Capturing expert knowledge before staff retire: interviews, transcripts and an assistant

Recording what experienced staff know, with AI on your own server

Capturing the knowledge of staff who retire soon works best as a method with five steps: structured interviews at the machine, transcription on the company's own server, how-to notes drafted by a language model, a check by the expert, and an assistant that answers successors only from approved notes. The AI does the typing and the sorting; the expert decides what is correct. All of it runs on one server with two NVIDIA RTX PRO 6000 Blackwell Server Edition cards, so recordings of your people and your processes stay in the building.

We wrote this for the owner or managing director of a manufacturer whose senior setters, process engineers and service technicians will leave in the next few years, and whose knowledge is in their heads more than in any document. The method comes first, then the law on recordings, then the server for the IT contact or the service provider who will run it.

Which tacit knowledge to capture first

Not every piece of experience is equally at risk. Three kinds of knowledge are usually missing from manuals and work instructions:

• Setting up and running in. How a setter brings a machine to a good part after a tool change, which values to adjust first, what the first parts look like when something is off.
• Fault patterns. Which noise, smell or pattern on the part points to which cause, and which repairs were tried before and did not help.
• Exceptions. The customer whose material behaves differently, the old machine that needs a different sequence, the reason a step in a work instruction is done the way it is.

Start with the people who leave first and with the processes that have no documentation at all. A list of names, retirement dates and the three machines or tasks each person knows best is a better project plan than any software choice. Two or three interviews with one expert show quickly how much the method yields in your company.

Knowledge transfer before retirement in five steps

1. Plan the interview. One topic per session: one machine, one product family or one fault type. Prepare questions about the normal sequence, the warning signs and the exceptions, and have a successor in the room who asks about anything unclear.
2. Record at the machine. A clip-on microphone for the expert, a camera or a phone for the hands and the display, photos of settings and gauges. Note the machine number and the drawing or part number at the start of each recording, so the note can be linked to them.
3. Transcribe on your own server. A speech recognition model turns the audio into text with timestamps; a diarisation model labels who spoke when, so the expert's words are separated from the questions.
4. Draft the note. A language model turns the transcript into a how-to note with fixed fields and marks contradictions and open questions. Each statement keeps a reference to the minute of the recording it came from.
5. Check, approve, publish. The expert and the successor read the note, correct it and approve it. Only approved notes go into the assistant, each with an owner and a date for review.

The steps that need a person, planning and checking, are where the time goes. Transcription takes less than half a minute of GPU time per hour of recording by the published figures below.

Interviews at the machine in practice

People explain better while they show. A setter standing at the machine remembers the step that gets skipped in a meeting room, and the camera catches the gesture that words miss. Keep sessions to one topic and stop when the expert tires. Ask for the last three times something went wrong and what was done about it; stories about real faults carry the knowledge that a description of the normal sequence leaves out.

Record in a quiet moment if possible. A machine running next to the microphone makes transcription harder, and a clip-on microphone close to the speaker helps the transcription. Photos of settings, gauges and the part are often worth more than a minute of explanation and go into the note as they are.

The language matters too. Many experts explain in dialect or switch between languages with colleagues. Whisper large-v3's model card notes that accuracy is uneven across languages and lower for low-resource languages, so test with two or three real recordings before you plan the whole programme.

Transcription on your own server

Speech recognition needs little GPU memory. OpenAI's Whisper repository gives about 10 GB for Whisper large and about 6 GB for the faster turbo version, with code and weights under the MIT licence; the large-v3 model card on Hugging Face lists 99 languages and is tagged Apache 2.0. NVIDIA's Parakeet TDT 0.6B v3 has 600 million parameters, transcribes 25 European languages with punctuation and word timestamps, and handles audio "up to 24 minutes long with full attention (on A100 80GB) or up to 3 hours with local attention" in one pass, under CC BY 4.0.

Hugging Face model cards openai/whisper-large-v3 and nvidia/parakeet-tdt-0.6b-v3, OpenAI's Whisper repository; RTFx on English test sets from the Open ASR Leaderboard paper (arXiv 2510.06961v3, 10 December 2025), one A100-SXM4-80GB, batch 64 where memory allowed; all read 25 October 2026.

RTFx is the number of seconds of audio transcribed per second of computing. An example programme of 15 experts with 12 hours of interviews each gives 180 hours of audio. At Whisper large-v3's 145.5 that is 180 × 3,600 ÷ 145.5 = 4,453 seconds, about 74 minutes of GPU time, and at Parakeet's 3,332.7 about 3 minutes, by our arithmetic on the paper's A100 and English test sets. Your card, your language and your recordings change these figures, but the order of magnitude is clear: transcription does not size the server.

Whisper's model card also states that "the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination)". That is one more reason why the expert reads the transcript of the interview before the note is approved. Our guides to the speech-to-text server for Whisper and Parakeet and to speaker diarisation on-premise cover the models in depth.

From transcript to a checked how-to note

A transcript is long and full of repetitions, half sentences and side remarks. The language model's job is to turn it into a note that a successor can follow, with the same fields every time.

Our example template; fields to be adapted with the expert and the quality department.

OpenAI publishes gpt-oss-120b under Apache 2.0 with 117B parameters, 5.1B of them active per token, and its safetensors files add up to about 65 GB, 60.8 GiB. By our users-per-card rule, which takes 90 per cent of the 95.6 GiB the driver reports on an RTX PRO 6000, less 3 GiB per card and less the weights, it leaves 22.2 GiB for the cache on one card, room for 39 conversations of 16,384 tokens at the same moment with a 16-bit cache. A long interview is processed in sections by topic, and drafting notes overnight is a small batch job for that card.

The model drafts; the expert approves. A note with a wrong torque, a skipped safety step or an invented cause is worse than no note, so every note carries the name of the person who approved it. The minute references let the successor listen to the original explanation when the note is not enough.

We build this server with two or four RTX PRO 6000 Server Edition cards and set up transcription, note drafting and the assistant through our Private AI/ML service; tell us how many experts and which machines you have in mind, and our team replies with a plan and a configuration.

An AI assistant for successors and onboarding new staff

Once notes are approved, the assistant answers questions from them: "what do I check first when the seal on line 3 leaks", "why does the second step on the old press differ". It searches the approved notes, answers with a reference to the note and the minute of the recording, and says so when no note covers the question. This is retrieval over your own documents, as our guide to RAG on company data explains; our guide to preparing company documents for an AI assistant covers the manuals and work instructions you already have, which can sit in the same index.

The same notes help with onboarding. A new setter reads the notes for the machines in the first weeks and asks the assistant what is unclear, while the expert, until retirement, reviews the questions nobody could answer. Questions without an answer are the list for the next interviews.

Recordings of employees and data protection

Audio and video of an employee at work are personal data, and a request from the employer is hard to refuse, which makes consent a weak basis. The European Data Protection Board's Guidelines 05/2020 on consent state that, because of the dependency between employer and employee, it is "unlikely that an employee would be able to respond freely" to a request for consent, and that employees can consent freely only in exceptional cases where "it will have no adverse consequences at all whether or not they give consent". Your data-protection officer chooses the legal basis, the information given to the people recorded and the retention period before the first interview.

Recital 155 of the GDPR adds that Member State law or collective agreements "may provide for specific rules on the processing of employees' personal data in the employment context". Rules on employee representation differ between Member States; our guide to the works council and an internal AI assistant is the place to start, and your legal counsel confirms what applies.

Three settings make the project easier to defend:

• Purpose. The recordings document processes. They are not used to evaluate the expert. Annex III point 4(b) of the AI Act lists as high-risk AI systems intended "to monitor and evaluate the performance and behaviour of persons" in work relationships, with those rules applying from 2 December 2027 under Article 113 as amended; keeping the purpose to documentation keeps the system away from that use.
• Retention. Keep raw recordings only as long as the note needs checking, and the approved note afterwards. Decide this with the data-protection officer.
• Access. Raw recordings are visible to the project team and the expert; the assistant searches approved notes only.

Whether a data protection impact assessment is needed is a question for the data-protection officer. This section is general information as of October 2026, not legal advice.

The server for transcription and the assistant

One server carries the whole method, because its parts run at different times: transcription and note drafting in batches at night, the assistant during the working day.

Reference configurations on our AI servers page; conversations by our users-per-card rule with gpt-oss-120b at 16,384 tokens and a 16-bit cache; power by our estimate (cards at 600 W, two AMD EPYC processors, plus 15 per cent for the rest of the server), to be confirmed by a measurement on the delivered system.

Two cards. The entry server has two RTX PRO 6000 Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives. MIG splits the second card into four isolated instances of 24 GB; Whisper large with its 10 GB fits one of them, and the embedding model for search and the diarisation model fit others. By our estimate the server draws about 1.8 kW and fits one 230 V 16 A feed, which carries 2.94 kW at 80 per cent.

Four cards. The knowledge assistant is rarely the only use. When purchasing, quality or sales want their own assistants on the same server, the inference node with four cards gives each a card, as our guide to AI servers for companies with 10 to 100 staff describes for the office side. Four cards need 32 A or three-phase.

The server needs a server room or a rack with cooling. Without one, our Private AI/ML service runs the platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU. Our guide to an AI server without an IT department covers who runs what when the company has no IT staff of its own.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build the server to order with two or four RTX PRO 6000 Server Edition cards, assembled and tested before delivery, delivered within the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up transcription, the note drafting and the assistant on it. For recorded meetings, our guide to meeting minutes with a local AI uses the same speech models.

FAQ