AI Strategy for Mid-Sized Manufacturers: From Goals to an Executable Agenda
Most mid-sized companies already use some AI, but rarely with a shared direction. Sales experiments with a chat assistant, IT is comparing licences, the data protection officer is worried, and the board wants to know what any of it is for. An AI strategy turns that patchwork into a plan. This guide covers the building blocks and, more importantly, how they become an agenda that actually gets worked through.
What is an AI strategy, and why do mid-sized companies need one?
An AI strategy states which business goals a company pursues with AI, in which processes, under which rules, with which people and on what budget. It is a decision document rather than a technology paper: it says what will be done, what will deliberately not be done, and how success will be recognised.
The gap is measurable. In Germany, a useful reference market for manufacturing-heavy mid-sized firms, the Federal Statistical Office (Destatis) reports that 26% of enterprises with at least 10 employees used AI in 2025. Among firms with 50 to 249 employees the figure was 36%; among firms with 250 or more it was 57%. Adoption grows with size, and smaller firms do not catch up by default.
The reasons for holding back are even more telling. Among German firms that had considered AI but were not using it, the 2025 Destatis survey found 72% citing a lack of knowledge, 62% unclear legal consequences and 60% data protection concerns. Cost came much further down, at 32%. The main barriers are knowledge, rules and trust, which is exactly what a strategy is meant to address.
Common triggers for writing one:
- Tool sprawl. Several departments use different AI services and nobody has an overview of data flows or contracts.
- Pilot purgatory. Two or three pilots have run, none has reached day-to-day operation.
- External pressure. Customers ask for an AI policy, competitors advertise AI features, regulators expect evidence.
- Retiring expertise. Experienced staff are leaving, and the company hopes AI can help capture what they know.
What are the building blocks of an AI strategy?
A workable AI strategy has five building blocks that depend on each other: a target picture, a use-case portfolio, governance, ownership and skills, and budget. Leave one out and a familiar failure follows: no target picture means drift, no portfolio means busywork, no governance means risk, no ownership means stalling, and no budget means slideware.
The value lies in the connections. A high-priority use case is only as good as the rules it runs under and the person accountable for it. That is why all five blocks belong on a single page, side by side.
1. Target picture
The target picture translates business goals into AI goals. It does not start with "we want to use AI". It starts with the three to five issues the leadership team is already worried about, such as lead times, quote turnaround, service quality, hiring or margins. For each, it describes what role AI could play and which metric should move. Equally important is what is out of scope, for example areas where data is too thin or the risk too high.
2. Use-case portfolio
The target picture produces a list of concrete use cases: pre-qualifying inbound enquiries, answering recurring service questions from manuals, summarising inspection reports. Each candidate is scored for value, feasibility, data readiness and risk, then sorted into a portfolio: a few quick projects for learning, a small number of larger ones that matter strategically. Finding and scoring candidates is a method of its own; the strategy captures the outcome, which is a justified order of work.
3. Governance
Governance covers which tools are allowed, which data may go where, who approves new uses and how outputs are checked. For a mid-sized company a short rulebook with a few clear roles is usually enough. Our guide to lean, auditable AI governance goes into detail. The principle to anchor in the strategy is simple: no use case goes live without a risk rating and a named owner.
If you want a reference structure, the NIST AI Risk Management Framework is intended for voluntary use and is organised around four functions: Govern, Map, Measure and Manage. The framework is extensive, but its outline condenses well into a few pages.
4. Ownership and skills
This block names who owns the strategy, who drives each use case on the business side and how staff will be trained. For companies operating in the EU it is also a legal matter. Article 4 of the EU AI Act requires providers and deployers of AI systems to take measures on the AI literacy of their staff, and according to Article 113 that obligation has applied since 2 February 2025. The 2026 AI Omnibus softened the wording without removing it: organisations no longer have to guarantee a specific level for every individual, but must still take measures to support AI literacy, as the KI-Campus summary explains. Take legal advice for your own situation; this guide is not a substitute.
5. Budget
The budget follows the portfolio, not the other way round. A practical split uses three pots: a baseline (licences, platform, training), the projects in the portfolio, and a reserve for opportunities that come up. Every project gets a stop criterion and a decision date on which it is either continued or closed. This is the simplest protection against pilots that never end. Public grants can support individual projects, but they should not dictate the order of the portfolio.
Which strategy fits which starting point?
There is no single AI strategy for mid-sized firms. The starting point sets the emphasis: companies with little AI use need rules and skills first, companies with scattered pilots need prioritisation, and companies already in production need scaling and operations. Three patterns cover most situations.
Getting started. AI use so far is minimal or uncoordinated. The focus is an acceptable-use policy, one approved tool for everyone, basic training, and two or three small use cases that teach the organisation quickly. The aim is confidence, not a big bang.
Consolidating. There are several pilots, a mix of tools and little that has reached production. The focus is an honest inventory, deliberately closing weak projects, a shared technical base and clear ownership of the survivors. Most of the value here comes from stopping things.
Scaling. One or more applications run in production, for example an assistant for technical service. The focus shifts to operations and monitoring, reusing components across departments, data quality, and deciding which capabilities to build in-house. For concrete application areas see AI agents for business functions and, where data control is a hard requirement, sovereign AI run on your own infrastructure.
What needs to be in place first?
Three conditions decide whether a strategy succeeds: a clear mandate from leadership, an honest view of the data, and one person with time to drive it. Technical prerequisites such as cloud access or on-premise servers can be solved; a missing mandate or missing capacity usually cannot.
Mandate. Leadership does not have to write the strategy, but it has to approve it and keep asking about it. Without that signal, departments treat AI as an extra task that loses to daily business.
Data. Many use cases fail not because of the model but because of scattered or outdated data: manuals as scanned PDFs, quoting know-how buried in personal inboxes, master data full of duplicates. A short inventory of the key data sources belongs before prioritisation, not after.
Systems and contracts. Which systems are authoritative (ERP, CRM, document management)? Which interfaces exist? Which AI contracts are already in place, including those bundled with office suites? The answers show what is possible without heavy integration work.
Employee representatives and data protection. Where a works council or other employee representation exists, as is common in Germany, involve it early, because AI tools often raise questions about performance monitoring. Involve your data protection officer from day one. Most use cases can be operated in line with the GDPR when designed that way from the start, rather than blocked later.
Capacity. At least one person needs a fixed share of their working time for this. If that capacity is not available internally, an interim AI lead can fill the role until internal structures are in place; see fractional head of AI.
How do you build the strategy in six steps?
An AI strategy for a mid-sized company can be built in six steps: clarify goals, take stock, collect use cases, score and rank them, set rules and roles, and approve an agenda with a budget. The length of the exercise matters less than what comes out of it: decisions, not just findings.
- Clarify goals. In one session, the leadership team agrees on the three to five most important business issues for the next two years and one metric for each. Output: a one-page target picture.
- Take stock. Which AI tools are already in use, officially or not? Where does the relevant data live? Which contracts exist? Short interviews with department heads are enough. Output: an honest baseline.
- Collect use cases. Workshops with business units look at real workflows, not technologies. Where do people search, retype, summarise or chase information? Output: a candidate list with a short description each.
- Score and rank. Each candidate is rated on value, feasibility, data readiness and risk. Output: a ranked portfolio and the first two or three projects.
- Set rules and roles. Acceptable-use policy, approval process, risk rating, owner per use case, training plan. Output: a short rulebook and a responsibility map.
- Approve the agenda. Leadership signs off portfolio, budget, owners and review dates. Output: a dated agenda that feeds into the regular management reporting.
The rhythm afterwards is what most strategies lack. A quarterly review works well: check each project against its stop criterion, score new candidates and adjust the ranking. Revisit the target picture about once a year. In our own operations, that fixed review slot has proved to be the difference between a strategy that lives and one that sits in a folder.
What are the limits and risks?
An AI strategy fails when it is too big, too technical or too vague. The usual mistakes are a document without decisions, a portfolio without stop criteria and rules nobody has read. External risks add to that: models and prices move quickly, dependence on a single vendor grows, and regulation keeps evolving.
Strategy as a substitute for action. A thick document can look like progress. If nothing reaches production after approval, the strategy was too abstract. The remedy is to start the first project immediately after sign-off.
Technology first. Strategies that open with a platform choice end up hunting for problems to fit the solution. The order should be goal, use case, then technology.
Vendor lock-in. Tying every application tightly to one provider reduces negotiating power and flexibility when prices or terms change. The strategy should state which data and components stay under your control and how a switch would work in principle.
Underestimated operations. A pilot is quick to build. Running it permanently, with monitoring, knowledge-base updates, user support and cost control, takes ongoing effort. A portfolio that only budgets pilots leaves out the most expensive part.
Adoption. Staff who see AI as a threat will work around new tools or turn to personal accounts. Being open about what AI is meant to take over, and what it is not, belongs in the strategy, as does early involvement of employee representatives.
Moving regulation. Under Article 113, the EU AI Act generally applies from 2 August 2026, some obligations follow later, and parts were already amended in 2026. Build a fixed date into the strategy for reviewing the legal position. This guide offers orientation, not legal advice.
When you do not need a full strategy. Very small firms, or companies that only use AI features built into standard software, often do fine with an acceptable-use policy and a training session. A full strategy pays off once several departments run their own projects or customer and operational data is processed.
Where do you start?
Start with two things in parallel: a quick inventory of which AI tools are already in use, and a leadership session that writes down the key business goals. Together they show which of the three patterns fits and which building blocks are needed first.
A practical sequence for the first weeks:
- Name an owner. One person with a mandate and time, internal or external.
- Put an acceptable-use policy in place. A provisional version beats none, because staff are already using AI services.
- Build skills. Basic training for everyone, deeper sessions for the teams running the first use cases; a hands-on workshop can bring leadership and department heads to a shared baseline.
- Pick the first use case. Small, with good data and a business unit that genuinely wants it.
- Book the first review. The first quarterly review is in the calendar before the first project starts.
For the rules themselves, read our guide to AI governance for mid-sized companies; terms such as large language model or AI agent are explained in the AI glossary. If running the programme alongside daily business is not realistic, the fractional head of AI page describes how to fill the lead role for a defined period.
Frequently asked questions
How long should an AI strategy document be?
Short enough for the leadership team to discuss in full in a single meeting. A few pages covering goals, portfolio, rules, owners and budget are enough; project detail belongs elsewhere.
Does a mid-sized company need a dedicated AI team?
Usually not at the start. What matters more is one named person with a mandate and protected time who connects business units, IT and data protection; delivery can be internal, external or mixed.
How often should an AI strategy be revised?
Review the portfolio quarterly and the target picture roughly once a year. Models, prices and regulation change faster than in traditional IT planning, so fixed review dates matter more than a perfect first draft.
Does the EU AI Act require a company to have an AI strategy?
Not as such. Organisations that deploy AI systems in the EU must, however, take measures to support the AI literacy of their staff under Article 4, and the strategy is a natural place to anchor that. This is not legal advice.
Sources
- Destatis (Federal Statistical Office of Germany): Enterprises using AI by size class, 2025 destatis.de
- Destatis: Reasons for not using AI, 2025 destatis.de
- NIST AI Risk Management Framework nist.gov
- EU AI Act, Article 4: AI literacy artificialintelligenceact.eu
- EU AI Act, Article 113: Entry into force and application artificialintelligenceact.eu
- KI-Campus: AI Omnibus recalibrates Article 4 – AI literacy remains mandatory ki-campus.org