Samsung published preliminary guidance for its third-quarter results on 8 October 2026. These estimates should be distinguished from the detailed results that follow. [1] The announcement provides an occasion to ask a broader question: how does value created in the technology supply chain translate into benefits for the organisations buying the tools and the people using them?
One company’s accounts cannot establish AI’s impact across the economy. Our argument is about the standard of evidence leaders should apply: any claim of progress needs to identify its beneficiaries, time horizon and costs. Without that discipline, a financial story, an operational experiment and a promise about better work can become misleadingly interchangeable.
Revenue, productivity and welfare need separate evidence
An infrastructure provider can earn revenue while a customer is still testing a use case. An employee can complete a task faster without the business knowing how to use the time released. A customer can receive a quicker response without receiving a more accurate one.
These outcomes may reinforce each other, but none proves all the others. Leaders should identify which outcome they are discussing and show how it was measured. Growth in a technology market is not, by itself, an account of the effects on work or customers.
Exposure is not a forecast of job losses
The ILO and NASK’s 2025 work measures potential occupational exposure to generative AI, not observed redundancies. It points to changes in tasks as an important possible outcome. [2] This distinction should remain visible when findings about the future of work enter corporate presentations.
An organisation can begin with specific activities: preparing a report, checking information or handling a customer request. Some stages may accelerate while others require more review. Broad job titles conceal these shifts. A task-level discussion is more useful for deciding what skills, controls and resources a deployment will require.
Count the work that moves elsewhere
In a hypothetical pilot, a report takes one hour to draft instead of two, but checking sources and correcting errors adds forty minutes. There is still a gain, though it is smaller than the drafting figure suggests. Licensing, training and process design introduce additional costs.
A meaningful comparison includes production, validation and rework at a comparable level of quality. It also checks whether effort has simply moved to another team. Communications leaders should ask for this full account before presenting a pilot result as evidence of organisation-wide productivity.
Sharing the gains is a management choice
Verified gains can support different decisions: improving service, reducing delays, taking on more work, investing in skills or changing staffing. Technology does not make that allocation on its own. Leaders should explain the choices they are considering and the consequences for affected groups.
Employees want to understand expectations and prospects. Customers may expect a service benefit. Investors will examine returns and investment needs. One upbeat message cannot answer all these questions. A credible narrative connects the measured gain with the decision about where its benefits will go.
International deployment needs local evidence
An English-language demonstration does not establish performance on French documents, Korean internal materials or a different customer base. Businesses should test the actual workflow, language and decision process before treating a global pilot as a local success.
For a company coordinating an Asian headquarters and a French subsidiary, a useful review would identify who approves the use case, who validates local outputs and who can stop a deployment. This is a hypothetical organisational example, not a claim about any named company. The aim is to make accountability visible across borders rather than assume that one central approval resolves every operational question.
Training includes knowing when to reject an output
Being able to produce a response quickly is only part of competent use. Teams need to recognise when a result requires verification, when a source is missing and when human judgement should take over. Otherwise, higher output can coexist with weaker control over commitments.
Training can start with errors observed in real tasks and specify review responsibilities. Staff should be able to report a problem without being penalised for detecting it. The relevant capability is collective: use, check, correct and sometimes decline to use the tool. Attendance at an introductory session is an incomplete measure of that capability.
Experimentation does not require certainty in public claims
Waiting for perfect knowledge could prevent useful trials. But accepting uncertainty in an experiment does not justify describing its results as settled. An organisation can specify a limited scope, an explicit hypothesis and a review date.
A first test might establish feasibility; subsequent use can examine ordinary working conditions and robustness. Public language should reflect the stage reached. This approach preserves ambition while helping leaders avoid claims they later have to retract or qualify substantially.
Ask four questions before announcing progress
We propose a practical leadership review: is the final result better; does the gain survive verification costs; who bears the new burdens; and who receives a tangible benefit? The answers will differ between tasks, markets and business models. They should therefore be documented rather than assumed.
A transformation note can record the uses tested, quality criteria, costs included and decisions on the gains. It can also explain why some uses were abandoned. Such an account gives employees and customers a basis for judging the organisation’s method. AI becomes a credible story of progress when financial performance, operational improvement and benefits for people are connected by evidence, not merely placed next to one another in an announcement.
Sources and context
Read in another language
These editions address the same subjects with context adapted for their readers.