Siemens Portugal's CEO estimates that only 10–15% of companies experimenting with AI get real operational impact, and blames unstructured data, unchanged processes and tools nobody uses. The first agent should pass four tests: enough hours, data that already exists, mistakes that can be caught before they cost anything, and a person approving every output. RFQ intake and quote drafting usually pass; production scheduling usually does not.
Most AI pilots in factories fail for reasons that have little to do with the model. In July, Siemens Portugal's CEO Sofia Tenreiro said that practically every company is experimenting with AI tools, but only 10–15% are getting effective impact in their operations, and she named three causes: unstructured data, processes nobody updated, and the difficulty of getting people to use the new tools [1]. The way out is to choose the first agent by four tests, not by what looks good in a demo: enough hours, data that already exists, mistakes that can be caught before they cost anything, and a person approving every output. In most plants that points to RFQ intake or quote drafting, not to production scheduling.
Key facts
- Only 10–15% of companies experimenting with AI reach effective operational impact, according to Siemens Portugal's CEO (14 July 2026) [1].
- Her three causes: lack of structured data, processes that need updating, and workers who do not adopt the tools [1].
- Bosch's CTO Mathias Pillin: to scale AI in industry "you can't stop at 80%, you must reach 99%" [2].
- 11.5% of Portuguese enterprises used AI in 2025, against 20% in the EU [3].
- The Portugal Digital Summit (21–22 October 2026, Lisbon) is titled "A Era dos Agentes de IA", the age of AI agents [4]; IAPMEI's SME hackathon in Braga on 26–27 October asks each firm to bring a real problem from its own operation [5].
Why do most AI pilots in factories fail?
The Siemens diagnosis is worth reading slowly, because none of its three causes is technical. Unstructured data means the price rules live in one estimator's head and the drawings live in email threads. Processes that were never updated means the agent is bolted onto a workflow that already had five manual workarounds. And tools nobody uses means the pilot ran in a meeting room, not on the desk of the person who answers RFQs.
Two more patterns follow from the same logic. The first is that nobody measured the work before the pilot, so nobody can prove a saving after it. The second is that the pilot tried to automate a whole process end to end. That is where Bosch's warning bites. Pillin's point was that industrial AI has to reach 99% reliability to scale, not 80% [2]. Simple arithmetic shows why long chains are fragile: if each of five steps is right 95% of the time and nobody checks in between, the whole chain is right only about 77% of the time (0.95 to the fifth power). A person approving the output at the end turns that 77% into a draft that is corrected before it leaves the building.
The vendors in Portugal's own industrial software scene say the same thing about data. Critical Manufacturing's CEO argues that execution, data and analytics "can no longer operate in isolation" [8]. An agent that cannot see the ERP, the price sheet and the drawing at the same time is guessing.
What makes a good first agent?
Tenreiro's own advice is to start with the most frequent decisions, with lower risk and higher cost [1]. We turn that into four tests. A candidate should pass all four before anyone writes code.
- Hours. Multiply how often the task happens by the minutes it takes. A task done 300 times a month for 30 minutes is 150 hours. Below roughly 40 hours a month, the saving rarely covers the build and the running cost (our rule of thumb, not a published benchmark).
- Data that already exists. The inputs must be there today: the mailbox, the drawings, the ERP tables, the Excel price sheet. If the data has to be created first, it is a data project, and it should be scoped as one. Siemens calls proprietary, reliable information the real differentiator [1]; the first agent should sit on top of the information you already trust.
- Reversibility. A mistake must be caught before it costs money. A draft quote reviewed by the estimator is reversible. A purchase order sent automatically is not. A machine parameter changed by an agent is not.
- Human approval. During the pilot, a person approves every output. The Model Context Protocol specification, which is how agents reach ERP data, says there should always be a human in the loop with the ability to deny tool calls [6]. It is also the only way to measure accuracy honestly.
Which factory workflows pass the four tests?
This is our assessment for a typical 20–200-person plant in Portugal. Your plant may differ, which is what the audit checks.
| Workflow | Hours | Data ready | Reversible | Easy to approve | Verdict |
|---|---|---|---|---|---|
| RFQ intake and triage | High | Yes: mailbox and portals | Yes | Yes | Good first agent |
| Quote drafting from drawings | High | Usually: price rules in Excel | Yes: it is a draft | Yes | Good first or second |
| Order-status answers from the ERP | Medium | Yes, read-only | Yes, if drafted | Yes | Good |
| Certificate and document packs | Medium | Often scattered | Yes | Yes | Second wave |
| Predictive maintenance | Varies | Needs sensor history | Partly | Hard | Data project first |
| Production scheduling | High | Routings often incomplete | Low: the plan goes to the floor | Hard | Not first |
| Automatic purchasing | Medium | Yes | Low: orders leave the company | Defeats the purpose | Not first |
The pattern is clear. Front-office and back-office paperwork around the order passes. Anything that acts directly on the shop floor or on suppliers fails the reversibility test until the upstream data is clean. We describe the seven workflows that pay back, with hours and payback maths, in AI agents in a factory: seven use cases that pay back.
How do you run the pilot so it proves something?
- Measure a baseline first. Two weeks of minutes per item, error rate and turnaround time, recorded by the people who do the work.
- Write the acceptance criteria before coding. For example: 80% of RFQs classified correctly without edits, and a quote draft ready within four working hours.
- Connect read-only. The agent reads the ERP, the mailbox and the shared drive; it writes nothing that leaves the company.
- Run two to four weeks on live work, with the estimator or planner approving every output.
- Decide on numbers. Scale it, fix it, or stop. A pilot that ends with "interesting" has failed.
What does it cost to find out?
Less than most firms expect. A GETMILK pilot agent starts at €3,900 excluding VAT for one workflow, with read-only connections, two to four weeks on live work and human approval on every output. The price is fixed in writing. 30% is paid at the start and the other 70% only when the agent works on your real data, so the risk of a pilot that goes nowhere sits mostly with us. You own the code and the data. Running it afterwards starts at €190 a month plus model usage at cost. The full price picture is in how much an AI agent costs for a factory.
If you do not yet know which workflow passes the four tests, start with the free 30-minute digitalisation audit. It ends with the three workflows we would automate first, ranked by hours and payback.
What to ask vendors at the Portugal Digital Summit and Web Summit
October and November put AI agents on every stage. The Portugal Digital Summit runs on 21–22 October at the Centro de Congressos de Lisboa [4]; IAPMEI's hackathon for SMEs follows in Braga on 26–27 October [5]; and Web Summit's AI Summit (9–12 November) has sessions titled "The agentic era" and "The ROI of AI", with the organisers noting that many organisations are not ready for agents [7]. Five questions separate the pilots that will work from the ones that will not:
- What accuracy do you reach on documents like ours, and how was it measured?
- Who approves the agent's 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 will it cost per month to run, including model usage?
A vendor who answers all five in numbers is selling a pilot. A vendor who answers with a demo is selling a demo.
References
- 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)
- European Commission. 2026 Country Report: Portugal, SWD(2026) 222 final (3 June 2026)
- ACEPI and GoingNext. Portugal Digital Summit 2026: A Era dos Agentes de IA (2026)
- IAPMEI. Laboratório de IA e Transformação Digital SpaceX AI (16 September 2026)
- Model Context Protocol. Model Context Protocol specification, Tools (revision 2026-07-28) (28 July 2026)
- Web Summit. AI Summit track (2026)
- ManufacturingTomorrow (company press release). Critical Manufacturing to demonstrate MES-powered industrial operations platform at Hannover Messe 2026 (19 March 2026)
Questions people ask
- Why do AI pilots fail in manufacturing?
- Mostly for reasons that have little to do with the model. Siemens Portugal's CEO names three: data that is not structured, processes that were never updated, and people who do not adopt the tool. Pilots also fail when nobody measured the work before, so nobody can prove a saving after.
- What is the best first AI agent for a factory?
- One that passes four tests: it takes enough hours a month, it works on data you already have, its mistakes can be caught before they cost anything, and a person approves every output. RFQ intake and quote drafting usually pass all four. Production scheduling and automatic purchasing usually do not.
- How long should an AI pilot take in a factory?
- Two to four weeks on live work is enough to measure hours saved and errors caught on one workflow, provided the data connections are in place and the baseline was measured before the pilot started.
- How much does an AI pilot cost for a small factory?
- With GETMILK, a pilot on one workflow starts at €3,900 excluding VAT, fixed in writing. 30% is paid at the start and 70% only when the agent works on your real data. The first 30-minute audit call is free.
- How do you measure whether an AI pilot worked?
- Measure minutes per item, error rate and turnaround time for two weeks before the pilot, then the same numbers during it, plus the share of drafts a person accepted without changes. If the hours saved do not cover the running cost, stop or change the workflow.