AI without a knowledge base produces interchangeable results

Linguistic quality does not replace robust corporate knowledge

Company-specific quality does not arise solely from a powerful model or a careful prompt.

It requires verified knowledge sources, clear terms, suitable examples, technical rules, authorizations, updates, and approvals. Without this foundation, generative AI often produces plausible average statements. They sound professional but hardly differentiate the company and can be factually or commercially incorrect.

Starting point

In April 2023, the discussion had changed within a few months.

ChatGPT had opened up broad access to generative language models. In March 2023, OpenAI introduced GPT-4. The outputs became more powerful, structured, and convincing in many tasks.

This created an obvious expectation:

As the model improves and the prompt is precise, the result will automatically become company-specific.

This expectation overlooked a fundamental difference.

A model can process general language and knowledge patterns. However, it does not automatically know:

• the current service definition of a company,
• internal technical terms,
• binding process rules,
• released evidence,
• Target group definitions,
• Exclusions and limitations,
• individual tonality,
• confidential experience values,
• applicable responsibilities,
• the source with precedence.

If this information is missing, the model fills gaps with general patterns. The result can be linguistically coherent and yet interchangeable.

What lies behind the problem

Model knowledge is not company knowledge

Language models are trained on large amounts of existing data. The knowledge contained therein is neither controlled company documentation nor a guarantee of timeliness and accuracy.

For a general explanation, this may be sufficient. For a binding statement about a specific company, it is not.

Example:

A model can describe what a digital strategy usually includes. It does not automatically know which services goeke.digital understands by this, which target groups are addressed, which method is used, and which terms are deliberately avoided.

Good language masks missing foundation

Weak raw texts are quickly noticed. Linguistically convincing texts are riskier if their factual basis remains unclear.

A generated text can:

• supplement an inappropriate service,
• describe a target group too broadly,
• claim an advantage that is not substantiated,
• use a technical term differently than the company,
• adopt outdated information,
• mix several sources contradictorily.

The linguistic quality makes it difficult to detect such errors because the text initially appears complete and plausible.

Prompts carry limited context

A prompt can specify the goal, format, tone, and individual facts. However, it quickly becomes confusing if every relevant detail, every exception, and every piece of evidence must be repeatedly incorporated.

Furthermore, a maintenance problem arises:

• Which version of the prompt is current?
• Who is allowed to change it?
• Which statement was taken from which source?
• Which rule applies to which use case?
• How are updates transferred to all prompts?

A prompt is an instruction. It is not a curated knowledge architecture.

Loose documents are not a reliable knowledge base

Companies often possess numerous documents:

• Website texts,
• Presentations,
• offers,
• Process descriptions,
• Training materials,
• E-mails,
• FAQ,
• internal notes,
• Contracts,
• Product information.

The volume of these files does not prove that an AI system can derive a correct answer from them.

Documents can:

• be outdated,
• use the same term differently,
• contain contradictory statements,
• only apply to an internal context,
• have unclear authorship,
• contain sensitive information,
• have no priority rule.

Before knowledge is technically retrieved, it must be factually clarified which source is binding for what.

Strategic Classification

Research into Retrieval-Augmented Generation had already described an important approach in 2020: a generative model is connected to an external retrieval system. This allows it to draw on additional knowledge from a document collection.

This approach was strategically relevant in April 2023 because it addressed two limitations of purely parametric models:

• Knowledge cannot be kept arbitrarily up-to-date and precise within the model itself.
• Origin and updating of external information remain open tasks.

However, RAG did not solve these tasks automatically.

A system can retrieve the wrong documents. It may fail to find important sections, prioritize outdated sources, or mis-summarize a found text. Even a correctly retrieved section may be unsuitable for the specific question.

Therefore, a company-specific knowledge base must do more than just search.

It must define:

1. which knowledge domain is covered,
2. which sources are permitted,
3. which source has priority,
4. which terms are binding,
5. which examples the application must understand,
6. which rules and limits apply,
7. who is technically responsible,
8. which content must not be used,
9. how updates and conflicts are handled,
10. which results must be human-approved.

Perspective from practice

In practice, work does not begin with a complete company database.

A clearly defined knowledge domain is more sensible. Examples:

• a specific service offering,
• a product area,
• a recurring consulting process,
• a defined FAQ section,
• a technical topic library,
• internal onboarding,
• a clearly defined support case.

Existing information is organized for this domain.

A typical finding:

• The website uses a different term than the offer.
• A presentation contains an older service level.
• An internal text explains an exception that is not intended to apply publicly.
• An example is helpful, but not released for publication.
• A subject matter expert has the current knowledge but no documented source.
• Two documents contradict each other and no one has established precedence.

This work initially seems less spectacular than selecting a model. However, it decides whether the later application can correctly map the company.

The system model of the AI knowledge base

Seven building blocks of a reliable knowledge base

1. Binding sources

Not every document has the same status.

A simple classification can distinguish:

• binding,
• supplementary,
• historical,
• under review,
• contradictory,
• locked.

The status must be recognizable to humans and technical systems.

2. Terminology standard

Services, target groups, products, and processes require clear designations.

Synonyms can be useful for search and communication. Internally, however, it must be clear which term has which meaning.

3. Technical Rules

An AI application needs more than facts.

It needs to know:

• which statement applies under which condition,
• which exception exists,
• which conclusion must not be drawn,
• which information may be missing,
• when it should not answer.

Rules must be formulated as testably as possible.

4. Examples and counter-examples

Examples show how knowledge is applied.

Counterexamples are just as important. They mark:

• impermissible formulations,
• unsuitable cases,
• typical misinterpretations,
• Performance limitations,
• necessary escalations.

5. Permissions

Not all knowledge may be used in every application.

It must be clarified:

• which people have access,
• which systems are allowed to access the content,
• which data is transferred to external services,
• which content remains confidential,
• which results are logged.

6. Update and Conflict Rule

Knowledge changes.

Every central piece of information requires:

• Responsibility,
• test interval or reason for change,
• Version status,
• Source with priority,
• Rule for contradictions,
• documented replacement of old content.

7. Technical Approval

The knowledge base can increase the quality of results. It does not negate professional responsibility.

Depending on the application, outputs will:

• fully checked,
• checked on a random basis,
• only used internally as a draft,
• escalates with uncertainty,
• automatically processed according to clear rules.

Knowledge quality must be evaluated after application

A knowledge base is not abstractly "good" or "bad." Its quality depends on the specific task.

For internal orientation, short, clearly marked sources may suffice. For a public expert answer, reliable evidence, current approvals, and unambiguous statement limits are required. For an automated system action, structured rules and secure termination conditions must also be present.

This distinction prevents two errors: unnecessary perfection for simple design tasks and insufficient diligence for consequential applications.

The role the prompt continues to play

A good knowledge base does not make prompts superfluous.

The prompt describes:

• the specific task,
• the desired recipient,
• the format,
• the current context,
• the applicable test criteria.

The knowledge base, on the other hand, provides:

• binding facts,
• Terms,
• Rules,
• examples,
• Limits,
• Sources.

Both levels must be maintained separately. Otherwise, company rules will be copied into individual instructions and diverge.

What companies should not do

Companies should not upload all existing files to a retrieval system without checking them.

This does not resolve contradictions, but makes them technically available.

You should also not try to control every answer exclusively through an ever-longer system prompt. Such a construction will be difficult to maintain and will remain limited by missing sources.

Likewise, the assumption that a source citation proves the correctness of the answer is also false. A system can cite a source and still interpret it incompletely or incorrectly.

Finally, companies should not build a knowledge base without clarifying responsibility and maintenance. A one-time import ages just like any other document collection.

Consequences for companies

Company-specific AI quality is a knowledge and governance task.

The model remains important. It decides on language ability, context processing, and technical possibilities. However, the business difference is created by connecting it with the right knowledge and controlled use.

A reliable foundation not only improves AI outputs. It also strengthens:

• Website and professional communication,
• internal orientation,
• Onboarding,
• Service,
• Sales documents,
• Search and answer capability,
• Consistency between teams.

This means that knowledge work does not become a technical side project. It becomes the common basis for several digital systems.

Subject-matter connection

Prepare corporate knowledge for AI applications

The SDC Knowledge Core organizes sources, terms, rules, examples, authorizations, and approvals for a clearly defined knowledge domain. Only on this basis is it decided how the knowledge is technically accessed and integrated into AI-supported processes.

Build an AI knowledge base in a structured way

Author

About Göke Frerichs

Göke Frerichs has been combining digital strategy, communication, technology, and implementation since 1999. As a digital strategist and Smart Digital Creative, he supports owner-managed B2B companies in developing clear and reliable digital systems from individual measures. His perspective is based on many years of consulting and implementation experience in the DACH region and North America.

More about Göke Frerichs