AI in Mid-Sized Companies, an Honest Look Without the Hype
Where language models deliver real value for mid-sized firms today and where they do not. A sober take on cost, data protection and a human in the loop.
The Hype and the Real Problem
Few technologies became a sales pitch as fast as generative AI. Many offers now carry the label "AI-powered", often without anyone able to explain what the software actually does and why it should beat the solution that came before. This is exactly where the problem starts for mid-sized companies. Buying a tool because the label sounds good means paying for a promise instead of a benefit.
An honest look begins with what these models technically are. A large language model uses vast amounts of text to compute the most likely next word. It understands no subject matter and checks no truth. This is not a flaw that gets trained away, it is the way the thing works. Almost everything worth saying about sensible and pointless use follows from this single fact.
A language model is an exceptionally good suggestion generator and a poor witness. It phrases, it does not prove.
Where AI Delivers Real Value Today
The strengths lie wherever language is processed and a person checks the result in the end. Three areas hold up well in practice.
First, text drafts. A first version of a product description, a reply template in support or the outline of a report appears in seconds. The value is not the finished text but the disappearance of the blank page. Editorial responsibility stays with the person who trims, corrects and approves.
Second, classification. Sorting incoming messages by urgency, routing requests to the right departments, pre-sorting documents by type. Models shine here because the result is verifiable and a mistake at worst costs a reassignment, not a false fact in a contract.
Third, search and summary across an own corpus. Instead of guessing keywords, a question in natural language finds the relevant passage in internal documents. What matters is grounding in real sources, so that every answer points to a readable document rather than being invented.
All three share one trait. The person stays in the loop, and a mistake is visible before it does harm.
Where AI Does Not Hold Up Today
The counter-list matters just as much. Wherever a statement must be correct and nobody checks it, a language model is the wrong choice.
- Legally binding or medical advice without professional review.
- Numbers, prices or deadlines that flow straight into an offer unchecked.
- Fully automated decisions with legal effect on individuals, which GDPR Article 22 restricts heavily anyway.
- Tasks where an invented but plausible-sounding answer costs more than no answer at all.
Hallucination is the core issue here. A model does not say "I do not know", it returns a fluent answer that sounds as confident as a correct one. This missing reliability can be dampened through source grounding and review steps, but never brought to zero. Ignoring that builds a system that helps in 95 percent of cases and quietly asserts something false in the remaining 5 percent.
A Sober Value-Risk Table
The table below sorts typical use cases by maturity and the oversight they require. The figures are experience-based and should be read as orders of magnitude, not as a guarantee.
| Use case | Maturity today | Human oversight | Time saved |
|---|---|---|---|
| Text draft for routine content | high | approval before sending | 40 to 60 percent |
| Classification and sorting | high | spot check is enough | 50 to 70 percent |
| Search across own documents | medium | source is included | 30 to 50 percent |
| Factual answers to customers | low | mandatory review of every answer | not advised |
| Autonomous decision | very low | legally constrained | not advised |
A glance at the oversight column shows the pattern more clearly than any marketing slide. The higher the risk of a false statement, the more tightly the human has to stay involved, and the smaller the honest automation gain becomes.
Costs That Rarely Make the First Slide
The price per request to a model looks tiny, often a fraction of a cent. That number hides the actual cost. Three items decide whether it pays off.
The first item is integration. A model on its own does nothing. It has to be connected to the company's own data, to existing systems and to a review workflow, and that is where the effort sits. The second item is ongoing upkeep. Results need to be monitored, input templates sharpened and failure cases worked through. The third item is human review, which never fully disappears in serious applications and belongs in the budget as a fixed operating cost, not as a short-lived transition.
Honestly calculated, automation only pays off above a certain volume of similar cases. With ten individual cases a month, the manual route is cheaper. With a thousand similar cases, the maths tips clearly the other way.
Data Protection and the Human in the Loop
Once personal data enters the picture, GDPR applies without exception. Every processing step needs a legal basis under Article 6, and anyone using an external model needs a data processing agreement with the provider. Customer data does not belong in a public chat service unchecked, where inputs may be reused for training.
Three routes lead to a clean position here. Anonymisation before processing, a provider with contractually assured data storage in the EU, or a self-hosted model on own infrastructure. Which route fits depends on protection needs and budget, and that trade-off belongs at the start of a project, not at the end.
Above all sits a principle we treat as non-negotiable. The human stays in the loop. AI may prepare, suggest and sort. Responsibility for a statement towards customers, authorities or partners rests with a person, not a model. This clear line protects legally, and it keeps quality high, because someone remains accountable.
What an Honest Start Looks Like
The useful path is unspectacular. First pick a single, clearly defined use case where a mistake is visible and cheap. Make it measurable, meaning decide in advance how success will be recognised. Then build small, review, and only expand once the benefit is proven. Whatever cannot be measured does not get rolled out.
This is exactly what we mean by "think further". First the honest question about real value, then a planned, narrow slice, then clean execution with a human in the loop, and only then the next step. More on this stance and the four movements "think further, plan further, build further, go further" is on the Mission page. Anyone wanting to assess a concrete use case in their own company will find the direct route to a sober conversation via Contact.
Conclusion
AI in mid-sized companies is neither a miracle cure nor a danger, it is a tool with clearly defined strengths. It helps with drafting, sorting and searching, and it fails wherever truth is assumed without oversight. Tuning out the hype, calculating honestly and keeping the human in the loop earns real time. Buying on buzzwords pays for a label.
Is a concrete AI use case worth it for a given company? We assess it soberly and say honestly where the value lies and where it does not.