow. Visit https://websphereinnovations.org/contact-us/ to get in touch.

Leave a Reply

Your email address will not be published. Required fields are marked *