Document analysis and draft responses
We build AI features for work that involves reading across documents, checking details and copying information between systems.
For a customer enquiry, for example, the application could bring together relevant product documentation and contract details to draft a response. The person handling the enquiry can check the sources, fill in any gaps and approve the reply. The enquiry, supporting documents and draft stay together in the same case.
We test whether this saves work by comparing it with the way your team handles real cases today. Are the details correct? Is anything missing? How much editing does the draft need? The feature has to earn its place in everyday use through the quality of its output and the time it saves.
AI features we can build
These are examples of possible features, either within an existing application or as part of a new one.
Answers with links to your source documents
Staff can search manuals, product documentation or internal policies. The application finds relevant passages and uses them to form an answer with links to the originals. Access permissions still apply, and the interface makes clear when the available sources do not support an answer.
Assistants for customer and partner portals
An assistant can explain documents, help with an enquiry or retrieve the status of a case. Current information comes from the system responsible for it. Actions such as updating customer details require the appropriate permissions and approvals.
Document extraction and classification
Required fields are extracted from PDFs, emails or scanned documents, classified and assigned to a case. The application checks mandatory fields and formats. Uncertain values are held for review rather than passed unchecked to the business system.
Editorial assistance inside the CMS
Drafts, summaries, suggested translations and tags can be created alongside the content. Templates and approved sources set the boundaries; editors review and publish through their usual workflows.
Prepare the next step in a workflow
An AI agent can classify a request, identify missing information, bring together data from several systems and prepare the next step. We define which tools it may use. Orders, publication and other consequential actions can require explicit confirmation.
Finding products and content
Search can respond to a description of a task or related terms, not just an exact keyword. We combine it with filters and structured data for applications such as product catalogues, specialist portals and media libraries.
The technology behind it
For knowledge search, retrieval finds relevant passages and the language model uses them to draft an answer. APIs supply current data from your business systems.
Language models and model APIs
Language models can write answers, summarise text or produce structured suggestions. We connect them through APIs or assess a model that can run locally. Selection depends on language quality, data handling, response time and cost per task, among other factors.
RAG: answers grounded in your documents
Retrieval-Augmented Generation combines search with a language model. Relevant passages are retrieved from approved documents and supplied to the model as the basis for its answer. The documents do not need to be trained into the model. Updates, citations and permissions are part of the implementation.
Embeddings and vector search
Embeddings represent the meaning of text as numerical vectors. Vector search can then find related content even when different words are used. Where precise names, part numbers or filters matter, we can combine it with conventional full-text search. A vector database does not replace the system that owns the business data.
MCP and existing APIs
The Model Context Protocol gives AI applications access to defined data sources and tools. We can connect existing APIs or provide purpose-built MCP tools. Authentication, permissions and approvals are enforced by the application, not left to instructions in a prompt.
WebMCP for browser-based features
WebMCP exposes selected website functions in a structured form to supporting browser agents. We are already testing it on our own website. Browser support is still developing, so the site must remain fully usable without WebMCP.
Local testing on NVIDIA DGX Spark
We use our own NVIDIA DGX Spark to test locally runnable models and AI prototypes. It helps us assess answers, runtimes and integrations before production use. The production environment is a separate decision made with your IT team: on premises, in a suitable cloud or a combination of both.
From prototype to production
For the first prototype, we choose one task and an approved dataset. Your team evaluates the results against real questions and cases: is the answer correct, are the sources relevant, and does it save work? We also test missing information, conflicting documents and requests that should be refused.
Before launch, we test data access and actions across different roles. Content from documents or external systems must not override permissions. We limit tool calls and costs, handle failures and preserve a way to work without AI. Sensitive data, logging and retention periods are agreed with the responsible teams.
Operations include versioned configurations, automated checks, CI/CD, monitoring and maintenance. Changes to the model, search or data sources trigger a fresh evaluation. Code, configuration and operating documentation remain accessible. You contract directly with hosting and platform providers.
Our approach to AI
Which task is worth tackling? What data may be used? How much automation is appropriate? Applied AI sets out how we work through those questions with you and the responsibilities we take on.
Connecting AI to your existing systems
- Semantic search in TYPO3Search relevance, content and permissions
- Web applicationsCustom interfaces, data models and workflows
- Intranets & extranetsProtected content, roles and partner access
- Legacy modernisationConnecting or gradually replacing older applications
- Maintenance and developmentUpdates, monitoring and ongoing technical support
- AccessibilityNavigation, forms and accessible web applications