Seven agents pay back in a typical Portuguese plant, all around the order: RFQ intake, quote drafting, drawing checks, supplier lookups, order-status answers, certificate packs and a knowledge assistant. At 300 RFQs a month, intake plus quote drafting frees about 150 hours, worth €2,475 at Eurostat's €16.5 an hour, against a model bill of about €64. Below roughly 100 RFQs a month, hours alone rarely justify the build.
Seven AI agents pay back in a typical Portuguese plant, and all of them work on the paperwork around the order, not on the machines: RFQ intake, quote drafting, drawing and data checks, supplier lookups, order-status answers from the ERP, certificate packs and a knowledge assistant. At 300 RFQs a month, intake plus quote drafting frees about 150 hours. At Portugal's 2025 manufacturing labour cost of €16.5 an hour [1] that is €2,475 a month, while the model bill for the same 300 RFQs is about €64 in GETMILK's worked example [2] [3]. The limit is volume: below roughly 100 RFQs a month, hours alone rarely justify a build.
Key facts
- Hourly labour cost in Portuguese manufacturing was €16.5 in 2025, against €35.0 in the EU [1].
- 11.5% of Portuguese enterprises with 10 or more staff used AI in 2025, against 20.0% in the EU; among firms with 10 to 49 staff the Portuguese share was 9.4% [6] [7].
- 24% of Portuguese firms make systematic use of generative AI to improve processes, against 37% in the EU; 88% name the availability of skilled staff as an obstacle to investment, against 79% (EIB survey, December 2025) [5], and the metalworking association AIMMAP says talent shortage is still the biggest obstacle to growth [4].
- Only 10–15% of companies experimenting with AI reach effective operational impact, according to Siemens Portugal's CEO (July 2026) [8].
- GETMILK's worked example: 300 drawing-based RFQs a month cost $72.55 in model usage on Claude Sonnet 5, about €64 at the ECB rate of 1.1403 dollars per euro on 25 September 2026 [2] [3].
Which AI agents actually pay back in a factory?
Siemens Portugal's CEO advises starting with "as decisões mais frequentes, com menor risco e maior custo", the most frequent decisions, with lower risk and higher cost [8]. In a 20–200-person plant those sit around the order. The table shows the seven we build most often. Hours are GETMILK's assumptions for 300 RFQs a month.
| Use case | What the agent does | Hours freed a month (our assumption) | Data it needs | Risk if it is wrong |
|---|---|---|---|---|
| 1. RFQ intake and triage | Reads email and portals, extracts customer, part, quantity and date, checks completeness | 50 h (300 × 10 min) | Mailbox, customer list | Low: a person sees every item |
| 2. Quote drafting | Reads the drawing or BOM, applies your price rules, drafts the quote in your template | 100 h (300 × 20 min) | Price rules, machine rates, past quotes | Medium: nothing is sent without approval |
| 3. Drawing and data checks | Checks material, tolerances and finishes against what the plant can make; flags gaps | 75 h of senior engineer time (300 × 15 min) | Capability table, drawings | Medium: the engineer checks the flags |
| 4. Supplier and material lookups | Finds stock, lead times and last prices; drafts supplier RFQs | 20 h | ERP purchase history, supplier list | Low: the buyer sends |
| 5. Order-status answers | Answers "where is my order" from the ERP and drafts the reply | 15 h (180 questions × 5 min) | Sales, production and shipping records | Low if read-only and drafted |
| 6. Certificate and document packs | Assembles material certificates and inspection documents per shipment | 25 h (60 shipments × 25 min) | Document store, lot traceability | Medium: checklist and sign-off |
| 7. Knowledge assistant | Answers from work instructions, manuals and maintenance logs, with the source | 20 h, hard to measure | Your documents | Low with sources; never alone for safety decisions |
Use cases 1 to 3 are the core of RFQ and quoting automation. Use cases 4 to 7 are separate agents that reuse the same connections to the ERP.
Bosch showed where use case 7 leads: its chief digital officer described a maintenance worker with 27 years in the job who now diagnoses faults with a team of AI agents and still makes the decision himself [9]. The retirement side of that story is in capture what your best engineer knows.
Why do so many AI pilots in factories stall?
Portugal is behind on adoption, not ahead. Eurostat counts 11.5% of Portuguese firms using any AI technology in 2025, against 20.0% in the EU, and the gap is widest in small firms [6] [7]. Those that try often stall. Siemens Portugal's CEO put effective impact at 10–15% of experimenting companies and named three causes: unstructured data, processes nobody updated, and people who do not adopt the tools [8].
Two more patterns follow: nobody measured the work beforehand, and the pilot tried to automate a whole chain. Bosch's CTO Mathias Pillin said that to scale AI "não se pode ficar nos 80%, é preciso chegar aos 99%", you cannot stop at 80% [9]. The arithmetic shows why: five steps that are each right 95% of the time, unchecked, are right together only about 77% of the time (0.95 to the fifth power, our calculation). Critical Manufacturing's CEO makes the data point from the software side: execution, data and analytics "can no longer operate in isolation" [10].
Which one should you build first?
Pick the use case that passes four tests before anyone writes code:
- Hours. Frequency times minutes. Below roughly 40 hours a month the saving rarely covers build and running cost (our rule of thumb, not a published benchmark).
- Data that already exists. The mailbox, the drawings, the ERP tables, the Excel price sheet. If the data must be created first, it is a data project.
- Reversibility. A draft quote reviewed by the estimator is reversible. A purchase order sent automatically or a machine parameter changed by an agent is not.
- Human approval. The Model Context Protocol, which is how agents reach ERP data, says there should always be a human in the loop with the ability to deny tool invocations [11].
RFQ intake and quote drafting pass all four in most plants. Production scheduling and automatic purchasing fail reversibility until the upstream data is clean.
How do you calculate the payback?
Net monthly saving equals hours freed times the cost of an hour, minus the monthly running cost. Payback in months equals the build price divided by the net monthly saving. We use Eurostat's €16.5 for the hour [1]; estimators cost more than the average, so this understates the saving on use cases 2 and 3.
At that rate, an integrated agent at GETMILK's starting price of €7,900 pays back in 12 months if it frees about 55 hours a month: €7,900 over 12 months is €658, plus €190 running and €64 of model usage is €912, and €912 divided by €16.5 is 55 hours.
What does it cost to run an agent each month?
In our worked example a plant receives 300 RFQs a month, each with a three-page drawing. For each one the agent triages the email, extracts the drawing into structured data, runs six ERP lookups and drafts the quote and reply; we add 25% for retries. On Claude Sonnet 5, at $2 per million input tokens, $10 per million output tokens and $0.20 for cached input [2], with a small model for triage, the month comes to $72.55: about €64 at the ECB reference rate [3], or €0.21 per RFQ. The model bill does not decide the payback. Hours do.
GETMILK's worked example: 300 RFQs a month
- Assumption: an estimator spends 45 minutes per RFQ today and 15 minutes reviewing the agent's draft. Saving: 30 minutes × 300 = 150 hours a month.
- Value: 150 × €16.5 = €2,475 a month.
- Cost: RFQ and quoting automation from €4,900 to set up and from €290 a month per plant, model usage up to an agreed volume included.
- Net: €2,475 − €290 = €2,185 a month. Payback: €4,900 ÷ €2,185 ≈ 2.2 months.
- Stress test: a setup tripled to €14,700 still pays back in under 7 months.
These are our prices and assumptions, not a quote. The full price range is in how much an AI agent costs for a factory.
When does an agent not pay back?
- Low volume. At 80 RFQs a month and 20 minutes saved each, the agent frees about 27 hours, worth €440. A pilot at €3,900 with €190 running and about €17 of model usage nets €233 a month and takes around 17 months to pay back. Then the case rests on speed winning orders, which you must measure.
- Messy data. If the price rules exist only in one person's head, the first project is writing them down. How to price a machined part shows what a written rate sheet looks like.
- Cheap hours. Portuguese labour costs less than half the EU average [1], so the same agent pays back about twice as fast in a plant paying €35 an hour.
What should you ask vendors this autumn?
The Portugal Digital Summit on 21–22 October in Lisbon is titled "A Era dos Agentes de IA" [12]; IAPMEI's free SME lab in Braga on 26–27 October asks each firm to bring a real challenge from its own operation [13]; and Web Summit's AI Summit (9–12 November) has sessions on "The ROI of AI" and "The agentic era", with the organisers noting that many organisations are not ready for agents [14]. Ask every vendor:
- What accuracy do you reach on documents like ours, and how was it measured?
- Who approves each output, and where is every decision logged?
- What does the agent do when data is missing?
- Can we stop after the pilot and keep the code and the data?
- What does it cost per month to run, including model usage?
A vendor who answers with a demo is selling a demo.
What to do next
- Count one workflow for two weeks. Pick RFQ intake or quote drafting and record items per week, minutes per item and turnaround time.
- Write the acceptance criteria before any code. For example: 80% of RFQs classified correctly without edits, and a quote draft ready within four working hours. If the counts show fewer than about 40 hours a month, stop here: you do not need an agent yet.
- If the numbers clear the bar, a GETMILK pilot agent on one workflow starts at €3,900 excluding VAT: read-only connections, two to four weeks on live work, a person approving every output. The price is fixed in writing, 30% is paid at the start and 70% only when it works on your real data, and you own the code and the data.
References
- Eurostat. Labour cost levels by NACE Rev. 2 activity (lc_lci_lev), manufacturing, 2025 (data updated 23 April 2026)
- Anthropic. Claude API pricing (accessed 28 September 2026)
- European Central Bank. Euro foreign exchange reference rate, US dollar (EXR.D.USD.EUR.SP00.A) (25 September 2026)
- CIP. AIMMAP: indústria, fileira exportadora (4 August 2026)
- European Investment Bank. EIB Investment Survey 2025: Portugal overview (9 December 2025)
- Eurostat. Enterprises using AI technologies, by size class (isoc_eb_ai), 2025 (2025 data, accessed 28 September 2026)
- Eurostat. 20% of EU enterprises use AI technologies (11 December 2025)
- ECO. IA a alucinar em ambiente industrial "pode custar vidas", alerta CEO da Siemens (15 July 2026)
- Business-IT. Bosch ConnectedWorld 2026: a fábrica que aprende e o talento que falta (28 July 2026)
- ManufacturingTomorrow. Critical Manufacturing to demonstrate MES-powered industrial operations platform at Hannover Messe 2026 (company press release) (19 March 2026)
- Model Context Protocol. Model Context Protocol specification, Tools (revision 2026-07-28) (28 July 2026)
- ACEPI. Portugal Digital Summit 2026: A Era dos Agentes de IA (accessed 28 September 2026)
- IAPMEI. Laboratório de IA e Transformação Digital SpaceX AI (16 September 2026)
- Web Summit. AI Summit track (accessed 28 September 2026)
Questions people ask
- What are the best AI agent use cases in manufacturing?
- For a 20–200-person plant, the ones around the order: RFQ intake and triage, quote drafting from drawings, drawing and technical-data checks, supplier and material lookups, order-status answers from the ERP, certificate and document packs, and a knowledge assistant that answers from your own documents. All of them draft for a person to approve.
- Why do most AI pilots in factories fail?
- Mostly for reasons that are not technical. Siemens Portugal's CEO said in July 2026 that only 10–15% of companies experimenting with AI get effective impact, and named unstructured data, outdated processes and tools nobody adopts. Pilots also fail when nobody measured the work beforehand or when they try to automate a whole chain without a person checking.
- How much does it cost to run an AI agent every month?
- The model bill is small. In GETMILK's worked example, 300 RFQs a month with three-page drawings and six ERP lookups each cost about €64 a month on Claude Sonnet 5 with prompt caching, or €0.21 per RFQ. GETMILK's run and support fee starts at €190 a month, with model usage billed at cost.
- Is an AI agent worth it for a small factory?
- It depends on volume. At 80 RFQs a month, an agent that saves 20 minutes each frees about 27 hours, worth about €440 at €16.5 an hour, and a pilot takes around 17 months to pay back on hours alone. Below about 100 RFQs a month the case rests on faster answers winning more orders, which you should measure.
- What should I ask an AI agent vendor?
- What accuracy it reaches on documents like yours and how that was measured, who approves each output and where decisions are logged, what happens when data is missing, whether you keep the code and data if you stop after the pilot, and the full monthly running cost including model usage.