AI-driven tender process for cost and time optimization

For companies from the financial industry, participation in tenders is a way of acquiring new customers, but unfortunately, it always requires an involvement of many specialists in preparation of an offer, what significantly increases operating costs.

Not every tender ends with success, which, on an annual basis, can result in large investment costs that do not yield any returns – notes Tomasz Gdula, Software Engineer AI/ML, who uses AI to streamline and automate the process of creating responses to requests for proposals (RFPs).

In this interview, Tomasz shares his experience of implementing a solution based on the language models. He also explains how preparation of the tender documents has been improved thanks to the tool created, resulting in reduction of the working time of the people involved while still maintaining high quality of answers provided in the forms.

When bidding for obligatory reporting systems, the sales team receives a list of detailed questions from the potential client concerning a variety of areas – from security and compliance issues, to the architecture, where the data is stored and how it is processed, to the after-sales service provided by the support crew. What is the process for answering these customer questions?

The questionnaires are answered by our experts from the relevant departments. Some of the issues are standard for entities from the financial market. However, more often than not, questions are specific, very detailed and concrete, what requires going a little deeper, by e.g. verifying the validity of the information. Sometimes it is even necessary to combine the knowledge of several people with various specialisations.

Depending on the number of questions (and there are often dozens of them) as well as their complexity, it often takes more than a week to collect all the information. This process engages many people from various fields, including translators more proficient in the nuances of English used in financial industry, banking, or law . It is worth pointing out that not every tender is successful, which, on an annual basis, can involve large investment costs with no return.

Is that why you've decided to reach out to modern AI technology?

In order to assist our specialists in the tendering processes, we decided to create a tool based on language models. The application uses a variety of data sources, based on documentation stored in various locations and formats. On the basis of the content contained therein, it answers the questions concerning the forms as briefly and precisely as possible.

The expert's role is limited to verifying the correctness, making minor corrections or completing missing answers if a given information is missing from the available sources.

Why is it important for experts to fill in missing information in the knowledge base?

The app collects information from our internal resources. These can be documents describing processes or technology, questionnaires already completed for other clients in the past, spreadsheets uploaded on SharePoint, etc. However, it may turn out that the model does not find an explanation for a question. The expert must then provide the answer individually and complete the knowledge base.

We see how important it is to document knowledge when the expert concerned is absent and the team has to complete and return the questionnaire within a strict deadline. Hence, it is no less important that each time an employee completes the internal documentation with the missing information. This allows the model to learn and give better answers in the future.

Since the model is learning all the time, will the app ever be finished?

There may always be customer enquiries to which even the experts will not know the answer. In such a situation, the team will have to determine our position concerning the issue, and the tool will only then make use of the newly created knowledge. In this sense, the tool is constantly being developed.

What are the differences between ChatGPT and the tool developed by you?

Our tool offers a diverse methods of communication. Of course, it is still possible to use chat, but we also enable users to comfortably work directly on Excel files. This allows for easier interaction with the tool and effective answers to many questions.

We cannot and do not want to make our knowledge public, so we use Microsoft Azure solutions and the latest OpenAI models implemented in terms of our private subscription. This allows us to guarantee the highest level of data security and confidentiality.

Furthermore, it is our priority that the tool provides precise answers based on our internal sources of knowledge. To this end, we have applied carefully tailored commands and further process the queries to obtain the most accurate results. Furthermore, our tool, unlike ChatGPT, has mechanisms to prevent the generation of erroneous results, or so-called hallucinations, which include dedicated safeguards and a systematic evaluation of the responses.

And finally, the solution is only used by selected company employees who are faced with completing questionnaires in their daily work. Some information is confidential and not all employees, even at our company, should have access to it.

Can you reveal how much it costs to use this app?

The cost of using the app depends on the amount of data we feed the model with, the number of queries, and the capabilities of the chosen model. As part of the subscription, we pay for tokens, which are then converted into money. For example, one thousand tokens can be converted into approximately 750 words. Currently, the costs range from a few dozen to a hundred dollars. Thus, the monthly costs of the tool are incomparably lower than the man-hours of the specialists who would be expected to answer the questions themselves.

Will this tool be available to your customers?

Due to the confidentiality of the information on which the app is based, this tool is only available to our selected employees. On the other hand, we are already discussing the issue with customers who need a similar solution in their company and, importantly, have knowledge that is largely documented. This is essential to ensure the correctness and accuracy of the generated responses.

Where else in banking do you see the use of tools based on language models?

Tools based on the language models and meant for the banking sector can support processes in many areas. In customer service, they can e.g. automatically analyse a person's communication history, suggesting proven solutions to specific problems, tailored to their needs or specific situation.

In the process of analysing and reviewing contracts, they can search for inconsistencies in the document or non-compliance with legal requirements, including the identification of abusive clauses.

Finally, given the ever-changing regulations, such tools may monitor legal changes and support adaptation.

What do you think the future will look like in terms of the application of tools that use Machine Learning?

The future looks extremely promising, with new models including new capabilities emerging almost every week. Furthermore, the manner in which we communicate with AI systems continues to evolve, regularly changing our approach to this technology.

Currently, we mainly use language models in our work, but we also use spoken language or images in our daily communication – to express complex ideas better.

I believe we will be using multiple assistants at every step in the future. Such as the tool we have been discussing, or others like it – such as the GitHub Copilot. These are tools that complement each other and help us in our daily work.

Thanks to this, people will be able to spend more time creating, testing, and developing new ideas!

Thank you for the interview.

Tomasz Gdula, a software engineer with many years of experience, specialises in developing solutions based on language models. He is currently responsible for creating and developing a tool at FINGO (part of the Regnology Group) that automates completion of complex tender questionnaires required during tenders by banks and other financial institutions.

‍