AI in SMEs: Business Case instead of Tool Hype

Core message

KfW Research finds a positive correlation between AI use and company performance, but no proof of causality. Therefore, your own business case counts for your decision: baseline, clear process, full costs, quality limit, and verifiable benefit.

Starting point

The question for many companies has shifted. It is less about whether generative AI can fundamentally process texts, analyses, code, or process steps. This has been shown in numerous applications. The more difficult question is: Which application is economically beneficial for one's own company?

The KfW analysis of September 15, 2026, examines AI usage in 2022-2024 and company key figures for 2024. For comparable AI users, revenue growth is on average 1.1 percentage points higher and profit margin 0.9 percentage points higher; labor productivity around 3.4 percent. The latter refers to revenue per full-time employee, not a measured time saving of an AI workflow.

Whether AI causes better performance or successful companies use AI more often remains open. The correlations with growth and profitability are mainly evident in companies that are already growing or have strong profitability. ^[1]

This evaluation therefore does not replace an examination of whether a specific use case is economically viable in one's own company.

What lies behind the problem

Usage is not yet impact

Many internal AI reports count licenses, active users, generated content, requests, or tokens. These are activity data. They show that a system is being used. They do not show whether a process is working more economically.

A team can generate more content and simultaneously require more time for review, correction, and coordination. An automated process can become faster, but cause higher follow-up costs due to errors. An agent can take over more work steps without the overall process lead time being shortened.

The technical price is only part of the cost

On September 29, 2026, OpenAI announced GPT-6.1 Sol with one-fifth of the regular input and output token prices of GPT-6 Astra, as well as permanently working agents. These are provider specifications regarding technical capabilities and prices. ^[2] Microsoft introduced additional Copilot features for delegated work and automation on September 25; the rollout will be gradual via the Frontier program. ^[3]

A lower model price can facilitate access to an application. The business case also includes:

• Data must be discoverable, current, and permissibly usable.
• Systems and interfaces must be connected.
• Permissions and authorizations must be defined.
• Results must be reviewed and errors processed.
• Processes, roles, and responsibilities may change.
• Consumption and follow-on costs must be controlled.

A cheap model call can therefore be part of an expensive or ineffective overall process.

Successful companies bring better prerequisites

KfW names digital infrastructure, data quality, competencies, and interdisciplinary cooperation as important prerequisites. Such foundations can also contribute to corporate performance without AI.

This explains why a statistical correlation does not automatically result in a transferable business case. Those who only copy the tool do not copy the organizational prerequisites.

Strategic Classification

The consequence for investment appraisal is: The availability of a model is only one prerequisite. In one's own application, four areas are crucial:

1. the selection of an economically relevant problem;
2. the quality of process, data, and knowledge;
3. the control of permissions, risks, and responsibility;
4. the ability to measure impact and make adjustments.

For medium-sized companies, this is good news. They do not have to follow every technical development. They need a reliable decision on which limited use case solves a real bottleneck and how success is determined.

Perspective for decision-making practice

An AI project should not start with the selection of a provider. The first checkpoint is a process where effort, quality problem, delay, or missed demand can be concretely described.

This allows us to decide whether AI is the right lever at all. Sometimes a clear template, a better data structure, or conventional automation is sufficient. In other cases, a model may be useful because information is variable, language needs to be processed, or multiple sources need to be merged.

The decision becomes robust when it is determined before the test which result should be better and which deterioration is not acceptable.

Framework: five checks for a robust AI business case

1. Record the baseline value

Without a baseline value, a claimed improvement is difficult to verify. For the current process, record at least:

• Processing and waiting time;
• direct internal and external costs;
• Errors, rework, and escalations;
• Quality criteria;
• Business key figures, such as qualified inquiries, closing rate, throughput, or avoided failures.

The baseline doesn't have to be perfect. But it must be sufficient to distinguish a relevant change from normal fluctuation.

2. Select a clear workflow

"AI in Marketing" or "AI in Customer Service" is too broad. A verifiable workflow has a defined input, a sequence of steps, an expected outcome, and a responsible person.

Examples of the necessary precision are not "automate content", but "generate a first draft for an FAQ from approved product documentation and have it reviewed by the product owner". Not "qualify requests", but "pre-sort incoming requests according to defined criteria without giving a binding commitment".

3. Define quality and risk limits

A faster process is not a success if unnoticed errors increase. Define in advance:

• which statements must be supported by sources;
• which actions require human approval;
• which data may not be used;
• at which error rate the test is stopped;
• who is responsible for correction and escalation.

A technical review is still necessary for legal, medical, financial, or safety-related statements.

4. Calculate full costs instead of model price

In addition to licenses and usage prices, consider:

• Process recording and conception;
• Data preparation and system integration;
• Tests and quality assurance;
• Training and change management effort;
• Ongoing monitoring, rework, and error costs;
• Internal coordination and responsibility time;
• Costs of changing provider or model.

Only these full costs can be compared with the expected benefits.

Free working time is initially additional capacity. It only becomes an actual cost saving when, for example, overtime or external expenses are eliminated. If it is used for additional orders, the achievable contribution margin and the necessary additional resources count. Do not count the same hour as both saved costs and additional revenue.

5. Measure results in a limited phase

A pilot needs a timeframe, a comparison basis, and a completion criterion. Four to eight weeks can be sensible as a planning framework for a narrowly defined process. This is a recommendation, not a general proof of effectiveness. The number of cases and process duration determine whether the observation is meaningful. A longer period may be necessary for revenue or customer loyalty effects.

The separation of key figures is crucial:

In the end, there are three legitimate decisions: expand, specifically revise, or terminate. Even a stopped pilot can be economically sensible if it prevents a larger misinvestment.

What companies should not do

• Explain AI usage based on licenses, tokens, or generated content to achieve success.
• Automate an unstable or unclear process.
• Scale an impressive demonstration to the entire company.
• Leave recommendation, execution, and final evaluation to the same model.
• Start a pilot without an owner, baseline, budget limit, and stop criterion.
• Confuse short-term time savings with long-term business value.

Consequences for companies

Statistical correlations provide a reason for examination. Whether the investment pays off must be shown by your own process.

Companies should therefore not ask how much AI they are using. The better question is: Which specific process improves which business result, under which conditions, and at what total cost?

Anyone who answers this question before technical implementation reduces risk and at the same time creates the basis for scaling. Once the first workflow delivers measurable results, further applications can follow using a proven decision logic.

Subject-matter connection

From application request to robust decision

With SDC-Discovery, I can help you clarify the initial situation, bottleneck, target vision, and a suitable initial testing area. The focus is not on the model, but on the decision of which process is economically relevant and realistic under the given conditions.

Clarifying the initial situation in an SDC Discovery

Business priorities and budget decisions remain with your management team. You contribute process knowledge, available data, and the responsible experts. Subsequent practical implementation and results verification are agreed upon separately.