Big Business Shows Small Firms What to Do - and Avoid - with AI

Google News surfaced a blunt warning from Gene Marks on August 30: small firms should copy enterprise AI wins, not enterprise spending habits.

His Guardian column argues that large companies have already paid for many costly experiments. Smaller businesses can now observe which projects deliver useful results and which consume money without changing operations.

That sounds like a simple advantage for small firms. However, the underlying evidence reveals a harder conflict. Large organizations use AI more often, yet only a small group reports meaningful companywide financial gains.

The divide is not between businesses that use AI and those that ignore it. It is between focused workflow changes and broad programs built around expectations, executive pressure, or fear of missing out.

Large companies have found practical uses in software development, customer service, security, and voice applications. They have also encountered rising computing costs, unreliable outputs, employee resistance, and disappointing returns.

Small companies face the same technical risks with fewer financial cushions. Their advantage is not better technology. It is the ability to choose a narrow problem, change one workflow, and stop quickly when results disappoint.

That distinction matters because AI adoption is rising across the economy. The winners will not necessarily be the businesses with the largest budgets. They will be the ones that connect spending to measurable work.

What Google News Left Behind the Headline

The central event is a shift from AI experimentation toward selective adoption based on evidence from large corporate deployments.

Marks published his argument in small business AI after several years of intense corporate investment. His point is observational rather than technological. Big companies tested AI at a scale that small businesses could rarely afford.

Some of those tests produced useful patterns. AI-assisted coding can help developers draft, review, document, and test software. Customer service teams can use models to summarize conversations or propose responses for human approval.

Security systems can use machine learning to flag unusual activity. Voice tools can transcribe calls, identify action items, and make recorded information searchable.

These applications share an important characteristic. Each begins with a recognizable task that already exists inside a business. The company can compare the new process with an established baseline.

A software team can measure review time, defect rates, and delivery speed. A support manager can monitor response quality, resolution time, escalations, and customer satisfaction.

A sales team can compare the time required to process calls and update records. Security staff can measure false alerts, investigation time, and the number of incidents caught.

The weaker projects often begin differently. Leaders announce an enterprise AI strategy before identifying the work that must improve. Teams then purchase tools, build infrastructure, and search for uses afterward.

That sequence creates expensive organizational theater. A business can accumulate pilots, presentations, vendor contracts, and training sessions without creating a dependable production workflow.

Google News makes these stories appear adjacent, even when their underlying evidence differs. A successful coding assistant deployment and an ambitious autonomous-agent program can both receive the label “enterprise AI.”

For a small business owner, that label is too broad to guide a decision. The relevant questions concern the task, the data, the human review process, and the economic result.

The original column also separates productivity from payroll reduction. Marks argues that small employers should use AI to increase employee capacity, not advertise immediate staff cuts.

That choice has operational value. Workers have less reason to hide problems when the stated objective is better output rather than fewer jobs. Their feedback can expose errors before customers encounter them.

It also reduces the temptation to count projected labor savings before a system works. A forecasted reduction is not realized value, especially when employees must spend time correcting unreliable output.

The headline therefore marks a broader change in AI decision-making. Adoption is moving from abstract enthusiasm toward choices about where automation deserves trust.

Small firms are not merely late adopters in this story. They are observers entering after the first round of corporate evidence has become visible.

Large Companies Have Adoption, but Value Remains Concentrated

Big companies hold a clear adoption advantage, while broad financial impact remains much harder to establish.

The United States Census Bureau reviewed business survey data collected from December 2025 through May 3, 2026. Overall AI use stayed between 17% and 20% during that period.

The same business AI data showed a major difference by company size. Thirty-seven percent of firms with at least 250 employees reported using AI.

Among businesses with 100 to 249 employees, 32% reported usage in the period ending May 3. Those figures confirm that larger firms still have more capacity to adopt new systems.

They can hire specialists, negotiate vendor agreements, prepare data, and create internal governance programs. They can also keep experiments alive longer when early returns remain uncertain.

Yet adoption does not establish value. It can describe anything from an employee drafting an email to a company rebuilding a major operational process.

The Federal Reserve highlighted this measurement problem in its April 2026 review. Census data showed about 18% of firms had adopted AI by the end of 2025.

However, an employment-weighted executive survey estimated that 78% of workers were employed by companies using AI. About 54% worked at firms using large language models.

The adoption measurement changes because surveys count different units. A small firm and a national employer each count as one business in firm-level data.

The result explains how AI can feel common at work while remaining absent from most companies. Large employers cover a substantial share of the labor force.

Question wording introduces another difference. The Census Bureau broadened its survey in November 2025 from production activities to AI use in any business function.

That change captures more activity, including administrative work. It also makes comparisons with older results less direct.

Small firms should treat adoption statistics as context, not as a mandate. A high percentage does not show that a particular tool suits their customers, data, or operating model.

Financial evidence is more revealing. McKinsey’s 2026 global survey received responses from 1,719 participants across 97 countries.

Thirty-seven percent attributed at least some earnings impact to AI. However, only about 6% qualified as high performers with significant impact and at least 5% of earnings tied to AI.

The enterprise AI results also showed that this high-performing share had not increased from 2025. Adoption widened faster than top-level financial results.

That gap creates the pressure behind this story. Vendors want broader deployment, executives want visible progress, and employees want tools that make daily work easier.

Finance leaders want verified returns. Security and legal teams want controls. Customers expect accuracy and responsible handling of their information.

A small business cannot resolve these competing demands with a larger strategy document. It needs a smaller economic unit that everyone can inspect.

That unit is a workflow. It has an input, an output, an owner, a quality threshold, and a cost.

For example, a support team might use AI to draft replies while an employee approves every message. The business can measure saved time and track correction rates.

It can also stop the test without disrupting the entire company. That reversibility is valuable because models, prices, policies, and vendor capabilities continue to change.

The strongest lesson from enterprise adoption is therefore not “use more AI.” It is “make each use case prove itself.”

Focused Workflows Beat Enterprise AI Theater

The primary contest is between measurable workflow redesign and AI programs that mistake visible activity for operational progress.

Workflow redesign means changing how a defined task moves from start to finish. It does not mean adding a chatbot beside an unchanged process.

Consider a sales call. A superficial deployment gives representatives a general writing tool and records estimated time savings.

A redesigned workflow captures the conversation, produces a structured summary, identifies commitments, and proposes customer-record updates. A person verifies the information before it enters the system.

The second approach creates clear checkpoints. Managers can measure whether records become more complete, follow-ups happen sooner, and employees spend less time on administration.

The same principle applies to internal knowledge. A company may own thousands of documents while employees still cannot locate the correct policy or project decision.

Connecting AI to unorganized files does not automatically solve that problem. The system needs appropriate access controls, current source material, and a way to show where an answer came from.

Employees also need a defined response when the evidence is missing. An AI system should express uncertainty rather than invent a plausible answer.

A small firm can test this pattern within one team. It might begin with approved sales materials, technical documentation, or internal procedures.

A searchable AI knowledge base can support that work when its sources and permissions match the workflow. It cannot replace information ownership.

Large organizations have shown where focused deployments often work. Software development is one example because code can pass through tests, review, and version control.

AI can propose code without receiving final authority. Developers can inspect changes and reverse them when necessary.

Customer service offers another useful pattern. Models can classify requests, summarize long conversations, retrieve approved information, or draft responses.

The risky step occurs when a company removes human review before measuring accuracy across ordinary and unusual cases. A fluent answer can still be wrong.

Voice applications follow a similar pattern. Transcription and summarization can reduce clerical work, but sensitive recordings require clear retention and access rules.

Security applications can detect unusual behavior across large volumes of activity. However, false positives still require investigation, and automated responses can interrupt legitimate work.

These examples explain why small business AI should begin with assistance rather than unchecked autonomy. Assistance keeps authority with someone who understands the customer and business context.

Autonomy raises the cost of every error. A model that drafts one poor message creates a review task. A model that independently sends it creates a customer problem.

The difference becomes greater when AI agents take actions across several systems. An agent is software that plans and executes multiple steps toward a goal.

An agent might read an email, update a database, prepare a quote, and schedule a follow-up. Each connection expands both potential value and potential damage.

Large firms can fund testing environments, access-control reviews, monitoring systems, and incident-response teams. Many small businesses cannot reproduce that infrastructure.

They should therefore reduce the system’s authority. Read-only access, narrow data scopes, approval gates, and transaction limits can make a pilot safer.

This approach is not anti-innovation. It treats reliability as part of the product rather than an obstacle added after deployment.

The organization should also measure total work, not only the minutes saved by a first draft. Review time, correction time, retraining, integration, and vendor management all belong in the calculation.

A tool that drafts content quickly can still create negative value if employees must verify every unsupported claim. Speed at one step can move labor elsewhere.

That is why the phrase “productivity gain” requires a denominator. Businesses should specify whether they mean more completed cases, shorter cycle times, higher quality, or lower operating cost.

They should also measure the result over enough time to capture exceptions. A demonstration using clean examples cannot represent an ordinary month of messy customer activity.

Google News coverage can make an isolated success look like a general formula. Small firms should look past the headline and identify the conditions that made the result possible.

Those conditions often include prepared data, engaged process owners, clear quality metrics, and executive support. The model itself is only one component.

Cost, Accuracy, and Trust Can Erase the Savings

AI creates value only when its operating costs and failure risks remain below the value of the work it improves.

Computing cost has become part of the adoption debate. AI systems often charge according to tokens, which are units used to process and generate information.

A short interaction may use few tokens. A long document, repeated agent loop, or large retrieval process can consume far more.

McKinsey found that one in five survey respondents said operating costs were limiting their organization’s AI use. The constraint included token expenses and appeared across company sizes.

About one in ten reported cost limits for each of three categories: chatbots, AI agents, and coding agents. These figures challenge the idea that software automatically becomes cheaper after deployment.

Small firms need cost controls before expanding a pilot. They should know the average expense per task and how usage changes during busy periods.

They should also test what happens when the system fails. Retries, longer prompts, repeated searches, and escalations can increase both computing and labor costs.

A process that appears affordable at demonstration volume may become uneconomic at production volume. The relevant metric is cost per accepted result, not cost per generated answer.

Accuracy presents another risk. Generative AI predicts outputs from patterns in data, which means confident language does not guarantee factual correctness.

A customer-facing error can trigger refunds, complaints, or lost trust. An incorrect financial or legal statement can create more serious exposure.

NIST’s AI risk framework gives organizations a voluntary structure for managing these problems. Its core activities concern governing, mapping, measuring, and managing risk.

NIST also released a generative AI profile in July 2024. The profile identifies risks specific to generative systems and proposes actions that organizations can adapt.

A small company does not need a large compliance office to apply the central logic. It does need a named owner and documented boundaries.

The owner should know what data enters the system, who can see outputs, and which actions always require approval. The business should also record failures.

That record turns anecdotes into evidence. If employees repeatedly correct the same error, the workflow needs different instructions, better source data, or a different tool.

Some problems cannot be solved through prompting. A model may lack current information, reliable access controls, or sufficient accuracy for a high-stakes task.

Stopping a weak project is a successful decision when the evidence supports it. Small firms should not preserve pilots merely because time has already been invested.

Employee trust matters as much as technical performance. Workers will resist tools presented mainly as a justification for headcount reductions.

They may also avoid reporting defects if every correction appears to weaken their own job security. That behavior deprives management of the information needed to improve the system.

The enterprise record offers a useful reality check. McKinsey reported that 14% of respondents at AI-using organizations saw AI contribute to an overall workforce decline during the prior year.

That was less than half the 32% who had expected reductions. Two-thirds reported little or no AI-related change in total employment.

Expectation and reality therefore remained far apart. Yet 39% still expected AI to reduce headcount during the following year.

Small businesses should resist counting those anticipated reductions as returns. A tool must work reliably before a company can redesign staffing around it.

Keeping employees involved produces better evidence. They know which cases are routine, which customers require care, and where source information becomes unreliable.

They can help define acceptance criteria before a vendor demonstration influences the project. They can also identify work that should never be delegated.

The skeptical view is that small firms lack the time, data, and technical expertise required even for disciplined pilots. That concern is legitimate.

A narrow test still requires process knowledge, evaluation, training, and oversight. Buying a familiar subscription does not remove those obligations.

However, limited resources make discipline more important, not less. A small company should choose fewer use cases and demand clearer evidence from each one.

It should also prefer reversible decisions. Exportable data, short evaluation periods, human approval, and limited permissions reduce the consequences of a bad choice.

These controls do not guarantee a return. They make failure visible before it becomes an enterprise commitment.

Small Business AI Has a Core-Work Problem

Small firms often use accessible AI around the edges of work, while deeper operational adoption remains limited.

The OECD’s 2025 review found a persistent adoption gap between small and large enterprises. Across member countries, 40% of firms with at least 250 employees used AI.

The figure fell to 20.4% among firms with 50 to 249 employees. It reached only 11.9% among firms with 10 to 49 employees.

Those figures used 2024 or the latest available national data. Definitions varied across countries, so the OECD cautioned against careless comparisons.

The broader pattern was still consistent. Smaller firms adopted AI less often than large firms across the G7.

The gap changed by application. European data showed a smaller difference for natural language generation than for autonomous machines.

That makes sense because text tools are easier to access. They often require no specialized hardware and can enter a business through individual employees.

Yet access does not equal operational depth. The SME adoption review found that only 29% of AI-using SMEs applied generative AI to core activities.

Many used it for peripheral tasks that supported operations without changing production. Marketing and sales stood out as an area where small firms used AI relatively actively.

Peripheral work can still deliver value. Drafting campaign variations or summarizing a meeting may save meaningful time.

The problem arises when a business mistakes scattered individual use for an operational capability. Employees may use different tools, upload different information, and apply inconsistent quality checks.

Managers then lack visibility into risk and results. The company cannot tell whether AI improved the workflow or simply changed how employees completed isolated tasks.

Shadow AI, meaning unapproved employee use of AI tools, also creates data concerns. Staff may paste confidential material into services without understanding retention policies.

A small firm needs a simple inventory before pursuing a larger strategy. Which tools are in use, what information enters them, and which outputs reach customers?

That inventory can reveal opportunities. Several employees may spend time summarizing similar documents or transferring the same information between systems.

It can also expose unsuitable uses. A worker may rely on generated advice in a context where every statement requires verification.

After mapping current activity, the company can select one workflow with enough volume to measure. It should avoid beginning with the most sensitive or complex process.

A useful candidate has repetitive inputs, a clear output, and a person capable of judging quality. The result should matter enough to justify evaluation.

The company should establish a baseline before adding AI. Without baseline time, cost, and quality data, improvement becomes a matter of perception.

Next comes a controlled pilot. The system handles a defined share of work while people compare its output with the existing process.

Evaluation should include ordinary cases and exceptions. Teams should record incorrect facts, missing information, inappropriate tone, access failures, and unexpected costs.

The business can then decide whether to stop, revise, or expand. Expansion should follow evidence, not executive impatience.

This sequence converts the Guardian’s observation into an operating principle. Let big companies expose broad patterns, then validate the relevant pattern inside a small business.

Google News can help owners discover those patterns. It cannot tell them whether a workflow fits their customers, records, and tolerance for error.

That judgment must remain local. It depends on the actual work and on employees who understand its consequences.

Three Signals Will Show Whether the Lesson Holds

The next phase will be decided by verified workflow returns, controlled operating costs, and the gap between expected and actual workforce change.

The first signal is whether more organizations report companywide earnings impact. McKinsey’s high-performer group remained near 6% in its 2026 survey.

If that share rises, the evidence for repeatable AI value will strengthen. The details will matter more than the headline percentage.

Businesses should look for workflow redesign, defined performance measures, and risk controls behind the gains. If results remain concentrated, broad adoption will still exceed proven transformation.

For a small company, this signal affects expectations. A larger group of high performers would provide more tested patterns to examine.

A flat result would support continued caution. It would suggest that access to better models alone does not solve organizational problems.

The second signal is the cost of production AI. One in five McKinsey respondents already reported limits linked to operating costs.

Watch whether vendors reduce inference costs and whether companies improve control over agent usage. Inference is the computing work required to produce an AI output.

Lower unit costs would make high-volume tasks more attractive. However, cheaper outputs can also encourage unnecessary usage and hide weak process design.

The better signal is falling cost per accepted result. That measure combines model expense with quality and human review.

If token expenses fall while correction work rises, the business has not created a dependable saving. If both costs decline, scaling becomes more credible.

The third signal is the difference between workforce forecasts and realized employment changes. Prior expectations substantially exceeded reported reductions in the 2026 survey.

A continued gap would weaken claims that AI automatically removes large numbers of roles. It would support the productivity-first approach described by Marks.

If verified reductions begin matching forecasts, small firms will face different planning questions. They will need to understand which workflows changed and which capabilities remained with employees.

Either outcome requires more precision than an earnings-call promise. Headcount can change because of demand, restructuring, outsourcing, or earlier overhiring.

Businesses should not attribute every reduction to AI without evidence. They should examine output, service quality, workload, and financial results together.

For readers arriving through Google News, the practical action is straightforward. Choose one recurring workflow and write down its current time, cost, quality, and owner.

Then decide what an acceptable AI-assisted result would look like. Keep a person responsible for approval and track every correction during the test.

Ask whether the system improved completed work, not whether employees generated more text. Review total cost, including oversight and failures.

If the evidence stays weak, stop. If it holds across ordinary and difficult cases, expand carefully.

Big business has already shown that AI spending can produce both useful systems and expensive disappointment. Small firms now have a chance to copy the discipline without copying the scale.