As SMEs embrace AI, rethinking work and judgment becomes essential
For SMEs, generative AI can be a blessing – removing one of the traditional constraints on business: the cost of producing an answer. But as answers become abundant, the competitive advantage shifts elsewhere, writes Aylin Aydin of Mondial & Co, who argues that value increasingly lies not in more information, but in applying judgment, context and human-led strategic thinking.
For most of business history, producing a credible answer was expensive because the answer carried the cost of finding information, interpreting it and converting it into a form that someone else could use. A market assessment required research; a proposal drew on commercial experience; a customer response depended on an employee who understood both the question and the relationship behind it. Even when the result was imperfect, the effort involved imposed a useful constraint: organisations had to decide which questions deserved attention.
Generative AI weakens that constraint. A small company can now produce a market summary, cash-flow commentary, campaign concept and customer reply in the time previously required for one of them. This is a genuine expansion of capacity, but it also changes the nature of the bottleneck.
When plausible material becomes abundant, the scarce capability is no longer production alone. It is the ability to distinguish a useful answer from a merely fluent one, connect it to the realities of the business and decide whether the organisation should act on it.
This distinction matters more for SMEs than the usual debate about whether AI will replace jobs. Smaller firms rarely contain surplus layers of knowledge. The same employee may manage a client relationship, recognise an operational exception and know why a process that appears inefficient was designed that way.
If AI adoption is treated principally as a route to headcount reduction, a firm may remove precisely the context it needs to evaluate the new volume of output. The managerial task is therefore to redesign work around a new division of labour: machines can extend the organisation’s capacity to produce and compare, while people retain the responsibility to interpret, challenge and commit.
What the SME evidence shows
Türkiye makes the operating question unusually concrete. According to TÜİK, SMEs represent over 99% of enterprises and 68% of employment. Their use of AI therefore concerns far more than a narrow group of technology-oriented companies: it affects how a large share of the country’s productive knowledge is organised, transferred and retained.
Adoption is growing, although it remains uneven. TÜİK’s 2025 enterprise survey found that 7.5% of firms with at least ten employees used AI, with adoption materially higher among larger organisations. The gap is usually described as a technology-access problem, yet scale also changes what a company can place around the technology: cleaner data, formal approval structures and enough management capacity to redesign a process rather than simply add a tool to it.
International evidence helps explain what may happen after adoption. In a 2025 OECD survey covering more than 5,000 SMEs in seven countries, 65% of Generative AI users reported better employee performance, while 83% reported no change in overall staffing needs. The combination is more informative than either figure alone: AI can expand what an existing workforce is able to do without making workforce removal the principal source of value.
These findings suggest that the first organisational effect is usually a redistribution of capacity rather than a simple substitution of labour. Employees can handle more cases, explore more alternatives or bring work back inside the firm, but the economic value depends on how management uses the released capacity.
The Turkish and international datasets should also be read with discipline: TÜİK establishes the extent and distribution of adoption, while the OECD records firms’ reported experience. Neither proves return on investment at the level of an individual company. That requires a baseline and a direct comparison of cost, quality, revenue, risk and management time before and after a specific workflow changes.
AI is strongest where the work can be explained
Evidence from workplace studies shows that AI performs best when the task has an identifiable output, visible quality standards and enough repetition for previous examples to be useful. Under those conditions, it can reduce the distance between a less experienced employee and the organisation’s established practice. Its role is therefore broader than producing text quickly: it can become a mechanism for distributing knowledge that the company has already managed to codify.
The result changes when the task cannot be described or evaluated so cleanly. In an experiment with consultants, AI improved performance on assignments that fell within the technology’s capabilities, but made participants materially less likely to reach the correct answer on a task designed outside that frontier. The problem was not simply that the system could be wrong; every analytical method can fail. The more consequential problem was that fluency made the failure easier to accept.
Human review therefore cannot be placed at the end of the workflow as a ceremonial approval step. Judgment has to shape the task before the system is used, test the assumptions while the work is being produced and determine whether the result is suitable for the particular decision. The more ambiguous or consequential the task, the less credible a process becomes when responsibility is separated from production.
A job contains more than its visible outputs
A payroll-first AI strategy usually begins with the outputs management can count: reports written, cases answered, invoices processed or presentations prepared. Yet a role also contains judgments that rarely appear in a process map. Employees remember why a supplier exception was allowed, which client will reject a technically correct but commercially insensitive answer, which forecast assumption has been chosen for convenience and which minor delay is an early sign of a larger operational problem. These observations are not incidental to the role; they are part of how the firm manages uncertainty.
AI is most effective with knowledge that has already been expressed in data, documents and examples. Many SMEs, however, compete through knowledge that has never been formally recorded because it resides in relationships, routines and the memory of a small number of people. Removing those people before understanding what they know does not eliminate cost so much as transfer it. The cost returns later through slower exception handling, weakened customer relationships, repeated mistakes or greater dependence on senior management.
Cross-border business makes this distinction particularly visible. Language can be translated immediately, while commercial context is acquired through exposure. A response may be grammatically precise and still misread the customer; a market summary may describe demand accurately while overlooking the role of trust, hierarchy, payment practice or informal influence in the buying decision.
At Mondial & Co, we encounter this shift across assignments involving Türkiye, Central Asia and the Gulf. The same evidence can lead to a different recommendation in each market because the route from interest to commitment is different. AI can collect the material, compare alternatives and reveal patterns that deserve attention. The interpretation still depends on people who understand how a decision is actually made in that market, and who can defend the recommendation when its consequences become real.
The danger is producing more while understanding less
Many AI programmes are evaluated through activity: active users, prompts, generated documents or hours reportedly saved. Such measures are useful for tracking adoption, but they say little about whether the economics of the business have improved. A team can produce twice as many proposals without increasing conversion, answer customers faster while creating more escalations, or bring more analysis into a management meeting while making the decision itself less clear.
This creates a capacity paradox. The first draft becomes cheaper, so the organisation produces more of them; as volume rises, review, coordination and correction may expand elsewhere in the process. Employees spend less time creating material and more time establishing whether it is accurate, relevant and consistent with other work. Unless management removes low-value output or changes the decision process, local time savings can be absorbed by additional activity rather than appearing in profit, service quality or cash flow.
Early labour-market evidence illustrates the point. An NBER study linking adoption surveys with administrative records in Denmark found modest reported time savings, but no significant effect on earnings or recorded hours two years after adoption. This does not mean that the technology has no value. It means that saved time is an intermediate resource, not a business result. Someone must decide whether that resource will support higher service levels, deeper analysis, additional customers, shorter lead times or simply more work of uncertain importance.
Redesign work around the point of judgment
The practical response is to examine work at the level of decisions rather than job titles. A single role may contain highly repeatable production, pattern recognition, relationship management, exception handling and formal accountability. Treating the entire role as either automatable or protected conceals these differences. A better design identifies where judgment enters the workflow, what information that judgment depends on and how responsibility should remain visible when part of the work is delegated to a system.
1. Automate production while keeping accountability visible
Repetitive, standardised and low-consequence activities are sensible starting points: classification, extraction, routine updates and initial document preparation. Each workflow should nevertheless retain a named owner, an explicit route for exceptions and a threshold beyond which the output cannot advance without review. Accountability becomes more important when production is distributed across people and systems because errors otherwise become difficult to trace and easy to normalise.
2. Preserve the knowledge that the process map cannot show
Before changing a role, management should identify the exceptions, relationships and contextual signals the employee handles, then decide how that knowledge will remain available to the organisation. Some of it can be documented and incorporated into the system; some requires apprenticeship, customer exposure or continued human ownership. If the firm cannot explain how the knowledge will survive, the apparent saving includes an unpriced capability loss.
3. Measure the decision, not the volume of output
Evaluation should follow the purpose of the workflow. Relevant measures may include correction rates, conversion, service recovery, cycle time, cash impact, avoided external cost and the share of outputs requiring substantial human revision. These indicators reveal whether AI changes an economic or operational result. Document counts and reported hours saved remain secondary unless the organisation can show where the additional capacity went.
What an AI project should produce
An SME should expect more from an AI initiative than a collection of licences and trained users. The project should leave behind a small set of operating assets that make the change visible, measurable and governable after the initial enthusiasm has passed.
1. A redesigned workflow
The first output is a before-and-after view of the selected process. It should show which activities move to the system, where employees add context, what information is required at each stage and which decisions remain outside the scope of automation. This turns AI from an individual productivity aid into an explicit part of the operating model.
2. A decision and accountability map
Every consequential output needs an owner who can approve, reject or escalate it. The map should define decision rights, review thresholds and the route for exceptions, including the situations in which the process returns to a fully human path. The purpose is not to add bureaucracy, but to prevent responsibility from disappearing between the user, the model and the software provider.
3. A knowledge register
The company should record the internal knowledge on which the workflow depends: customer exceptions, commercial rules, market context, risk signals and the people who currently hold them. The register reveals what can be documented for wider use and what still depends on experience, relationships or professional judgment. It also allows management to develop succession and training plans before changing roles.
4. A pilot scorecard and scale decision
The pilot should end with an evidence-based decision to scale, redesign or stop. The scorecard should compare the new workflow with its baseline using a limited set of business measures, such as cycle time, correction rate, conversion, service recovery, cash impact and management review time. A successful pilot is one that improves the economics or reliability of the process, not simply one in which employees used the tool.
What this means under economic pressure
The argument becomes more demanding when financial conditions are tight and domestic demand is slowing. The Central Bank of the Republic of Türkiye’s August 2026 Inflation Report describes demand conditions as supporting disinflation and maintains a restrictive policy setting. Under such pressure, SMEs have legitimate reasons to seek lower operating costs and greater throughput.
Yet the margin for implementation error is also smaller. A company with limited financial slack may not be able to recover quickly from a damaged customer relationship, a weak credit decision or the loss of an employee who held critical operational memory.
The strongest near-term use cases therefore expand the reach of a lean organisation without making its risk-bearing capacity thinner. AI can help a team examine additional markets, shorten response times, reduce selected contractor costs and make established internal practice available to less experienced colleagues. Those gains are more likely to endure when the project begins with a process constraint or service objective. When it begins with a payroll target, management has an incentive to count visible labour cost before it understands the less visible work that the role performs.
