Start now, in three layers: record the engineer doing real work, turn their rules of thumb into written tables, and put an AI agent on top that answers from those documents with sources while a person approves. It is urgent because Portugal has one of the oldest workforces in the EU, with 99 workers over 54 for every 100 under 35, and 88% of Portuguese firms name the availability of skilled staff as an obstacle to investment.
If your most experienced engineer retires next year, start capturing what they know now, in three layers: record them doing real work, turn their rules of thumb into written tables, and put an AI agent on top that answers from those documents with sources while a person approves. Twelve months is enough if you begin this quarter, and the date is rarely a surprise: the statutory retirement age is 66 years and 9 months in 2026 and 66 years and 11 months in 2027 [14]. It is urgent because Portugal has one of the oldest workforces in the EU: 99 workers over 54 for every 100 under 35 [2]. And, as Bosch's chief digital officer put it this summer, industry is losing its most experienced generation "para a reforma", to retirement, not to artificial intelligence [1].
Key facts
- The number of employed people aged 55–64 in Portugal rose 65.7% in twenty years, from 588,400 in 2003 to 975,200 in 2023 [2].
- In Portuguese manufacturing, people aged 55–64 were 18.7% of employment in 2025 (155,900 people), against 12.3% in 2015 (our arithmetic on Eurostat data) [10].
- Portugal has 99 workers over 54 for every 100 under 35, the fourth-oldest workforce in the EU [2].
- 88% of Portuguese firms name the availability of skilled staff as a long-term obstacle to investment, against 79% in the EU; in manufacturing the figure is 86% (EIB survey, December 2025) [3] [11].
- AIMMAP, the metalworking association, says companies still name the shortage of qualified talent as the biggest obstacle to growth (August 2026) [4].
- Bosch's CDO Tanja Rückert: "Estamos a perdê-la para a reforma", we are losing the most experienced generation to retirement [1].
Why is losing one engineer such a risk now?
Because the replacement is not on the market. The workforce has aged fast: in twenty years the youngest workers (up to 24) fell by more than 40% while the 55–64 group grew by two thirds [2]. Portugal also had the second-largest share of people aged 65 and over in the EU in 2024 [3]. AEP's July survey of 400 managers lists labour scarcity among the main constraints on business [5], and AIMMAP puts the shortage of qualified talent first [4].
Manufacturing shows the same shift. Eurostat counts 155,900 people aged 55–64 working in Portuguese manufacturing in 2025, up from 94,500 in 2015: from 12.3% of the sector's workforce to 18.7% [10]. The EU manufacturing share is 20.1%, so Portuguese factories are not older than the European average; they are ageing slightly faster (our arithmetic). AIMMAP's 2026 plan answers with talent-attraction and immigrant-recruitment programmes [12]. Both take years.
Large employers are responding by building their own pipelines, which tells you how long the fix takes. In June, Continental signed a memorandum with the municipality of Famalicão and the FORAVE vocational school, which trains young people in automation, robotics, industrial maintenance and mechatronics, to improve the school's facilities [9]. That produces technicians in years. A 40-person machine shop whose senior estimator leaves in 2027 cannot wait for them.
What knowledge actually walks out of the door?
Rarely the things that are already in the ERP. What leaves is the reasoning around them:
- Quoting rules. Which machine a part family really runs on, how long setup takes, what surcharge a ±0.01 mm tolerance deserves, when to refuse a job.
- Machine behaviour. The noise that means a spindle bearing is going, the parameter that must be changed after a mould change.
- Supplier knowledge. Who can anodise in five days, who is reliable in August.
- Customer meaning. What one customer means by "standard finish", which drawing notes another customer never enforces.
- Failure history. Why the last three non-conformities happened and what was changed.
Simoldes, a mould and plastic-components group with 7,000 employees, was asked exactly this question in an interview this year: how to balance investment in automation with preserving specialised technical knowledge [7]. Its answer was to redesign roles "para que operadores se tornassem técnicos especializados", so operators became specialised technicians [7]. For a smaller plant, the practical version is to write the knowledge down before the person who carries it leaves.
How do you capture what an experienced engineer knows?
No single method works alone. The table shows the ones we use, with the expert's time as our own estimate for a typical plant.
| Method | What it captures | Expert's time (our estimate) | What you get |
|---|---|---|---|
| Recorded walkthroughs | How they quote a real RFQ or diagnose a real fault, step by step | 1–2 hours a week | Transcribed, searchable recordings linked to parts and machines |
| Structured quoting rules | Setup times, machine choice, surcharges, refusals | Four to eight sessions of 2 hours | A rules table juniors and agents can use |
| Decision notes on real cases | Why an exception was made | 10 minutes per case | A case library |
| Document clean-up | Work instructions, maintenance logs | Review only | Current, versioned documents |
| Recorded Q&A with the successor | The gaps nobody thought to ask about | 1 hour a week | A growing FAQ |
| AI agent answering from the above | Everything above, retrievable in seconds | 30 minutes a week reviewing answers | Answers with a link to the source |
The order matters. Recordings come first because they need no preparation: a phone on a tripod next to the engineer while they price tomorrow's RFQ is enough. Rules come second, because the recordings show which rules exist. The agent comes last, because it can only be as good as the documents under it. Our guide on how to price a machined part shows what a written quoting rule looks like.
What does an agent that answers from documents look like?
Several large Portuguese manufacturers are already heading this way. In 2024 Sonae Arauco described its next project as an intelligent maintenance assistant that crosses the data captured from its machines with large language models in a question-and-answer format; its digital director stressed that every answer "passou pela nossa base de conhecimento", passed through the company's own knowledge base [6]. At Bosch, Rückert told the story of a maintenance worker with 27 years in the job who now relies on a team of AI agents to diagnose faults in seconds, and still makes the decision himself [1]. Siemens Portugal's CEO argues that the real differentiator is each company's proprietary, reliable information, built up over decades of maintenance and engineering [8].
For most small plants this would be the first agent of any kind: only 9.4% of Portuguese firms with 10 to 49 staff used any AI technology in 2025, against 17.0% in the EU [13]. A knowledge AI agent for a 20–200-person plant follows four rules:
- It answers only from your documents and shows the source for every answer: the recording, the rules table, the work instruction.
- It says "I don't know" when there is no source, instead of guessing.
- It logs every question. The unanswered ones are the list of knowledge still missing.
- The expert reviews answers weekly while they are still in the building. Every correction becomes a new rule or a new document.
It reads through read-only connections: the shared drive, the maintenance log, the ERP. Nothing is written back without a person approving it.
Why keep a human in the loop?
Because the knowledge belongs to the engineer, and the review is how it is captured. An agent left alone produces confident mistakes. Siemens' CEO put it plainly: "O humano vai ser sempre um pilar", people will always be a pillar [8]. After the expert retires, the successor takes over the review.
A 12-month plan if they retire next year
- Months 1–2: find the risk. List the 20 questions only this person can answer today. Record five real walkthroughs. Pick one area to start: quoting or the maintenance of one critical machine.
- Months 3–6: write it down. Build the quoting rules table or the maintenance playbook with the expert. Clean up the documents. Set up a read-only assistant on top, used first by the successor.
- Months 7–12: hand over. The successor answers first, the agent second, the expert third. Count every month how many questions still reach the expert. That number should fall. Knowledge capture is one of the seven workflows in AI agents in a factory: seven use cases that pay back.
GETMILK's worked plan: what does capture cost in a year?
This is our own worked plan for one area, quoting or one critical machine, at GETMILK's published prices, excluding VAT.
- Month 0: the free 30-minute call. We list the questions only this person can answer and pick the area. An on-site diagnostic, if you want it written down, is €490 and is credited in full against the project.
- Months 1–3: groundwork, €5,700. An engineer inside your team one day a week through factory modernisation, €1,900 a month with a three-month minimum: recording set-up, eight two-hour rules sessions with the expert, document clean-up.
- Months 4–5: the assistant, €3,900. An AI agent pilot that answers from the recordings, the rules table and the documents, read-only, two to four weeks on live questions, the expert approving every answer.
- Months 6–12: run it, €1,330. Run and support at €190 a month for seven months, model usage billed at cost.
- Total: €10,930 for the year, plus model usage. At Eurostat's average manufacturing labour cost of €16.5 an hour [15], that equals about 660 hours of an average worker's time (our arithmetic); a senior engineer costs more.
- The expert's time: about 70 hours over the year (39 hours of walkthroughs, 16 of rules sessions, 15 of reviewing answers).
Terms are fixed in writing: 30% at the start and 70% only when it works on your real data. You own the code, the documents and the data.
What to do next
- Write the list this week. Sit with your most experienced engineer for one hour and write down the 20 questions that only they can answer today. That list is the scope.
- Record one real job. A phone on a tripod while they price tomorrow's RFQ or diagnose the next fault. If your work instructions are current and the successor is already in place, recordings and weekly Q&A may be all you need, and you do not need us.
- If the list is long and the date is close, GETMILK runs the plan above: the groundwork from €1,900 a month and the AI agent pilot from €3,900, excluding VAT. The free 30-minute call tells you which parts you can do alone.
References
- Business-IT. Bosch ConnectedWorld 2026: a fábrica que aprende e o talento que falta (28 July 2026)
- PORDATA, Fundação Francisco Manuel dos Santos (data: INE, Eurostat). Pordata retrata o perfil do trabalhador em Portugal (press release, Dia do Trabalhador) (29 April 2024)
- European Commission. 2026 Country Report: Portugal, SWD(2026) 222 final (3 June 2026)
- CIP. AIMMAP: indústria, fileira exportadora (4 August 2026)
- AEP. Atividade. Boas sensações para o 2.º semestre (17 August 2026)
- ECO. Sonae Arauco quer "domesticar as fábricas" para serem mais eficientes (29 June 2024)
- AFIA (originally RHmagazine 163). Entrevista a Domingos Pinto: Quando os dados guiam a produção (Simoldes) (22 July 2026)
- ECO. IA a alucinar em ambiente industrial "pode custar vidas", alerta CEO da Siemens (15 July 2026)
- ECO. Continental Mabor contrata mais 200 e tem "vários investimentos em curso" em Famalicão (11 June 2026)
- Eurostat. Employment by sex, age and economic activity (lfsa_egan2), manufacturing, Portugal and EU, 2015 and 2025 (updated 10 September 2026)
- European Investment Bank. EIB Investment Survey 2025: Portugal overview (9 December 2025)
- AIMMAP. Plano de Atividades e Orçamento 2026 (17 December 2025)
- Eurostat. Enterprises using AI technologies, by size class (isoc_eb_ai), 2025 (2025 data, accessed 28 September 2026)
- ECO. Está confirmado. Idade da reforma volta a subir em 2027 para 66 anos e 11 meses (on Portaria n.º 476/2025/1) (29 December 2025)
- Eurostat. Labour cost levels by NACE Rev. 2 activity (lc_lci_lev), manufacturing, 2025 (data updated 23 April 2026)
Questions people ask
- How do you capture knowledge from retiring employees?
- In three layers. Record the person doing real work (a quote, a fault diagnosis) and transcribe it. Turn their rules of thumb into written tables, such as setup times per part family or surcharges for tight tolerances. Then put an AI agent on top that answers questions from those documents, cites its source, and is reviewed by the expert while they are still there.
- How long does it take to capture an engineer's knowledge?
- Plan for 12 months if the person is retiring next year: two months to find and record the critical knowledge, four months to write rules and build a read-only assistant, and six months in which the successor answers first and the expert reviews. It takes one to two hours a week of the expert's time, not a full-time project.
- Can AI replace an experienced engineer?
- No. An agent can make the engineer's knowledge searchable and draft answers from it, but decisions stay with people. Bosch's own example is a maintenance worker with 27 years' experience who now uses AI agents to diagnose faults and still makes the decision himself.
- Is it safe to put our technical documents into an AI assistant?
- It can be, if the documents stay in your systems, the assistant reads them through scoped read-only connections, answers carry a link to the source, and every question is logged. Model calls should go to providers whose business terms exclude training on your data, on an EU endpoint if you need EU processing.
- Where should a factory start with knowledge capture?
- With the 20 questions only one person can answer today. They show where the risk is concentrated, usually quoting, maintenance of one critical machine, or one demanding customer's specifications.