AI strategy 2027: 7 moves Luxembourg SMEs should make now

Learn more about AI implementation in Luxembourg.

October is when Luxembourg companies set next year's budget, and the question I hear most in these weeks is a version of "what should we actually do with AI in 2027?" Not "should we use AI": by 2025 a fifth of EU enterprises with ten or more staff already did, up from 13.5% a year earlier, according to Eurostat's survey on AI use in enterprises. Using AI is no longer an edge. Using it on the right work, with data and governance that let you move faster than competitors, is.

2027 is also the first year in which AI economics and AI regulation arrive together. Four dates sit between December 2026 and January 2028, and three of them reward the same groundwork. That overlap is the opportunity this article is about: one programme, several deadlines, and a budget that the state co-funds.

Before you budget: most of the moves below fit inside SME Packages – AI, which reimburses 70% of a €3,000–€25,000 project (Guichet.lu). Check whether your project qualifies in 2 minutes, no email needed.

What changes for AI in Luxembourg between now and 2028?

Four hard dates, each with its legal instrument. If a date below moves, the instrument is where it will move.

Article 50 transparency duties (telling people they are talking to an AI, labelling deepfakes) already apply since 2 August 2026. We cover the detail in our post on what the August 2026 AI Act date actually changed, and the high-risk file in is your AI high-risk under the new 2 December 2027 deadline.

The bill 8815 dates come from the government's own e-invoicing portal and the parliamentary dossier. They can still change before the vote; we track every step on our e-invoicing 2028 tracker.

What will separate the winners from the rest in 2027?

Not access to models. Every SME in Luxembourg can buy the same frontier models for a few dozen euros per seat. What will separate firms in 2027 is three things the competitor cannot copy by signing up to the same tool:

1. Workflows redesigned around AI, not AI bolted onto old workflows. A chatbot on top of a messy process produces fast mess.
2. Clean, owned data. Supplier files, client files and document archives that an AI system can read without a person fixing every third record.
3. Governance that speeds you up. A register, a classification and a sign-off rule mean a new use case is approved in a week instead of debated for a quarter.

The cautionary number for 2027 comes from Gartner, which in June 2025 predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs, unclear business value and weak risk controls. All three causes are decided before a line of code is written. That is why the moves below are mostly about choosing and preparing, not about technology. We looked at what the surviving projects have in common in why 40% of AI agent projects die and what the rest do differently.

The 7 moves to make for 2027

1. Fund two workflows with a measured baseline, and stop the rest

Pick two processes where you can count today's cost: hours per week, error rate, turnaround time. Invoice intake, client onboarding, quote preparation and recurring reporting are the usual candidates in a 20–200 person Luxembourg company. Write the baseline down before anything is built. A project without a baseline cannot prove it worked, and a project that cannot prove it worked is the first one cut when budgets tighten.

The harder half is stopping things. Most SMEs I meet have five to ten informal AI experiments running. Keep the two with a baseline and an owner; let the others end.

2. Build one AI register that serves three purposes

List every AI system the company uses or provides: the chat assistant, the CV-screening feature in your HR tool, the chatbot on your website, the meeting note-taker. For each, record who uses it, on what data, and how it is classified under the AI Act.

That single list does three jobs: it shows where Article 50 disclosures are missing today, it identifies any Annex III system that needs a full file by 2 December 2027, and it is the evidence base for your AI literacy measures under Article 4. One spreadsheet, three compliance answers. If you are unsure whether you are a provider or a deployer for a given tool, our 5-minute provider vs deployer test settles it.

3. Fix your master data once, for 2027 and 2028

E-invoicing needs correct VAT numbers and Peppol IDs for every supplier and customer. AMLR needs verified identification data and beneficial owners for every client of an obliged entity. AI automation needs records clean enough to match against. These are the same tables.

Doing the data work once, in early 2027, is cheaper than doing it three times under three deadlines. Start with the Peppol ID guide if e-invoicing is your first constraint, or with KYC remediation before AMLR if you are an obliged entity.

4. Move from chat to agents only where a person signs the output

Agents that read, check and draft are ready for production in 2027. Agents that decide on their own are where the Gartner cancellations will come from. The rule I use: an agent can do any amount of reading and preparation, as long as a named person approves what leaves the company or touches the ledger. That keeps the value and removes most of the risk.

5. Train people for the tools they actually have

Article 4 AI literacy has applied since 2 February 2025, and the Omnibus turned it into a duty to support literacy rather than guarantee it (see our Article 4 explainer). The business case is stronger than the legal one: the same licence in the hands of a trained and an untrained employee produces very different output. Budget a short awareness session for everyone and a few hours of role-specific practice for heavy users, and keep the attendance record.

6. Make sure AI answer engines can find and name you

Your next client may never see a list of blue links. Google switched on AI Mode and AI Overviews in France in July 2026, and AI answers are now a standard part of search in Luxembourg's neighbouring markets. If those answers cite your competitors and not you, no amount of traditional SEO fixes it. We measured what this looks like for a Luxembourg firm in Google AI Mode: cited is not the same as recommended.

7. Budget with the funding, not around it

Luxembourg co-funds exactly the kind of projects listed above. SME Packages – AI reimburses 70% of eligible costs for projects between €3,000 and €25,000, so at most €17,500 (Guichet.lu). Fit 4 AI co-funds an upstream study at 50% for SMEs. The aid is reimbursed after the project, not paid upfront, so cash flow still needs planning. Our SME Packages – AI guide covers eligibility and the application.

Mid-way check: if you already know your two workflows, the fastest next step is a written estimate with the grant deducted. Ask for one on our contact page, and we reply with a scope and a price, not a sales call.

How much should a Luxembourg SME budget for AI in 2027?

Using our published project tiers and the SME Packages – AI rate, here is what the three typical project sizes cost after the grant. These are our own "from" prices, excluding VAT; your project may cost more, and only the Ministry decides eligibility.

Two things the table hides. First, the grant is per project, so two well-separated workflows can each be a project. Second, the cost that most often surprises owners is internal time: someone has to own the workflow, test the output and change how the team works. Budget a few hours a week of a manager's time for the first two months, and count it. Our AI cost guide for Luxembourg SMEs breaks down the running costs too.

What should you stop doing in 2027?

Three habits cost Luxembourg SMEs the most in 2026, and they are worth naming because they feel productive:

• Buying seats without measuring use. Licences for a whole team, used seriously by three people. Measure usage after 60 days and reallocate.
• Pilots without an owner. A pilot owned by "IT" or by the vendor ends when the vendor's attention does. Every project needs a business owner whose numbers it changes.
• Treating the AI Act as a 2028 problem. Transparency duties apply now, and an Annex III file takes twelve to eighteen months. Sixteen months from August 2026 is a schedule, not a reprieve.

A 90-day plan for the first quarter of 2027

By the end of March you have a working system, a compliance register, cleaner data for 2028 and a measured result to justify the second project. That is a better position than most of your competitors will be in, and it costs less than a single unused licence programme.

If you want a second opinion on which two workflows to pick, our AI consulting service starts with exactly that assessment, and our step-by-step AI roadmap guide shows how to extend the plan beyond Q1.

Frequently asked questions

Is it too late to start with AI in 2027?

No. A fifth of EU enterprises used AI in 2025, so most of your competitors are still early. What is late is starting without a baseline and an owner: those projects are the ones that get cancelled.

Which AI deadline matters most for a Luxembourg SME in 2027?

For most SMEs, the 1 January 2028 e-invoicing receiving date (bill 8815, as drafted) needs the most preparation in 2027. For firms using AI in recruitment or credit decisions, the 2 December 2027 AI Act high-risk date comes first. Obliged entities under AMLR face 10 July 2027.

How much does an AI project cost a Luxembourg SME after funding?

From about €900 net for a €3,000 quick win to €7,500 net for a €25,000 project, if it qualifies for SME Packages – AI at 70%. The grant is paid after the project.