Comparing n8n and custom GPT

Purpose of this research

The purpose of this research is to compare the n8n solution with the custom GPT solution. Both solutions can help Sales collect company information and assess incoming requests. However, they use different sources and work in a different way. This research investigates which approach is more suitable for Bierens and what the advantages and limitations are.

The comparison focuses on cost, reliability of the information, privacy, security, flexibility and integration with the Sales Assessment Tool. It also investigates which tasks can be automated and which results still need to be checked by an employee.

Main finding

Based on the technical comparison, n8n is the stronger option for connecting the different steps of the Sales process. It can retrieve company data, apply agreed business rules and send the result to a dashboard. The custom GPT is useful for an initial company overview and explaining information, but information from the open web still needs to be checked.

The n8n approach is expected to provide better-supported results when it uses approved company data, KVK information and a credit report. This does not mean that the complete result is automatically 100% accurate. The correct company must still be selected, the data must be up to date and the business rules must be implemented correctly.

Status of the research

This document contains the comparison between the custom GPT and n8n solutions, including initial testing results and the proposed next steps. The custom GPT was tested by the Sales team on 20 Dutch companies and correctly classified 16 of them, achieving 80% accuracy on this test set. Testing on international companies is still ongoing.

The n8n prototype is showing promising results because it uses verified company information from APIs, such as KVK, instead of relying only on publicly available web information. However, its overall accuracy has not yet been measured, and using paid APIs makes the solution more expensive.

The cost estimates provided by Laurens are included in this comparison. Further testing is needed to compare both solutions in terms of accuracy, costs and time savings before making a final conclusion.

1.1 Project context

As established earlier Bierens receives approximately 20,000 requests per year, according to the project plan. Employees currently retrieve company information and read case descriptions manually. This takes time, especially when a description is long or important information is missing. The project investigates how automation and AI can support this initial assessment. [P1]

The solution should retrieve relevant company information, process the case description and classify the request based on agreed business criteria. It supports Sales before the intake process. It does not replace legal decisions or the final judgment of an employee.

1.2 Link to the project plan

This comparison mainly supports research question 4 from the project plan: which technical approach is most suitable for the automated Sales process, and which steps should be handled by automation, business rules or AI? It also supports question 6, which looks at cost, security, privacy and the hosting of an LLM. The questions below are comparison topics within those project questions, rather than a replacement for them.

1.3 Research methods

For this comparison, I use document research and available product analysis to investigate the technical options. The project plan provides the business context. The cost information reported from Laurens about the custom GPT that we’ve built provides an initial estimate. Prototyping and testing will be used to compare the actual results. Finally, feedback from Sales and the development team will be used to check whether the proposed solution fits Bierens.

The technical documents describe what the platforms can do. They do not prove that my own solution already does this 100% correctly. Therefore, the comparison separates available features, expected benefits and results that still need to be tested nationally and internationally.

2.1 Custom GPT solution

My custom GPT uses instructions and reference files to guide the company research and classification. It can search for public information about a company and create a readable overview. An employee can ask follow-up questions or request an explanation of the result.

This makes it useful for an initial assessment, or an estimation of the company profile. However, public information may be missing, outdated or related to another company with a similar name. A company website may also describe the company in a positive way without providing evidence for every claim. Information being publicly available does not make it 100% reliable.

The employee should therefore be able to see which source supports each important field. If an employee count or financial figure cannot be confirmed, the result should show that the information is unknown or estimated. The GPT should not present an estimate as a verified fact.

2.2 n8n solution

The n8n solution connects different steps of the Sales assessment process into one automated workflow. I have developed a working prototype that starts with test input and processes the incoming request through several connected nodes.

The workflow first normalizes the incoming data, translates the case description when needed and uses AI to summarize long case descriptions. It then identifies the company and performs online company research to collect relevant business information.

An important improvement compared to the custom GPT is the integration of the KVK API. The workflow uses KVK company verification and activity-status checks to retrieve more reliable company information. This addresses one of the main problems found during custom GPT testing, where inactive companies were sometimes not correctly identified through public-web research.

After collecting the information, the workflow uses an Information Extractor to structure the company data and an Account Segmentation Classifier to recommend a DAM, AM, PAM or KAM category. The results are combined into a structured final report containing company details, business activities, available employee information, sources, missing information and classification reasoning.

A planned next step is to integrate a credit-information API into the workflow. This API would provide additional company information, such as credit scores or credit limits where available. Currently, this integration has not been implemented. Once access is available, the workflow could use this information to support more reliable company assessments and reduce dependence on incomplete public-web information.

The prototype combines AI processing with external data retrieval. AI is used for translation, summarization, company research and information extraction. Fixed business rules should be applied separately through workflow logic, especially when a decision depends on verified information.

The workflow has been executed successfully during initial testing, including an example where the KVK API identified an inactive company. However, the complete solution is not finished yet. The credit-information API, live website input, internal system connections, Sales dashboard and human-review process still need to be implemented or tested.

The current prototype demonstrates that n8n can combine AI processing, verified company information and classification into one workflow. Further testing is required to measure its accuracy and reliability compared with the custom GPT solution.

Account segmentation and case acceptance should remain separate. A DAM, AM, PAM or KAM label describes the account segment. It does not automatically mean that a specific case should be accepted.

The table compares the intended solutions. The costs are estimates and the other findings describe the expected behaviour. Actual performance still needs to be tested.

4.1 Cost information from Laurens

One company’s search cost approximately €0.40 for the custom AI Agent. A more complete assessment through the intended n8n solution, including credit information, KVK information and an LLM, is estimated at €5–€10. These amounts are treated as project estimates, not confirmed supplier prices. [P2]

For the calculations, I assume that €5–€10 is the total service cost per full assessment and already includes the basic search. This needs to be confirmed. If the €0.40 is charged separately, it must be added. The estimate does not provide a confirmed breakdown for hosting, maintenance or employee time, so these costs are not included in the totals below.

4.2 What does this mean for Bierens

If every one of the approximately 20,000 yearly requests receive one n8n assessment, the estimated service cost is €100,000–€200,000 per year. This is approximately €8,333–€16,667 per month. It assumes one paid assessment per request does not account for repeated companies, cached results or additional paid retries.

A possible approach is to check basic requirements first and order the more expensive data only when needed. If all 20,000 requests receive a €0.40 basic search and 25% then receive a full assessment, the estimated total is €31,000–€56,000. This assumes the €0.40 search is included in the full assessment: 15,000 basic-only searches cost €6,000 and 5,000 full assessments cost €25,000–€50,000. This is a planning example, not a tested saving.

5.1 Why n8n is expected to give stronger evidence

The expected advantage of n8n is that the workflow can retrieve specific fields from approved sources instead of asking AI to estimate them from web pages. For example, a company status can be taken from the selected company record. The workflow can then apply the same approved rule each time.

However, the source and the rule are two different things. The source may be outdated, the wrong company may be selected or the rule may contain an error. Credit scores are assessments from a provider; they are not a guarantee that a company will pay. The AI summary can also leave out important information even when the company data is correct.

5.2 Why the custom GPT needs review

The custom GPT can combine public information into a useful company profile. This is helpful when Sales needs an overview quickly. However, it must be clear whether a statement comes from a source, is an interpretation or cannot be confirmed. A link alone does not prove that the page supports the statement.

For both solutions, important facts should have a source and a date. Missing data should remain missing. The system should not turn an unknown employee count into zero or give a low classification only because a search failed.

5.3 Risks of Using the Custom GPT

Although the custom GPT showed promising results during testing, there are several risks that Bierens should consider before using it as an important part of the Sales assessment process.

1Incorrect or missing company information

GPT mainly relies on publicly available information. This information can be outdated, incomplete or unavailable. During testing with 20 Dutch companies, the GPT produced four classifications that did not match the expected results. Some errors occurred because inactive companies were not correctly identified or important financial information was missing.

2AI hallucinations and incorrect classifications

An LLM can generate information that sounds correct but is not supported by reliable sources. If the GPT uses incorrect information to classify a company, it may recommend the wrong DAM, AM, PAM or KAM category. Therefore, important company information should be verified before the classification is used.

3Privacy and data security

Incoming Sales requests may contain personal or confidential information. Sending this information to ChatGPT without approval could create privacy and security risks. Bierens must check where the information is processed, how it is stored and which data may be shared with external AI services.

4Limited automation and integration

The current custom GPT requires an employee to start the research and review the results. It is not directly connected to the Bierens website, internal Company Search system or Sales dashboard. This limits its suitability as a fully automated solution.

5Inconsistent results

The GPT may produce different classifications when the same company is assessed multiple times. This could happen because search results, available sources or AI interpretations change. Repeated testing is needed to measure how consistent the classifications are.

6Dependence on external services

The custom GPT depends on ChatGPT and the availability of external information sources. Changes to the platform, its features or its pricing could affect how the solution works.

There is also a geopolitical risk. As a European law firm, Bierens should be aware of the growing European interest in reducing dependence on American technology providers.

The WIRED article “All the Ways Europe Is Ditching American Technology” describes how European organizations are investigating alternatives to American software. It discusses concerns following US sanctions against officials linked to the International Criminal Court (ICC), which raised questions about how political decisions can affect access to important technology services. [9]

Although there is no evidence that Bierens will lose access to OpenAI, this example shows why relying entirely on one foreign technology provider could create a long-term business risk.

Since Bierens has restricted the use of Chinese AI systems, it would also make sense to investigate European AI alternatives rather than becoming fully dependent on American providers.

One possible alternative is Mixtral, a model developed by the French AI company Mistral AI. Depending on the model, licensing conditions and available infrastructure, it may be possible to run a suitable Mistral model within an approved European environment. This could give Bierens greater control over where information is processed.

However, using a European model does not automatically guarantee better privacy, security or classification accuracy. The model would still need to be tested, and its hosting, licensing and data-processing arrangements would need approval.

Risk reduction

These risks can be reduced by requiring reliable sources for important company information, marking missing information as unknown and sending uncertain classifications to Sales for review.

For long-term technical independence, Bierens should also avoid placing all its business logic inside a single AI platform.

An advantage of the custom GPT developed during this project is that its instructions, Account Segmentation Brief and reference documents, can be saved separately from ChatGPT. These materials contain much of the knowledge and classification guidance used by the agent.

If Bierens decides to switch to another AI provider, these materials could be reused to build a similar solution. Migration would still require implementation changes, prompt adjustments and new classification testing.

This is another reason why n8n could be a suitable long-term solution. By keeping company data connections, business rules and workflow steps separate from the LLM, Bierens could potentially replace OpenAI with a European model such as Mixtral without rebuilding the entire assessment process.

For these reasons, I recommend that Bierens considers European AI alternatives and designs the Sales assessment system to remain independent of any single AI provider. This would reduce long-term dependency risks while allowing Bierens to choose the most suitable AI technology in the future.

5.4 Platform change to consider

OpenAI has announced a planned move from custom GPTs to plugins, including a planned retirement date of 11 December 2026 for affected Enterprise workspaces. Bierens should check its own account notice before continuing with the conversational solution. If the current custom GPT needs to be migrated, the replacement should be tested again. [1]

This is another example of the risks of depending on a single AI platform. Changes made by the provider can affect how Bierens uses the solution, even when the agent itself is working correctly.

However, an advantage of the custom GPT developed during this project is that its instructions, Account Segmentation Brief and reference documents can be saved separately from ChatGPT. These materials could be reused to build a similar agent using another AI model, including a European alternative such as Mixtral.

The migration would not be completely automatic. The prompts may need adjustments, integrations may need to be rebuilt, and the classification results will need to be tested again.

Using n8n could make this process easier because the workflow, business rules and company-data connections can remain separate from the AI model. This would allow Bierens to investigate different AI providers without rebuilding the entire Sales assessment process.

The following examples show how the solution should handle company information to reduce reliability risks.

6.1 Privacy and security

The project plan states that sensitive information should not be sent to an external AI service without approval from Bierens. It also gives preference to storing and processing data within the EU or EEA. These requirements apply to both solutions. [P1]

The custom GPT runs within ChatGPT, so Bierens does not directly control its hosting. For a personal ChatGPT account, information is stored and processed outside the EU/EEA, including in the United States. This means that the current custom GPT cannot be assumed to meet the preference for European hosting and processing. Its setup must first be checked and approved before sensitive case information is used.

The available privacy controls depend on the ChatGPT plan. ChatGPT Business offers regional storage options to some customers, but this does not guarantee that processing stays in that region. Safety and abuse-monitoring copies are still stored in the United States. Eligible Enterprise and Edu workspaces offer European storage and model-processing options, although these also have limitations. Therefore, the actual workspace settings must be checked.

n8n provides more control over how the workflow handles information. It can be hosted in an environment approved by Bierens, and the workflow can limit which information is sent to an LLM. For example, company information can stay within the approved environment while only the necessary text is sent for summarization.

However, hosting n8n in Europe does not automatically mean that all information stays in Europe. The LLM and other connected services must also meet the requirements. The server must be maintained, access must be restricted, credentials must be protected and stored workflow data must be deleted when it is no longer needed. [5, 8]

Based on this comparison, n8n offers more flexibility to build the solution around Bierens’ privacy and security requirements. The current custom GPT has not yet been shown to meet those requirements. Both solutions need approval of their actual setup before processing sensitive information.

6.2 Frontend and system integration

The custom GPT can be used inside ChatGPT, but it is not directly an embedded part of the Bierens ecosystem. A separate assistant built through an API would be a different implementation. [1] n8n can connect the website input, company data requests and the result shown in the Sales dashboard. These connections still need to be built and tested. [7]

The dashboard should show the company, classification, sources, missing information and the reason for the result. An employee should be able to correct or review it. Failed connections and duplicate requests must also be handled so that a retry does not create a second case or send the same message twice. n8n supports error workflows, but their behavior must be configured. [6]

7.1 Completed Custom GPT Testing – Dutch Companies

After presenting the custom GPT to Laurens and receiving promising feedback, we made it available to the Sales team for testing. The purpose was to investigate how accurately the GPT could classify companies using publicly available information and the Account Segmentation Brief.

The Sales team tested the GPT on 20 Dutch companies and compared its DAM, AM, PAM and KAM classifications with the expected classifications.

Test results:

GPT achieved 80% agreement with the expected Sales classifications on this test set.

Problems identified:

The four incorrect classifications showed limitations in the information available to the GPT. Some companies were inactive, but the GPT could not reliably identify their status through public-web research. In another case, important financial information was unavailable, which affected the classification.

This showed that the GPT could produce useful company profiles and classification recommendations, but it could not always retrieve the verified information needed to apply the business criteria correctly.

Findings

The results suggest that the custom GPT is a useful starting point for company research, but relying only on publicly available information creates risks.

Based on these findings, I started developing an n8n workflow that combines AI research with verified company information from the KVK API. A credit-information API is also planned but has not yet been implemented.

The goal is to investigate whether verified information can improve classification reliability while limiting the number of paid API requests.

Testing on international companies is still ongoing. The 80% result applies only to the 20 Dutch companies tested and cannot be used as a general accuracy percentage for all incoming requests.

7.2 Advised comparison Test – Custom GPT vs. n8n

The next step is to compare both solutions using the same set of 30 Dutch company examples. This will help determine whether the n8n workflow produces more reliable classifications than the custom GPT and whether the additional API costs are justified.

The planned test set will contain:

20 normal company examples with sufficient information.

5 examples with missing or conflicting information, such as unavailable financial data.

5 difficult examples, such as inactive companies or companies with similar names.

Each example will be tested twice in both solutions to check consistency. This means 60 runs per solution and 120 runs in total. Repeated runs will not count as additional independent companies.

The expected classifications should be confirmed with Sales before the test. Examples already used to improve the GPT should be kept separate from the final comparison where possible.

Evaluation criteria

Expected outcome

I expect the n8n solution to provide more reliable results for company information that can be verified through the KVK API. However, this improvement still needs to be measured through testing.

The n8n solution may also be more expensive because it uses paid company-data services. Therefore, I want to investigate whether the workflow can first use less expensive research methods and only request additional API information when important data is missing or uncertain.

The final comparison will help determine which approach offers the best balance between accuracy, cost, automation and reliability for Bierens.

The results of this planned experiment will be added after testing is completed.

7.3 Recommendation

If the testing results are as expected and the n8n solution demonstrates better reliability and classification accuracy than the custom GPT, I would recommend continuing with n8n as the main workflow for the Sales Assessment Tool.

The n8n approach would allow Bierens to combine verified company information, business rules and AI tasks in one automated process. The custom GPT could still be useful as a cheaper alternative for initial company research, but it should not be treated as a verified company database.

Because the estimated cost of a full n8n assessment is €5–€10, it is important to investigate ways to reduce unnecessary API requests. One possible approach is to perform basic company research first and only use paid KVK or credit-information APIs when important information is missing, uncertain or required for the classification.

Before making a final recommendation, both solutions need to be tested on the same company examples. The results should be compared based on classification accuracy, information reliability, processing time and costs. The cost estimates should also be confirmed with Laurens.

If n8n performs better and the additional costs are justified by the improvement in reliability and automation, it would be the preferred solution. However, if the custom GPT provides sufficiently reliable results for certain types of cases, a hybrid approach could be considered to balance accuracy and costs.

The final decision will depend on the testing results and feedback from the Sales team.

7.4 International cases

Bierens handles requests from companies in different countries. International cases can be harder to assess because the same company information is not available in every country. For Dutch companies, the solution can use KVK information and available credit information. For foreign companies, it depends on the local company register and the coverage of the data provider.

Some countries have public company registers where information can be checked relatively easily. In other countries, access is more limited. Information may not be available online, may require payment or may only be provided in the local language. The available records may also be incomplete or outdated. This makes it harder to confirm information such as company status, employee count and financial position.

7.5 Impact on the two solutions

The n8n solution depends on the coverage of its connected data sources. If a provider does not cover a particular country or company, the workflow may not be able to retrieve a credit score or verified company information. Additional sources or manual research would then be needed.

The custom GPT can help search public sources and interpret information in different languages. However, it cannot reliably fill gaps when the information is not available. Limited online information also makes it harder for an employee to check whether the GPT’s findings are correct.

7.6 Proposed approach

The workflow should first identify the company’s country and check which approved sources are available. Where possible, the company should be verified through the official local register or an approved international company-data provider.

If important information cannot be found, the result should clearly show what is missing and which sources were checked. Missing information should not automatically lead to a negative classification. Limited data availability does not itself mean that a company is unreliable.

International cases with insufficient evidence should therefore be forwarded to Sales for further review. Testing should include companies from countries with different levels of data availability, so that the research shows where each solution works and where additional support is needed.

8.1 Project sources

[P1] Arbi, R. (2026). Projectplan Template – English v3. Fontys University of Applied Sciences / Bierens Groep. Internal project document. Used for the project context, annual request volume, research questions and security requirements.

[P2] Van der Lof, L. (16 September 2026). Interview about technical alternatives and API costs. Internal discussion at Bierens Groep. Discussed the use of fixed business rules, AI-based company research and the potential costs of company-data APIs. The project uses approximate cost estimates of €0.40 for a basic AI search for the custom and €5–€10 for the APIs in the n8n solution.

[P3] Bierens Sales Team. (2026). Test Sheet NL – Custom GPT Classification Results. Internal Excel document. Contains the results of testing 20 Dutch companies, with 16 matching classifications and four mismatches.

[P4] Arbi, R. (October 2026). n8n Sales Assessment Workflow Prototype. Internal prototype. Used to investigate company research, KVK verification, AI processing, account segmentation and structured reporting.

8.2 Public sources

The following official sources were consulted on 16 October 2026. They support the platform comparison. Prices and features can change.

OpenAI. (2026). GPTs in ChatGPT. https://help.openai.com/en/articles/8554407-gpts-in-chatgpt

OpenAI. (2026). Enterprise Privacy. https://openai.com/enterprise-privacy/

OpenAI. (2026). API Data Controls. https://developers.openai.com/api/docs/guides/your-data

OpenAI. (2026). ChatGPT Business Overview. https://help.openai.com/en/articles/8792828-chatgpt-business-overview

n8n. (2026). Privacy and Security. https://docs.n8n.io/privacy-and-security/

n8n. (2026). Error Handling. https://docs.n8n.io/flow-logic/error-handling/

n8n. (2026). Webhook Node Documentation. https://docs.n8n.io/integrations/builtin/core-nodes/n8n-nodes-base.webhook/

n8n. (2026). Security. https://n8n.io/legal/security/

Burgess, M. (8 June 2026). All the Ways Europe Is Ditching American Technology. WIRED.

https://www.wired.com/story/all-the-ways-europe-is-ditching-american-technology/