Less manual work. More time for your core business.

01 / 06

The challenge

Each project comes with four extensive reports totaling several hundred pages. Staff have to pull the relevant information together and assess it against defined criteria. Synthesizing the data by hand takes several working days.

02 / 06Prepare

Step 1 – Prepare the information

Several AI-driven extraction processes handle the different document types and prepare the relevant content in a structured form. The information is processed chapter by chapter and topic by topic rather than analyzing all documents at once.

03 / 06Review questions

Step 2 – Structure the review questions

Together with the banking team, domain-specific review questions are defined for each report chapter. A central JSON structure maps every question to the information and assessment criteria it requires.

04 / 06Answer

Step 3 – Answer the review questions

A Python function on AWS Lambda orchestrates the analysis of each review question. Relevant information from the various documents is combined and passed to an LLM to answer the question precisely. Where needed, results from earlier chapters are taken into account.

05 / 06Assemble

Step 4 – Assemble the results

The individual answers are automatically assembled into a report with 18 chapters. A final AI step improves structure, readability and linguistic consistency. The expert review remains with the bank’s staff.

06 / 06Result

From several days to around 15 minutes

Automated data synthesis and report generation take around 15 minutes instead of roughly two to three working days. The result is a structured assessment report, which the responsible staff then review on its merits.

01 / 06

The challenge

Customer inquiries contain multiple technical line items, and each one has to be matched with suitable products from a catalog of thousands of items. Researching this by hand takes a lot of time.

02 / 06Prepare

Step 1 – Prepare the files

A Python service hosted on AWS converts incoming P93 files into structured JSON and prepares them for further processing.

03 / 06Understand

Step 2 – Understand the requirements

An LLM analyzes the content, excludes irrelevant line items and maps the technical requirements to a standardized schema that matches the database fields of the product catalog.

04 / 06Match

Step 3 – Match products

A Python matching service compares the standardized requirements with the product catalog. It applies hard exclusion criteria and flexible weightings, both defined in close collaboration with the client.

05 / 06Suggestions

Step 4 – Show suggestions

For each line item, the system suggests three suitable products, including key properties, a confidence score and a link to the product page. Users can also adjust the weightings and apply predefined filters.

06 / 06Result

The result

An inquiry with around ten line items is processed in about ten minutes instead of roughly two hours of manual research. The solution is in production use.

01 / 07

The challenge

An internal help desk answers numerous questions from employees every day. The knowledge it needs, however, is spread across PDFs, Excel files, documents and internal websites. Handling requests manually is time-consuming and ties up a large team.

02 / 07Prepare

Step 1 – Prepare the knowledge

A Python process ingests documents, websites and spreadsheets from various internal sources, prepares each according to its source and stores them in a vector database as a central knowledge base.

03 / 07Retrieve

Step 2 – Find relevant content

A user’s question triggers a RAG process that retrieves matching knowledge chunks from the vector database and prioritizes them through re-ranking. A second RAG process also searches an internal Excel knowledge source to provide relevant links for further reading.

04 / 07Answer

Step 3 – Generate answers

Based on the most relevant content, an LLM writes a precise answer. Employees get clear answers within seconds, along with direct links to further information.

05 / 07Routing

Step 4 – Route special cases

An upstream router detects questions the knowledge assistant should not answer. Irrelevant requests are filtered out, sensitive topics are forwarded to the responsible contacts and, in defined cases, a human is brought into the loop.

06 / 07Evaluate

Step 5 – Evaluate quality

Together with the subject-matter team, reference questions with ideal answers are defined. Several large language models are compared against these test cases to select the model with the most reliable answer quality.

07 / 07Result

The result

The knowledge assistant answers most internal requests within seconds. This takes considerable pressure off the help desk: instead of a team of seven, a single person now takes calls, and the rest of the team can focus on other tasks.

01 / 07

The challenge

Real estate appraisals require extensive location research and the evaluation of a large amount of market information. Data has to be gathered from different sources, compared and turned into clear analyses.

02 / 07Capture

Step 1 – Enter the property

The appraiser enters the property address and basic details such as year of construction and condition. The application then automatically starts several analysis processes tailored to that property.

03 / 07Location

Step 2 – Analyze the location

Several processes running in parallel examine the immediate surroundings: shopping, schools, kindergartens, universities, public transport and highway access. The results are automatically turned into location descriptions.

04 / 07Market data

Step 3 – Gather market data

The appraiser can upload their own market reports, select documents from a curated database of publicly available reports or run an online search. The application takes into account which reports fit the property and its location.

05 / 07Compare

Step 4 – Compare market figures

Python-based processes extract relevant figures such as yields, current prices per square meter and how they have developed over time. Values from different market reports are set side by side in a clear comparison table.

06 / 07Appraisal

Step 5 – Prepare the appraisal

Selected market figures are turned into a real estate market analysis. Location descriptions can be output as bullet points or as shorter or more detailed texts. All results can be reviewed, edited and exported in various formats directly in the web application.

07 / 07Result

The result

Automating the preparation of location and market information cuts the effort for key parts of the appraisal from an estimated two hours to about 15 minutes. The appraiser stays in control of the results and can adjust them as needed.