Engineering Verified Real deployment at a named or precisely described company, with a measured result from a primary source.

Engineers ask questions of their standards, drawings and past project files and get an answer with the source page (Flender, NETZSCH, Victaulic)

Added 2026-10-27 · Last updated 2026-10-27

Manufacturers in this case: Flender · NETZSCH Pumpen & Systeme · Victaulic · all manufacturers using AI

Role

Engineering (design, application and project engineering: finding answers in standards, specifications, drawings and past projects) · Typical users: design or application engineer, project engineer or project manager, product manager, technical customer support, new engineers in onboarding

The problem

Engineering knowledge is spread over thousands of documents: internal design standards, customer specifications, drawings, test reports, manuals, approvals and past project files, in shared drives, PLM and email. Keyword search finds files, not answers, so engineers spend a long time reading, ask the senior colleague who knows, or redo work that was already done. At Flender, a maker of industrial gearboxes and drives that builds to order, knowledge sat in fragmented repositories and in many formats (documents, drawings, spreadsheets), and interpreting a design standard or drawing took about 30 minutes. At Victaulic, a pipe-joining and fire-protection manufacturer with 5,500 employees and 50 facilities, customer care and application engineering teams searched by hand through manuals, submittals, white papers and approvals to answer technical questions, which slowed answers and made them inconsistent. At NETZSCH Pumpen & Systeme, project, product and development staff needed fast access to project-specific details buried in extensive documentation.

How AI helps

The company's own documents are indexed into a private assistant. An engineer asks a question in plain language (for example: what flow rate and differential pressure does this customer require, which noise standard applies, which approvals must we pass on to the supplier), and the assistant finds the relevant passages and writes a short answer with links to the source pages, so the engineer can check them. This is called retrieval-augmented generation (RAG): the language model answers from retrieved company documents rather than from general knowledge. At Flender it runs on premises and also reads drawings; at Victaulic it runs on AWS with Claude Sonnet; at NETZSCH it is the MAIA platform. The engineer still reads the source and decides.

Step by step

List the 20 technical questions your engineers ask most often and where the answer usually is (standard, spec, drawing, past project, a senior colleague).
Pick one document collection for the pilot, such as your internal design standards or the last 20 customer specifications, and remove outdated versions first; the assistant repeats what the documents say.
Check confidentiality: customer drawings and specifications often fall under NDAs and export rules, so use a tool that keeps data in your tenant or on premises, as Flender did.
Test the assistant on your 20 questions and score each answer: correct, partly correct, wrong, with or without the right source page.
Roll it out to a small team, require that answers show the source page, and collect wrong answers to fix the documents or the setup.
Measure search time on a few typical tasks before and after, and extend to other collections (test reports, past projects) once answers are reliable.

Tools

Atos sovereign agentic AI engineering assistant (RAG and multimodal AI, on premises): used at Flender China
MAIA (AI knowledge platform for technical documents): used at NETZSCH Pumpen & Systeme
Nagarro NIA agentic AI platform on AWS (Amazon Bedrock with Claude Sonnet, Amazon Textract, Amazon OpenSearch): used at Victaulic
Other tools to evaluate: Microsoft 365 Copilot or Copilot Studio over SharePoint, Google Gemini for Workspace, the AI search of your PLM, and enterprise search tools with source citations
For a first test: any LLM with file upload (ChatGPT, Claude, Copilot, Gemini) on two or three non-confidential documents, with the prompt below

In the AI tools index: Atos sovereign agentic AI · MAIA · Nagarro NIA · all AI tools used by manufacturers

Result

Flender China, sovereign agentic AI engineering assistant built by Atos and deployed on premises: time to interpret design standard documents and engineering drawings cut by 83%, from about 30 minutes to 5 minutes; about 2,000 engineering hours saved a year; knowledge discovery and document retrieval 30 to 50% faster; Atos also reports a sixfold increase in engineering design processing capacity (Atos case study, September 2026). NETZSCH Pumpen & Systeme (positive displacement pumps, about 600 employees in Germany), MAIA: more than 5,600 working hours saved since introduction, 235,500 pages of technical documentation searchable, about 1,400 queries a month; used to review customer specifications, write test procedures from standards and specifications, and pick out the specifications to pass on to suppliers; its head of project management says the team identifies critical details earlier in projects (MAIA case study). Victaulic, conversational assistant built by Nagarro on AWS with Amazon Bedrock and Claude Sonnet: searches dropped from minutes to seconds, 600 to 700 engineering hours saved a year, and response accuracy above 90% (Nagarro success story).

As reported by the source; most figures are vendor or self-reported and not independently verified.

Human review: cautions

The assistant can misread or mix up documents; every answer must show its source page, and the engineer checks it before using it in a design, a quote or a customer reply.
Outdated or contradictory documents give outdated or contradictory answers: clean up versions before indexing them.
Customer drawings and specifications are often under NDA or export control: keep them in a company-approved tool, not a public chatbot.
All figures are published by the vendors or integrators (Atos, MAIA, Nagarro); the hours saved are their estimates and the measurement method is not described.

Try it yourself

You are an experienced application engineer at a manufacturer.

I will attach or paste: (1) one to three technical documents [for example a customer specification, our internal design standard, a product manual or a past project report]; (2) a list of questions our engineers ask about them.

Do this:
1. Answer each question in two to four sentences, using only the attached documents.
2. After each answer, quote the exact sentence or table you used and give the document name and page or section.
3. If the documents do not answer the question, or two documents disagree, say so clearly and list what is missing or conflicting, instead of guessing.
4. At the end, list requirements in the documents that look critical or unusual (tight tolerances, special approvals, standards, test conditions) that an engineer should not miss.

An engineer will check every answer against the source before using it. Do not upload confidential customer documents to a public chatbot; use an approved company tool. Use only the data I paste in; if something is missing, write 'not found' or list it as a question instead of guessing. Do not send, order, post or change anything in any system: produce drafts for my review and approval.

Questions:
[paste]

Want more like this? Join the newsletter

Related cases

All 11 Engineering use cases →