AI Strategy for Mid-Sized Companies: From Pilots to Execution
In conversations with managing directors of mid-sized companies, we currently hear two sentences strikingly often. First: "We need to do something here." Second: "We have tried quite a few things, but little of it stuck." Both sentences describe the same problem from two sides. What is missing is not motivation or tools. What is missing is a path to execution that fits the size and risk capacity of a mid-sized business.
Three dead ends that bury pilots
Before we get to the path, it is worth looking at the dead ends. We find three patterns in initial conversations again and again.
Tool procurement without a task. Licences have been bought, a task was never defined. Six months later, a fraction of the staff uses the tool occasionally, and nobody can say what it achieves. The budget is gone anyway, and with it, often, the willingness for a second attempt.
The pilot without a metric. "The team quite likes it" is not a basis for a decision. Without a baseline measurement, every pilot ends in matters of taste, and matters of taste lose the next budget round against anything that has numbers.
The permanent experiment. The pilot simply keeps running past its planned end date, half supported, half forgotten. It ties up attention, produces no decision and serves internally as proof that "this AI thing doesn't really catch on here".
All three dead ends share the same root: the frame of task, owner and measurement is missing. That is exactly what the following four steps provide.
Step 1: Analyse tasks, not technologies
The most common mistake we see in projects is the order of events: first a tool is procured ("We have Copilot licences now"), then the search for use cases begins. Turn it around. List the ten most time-consuming recurring tasks in sales, administration and customer service. Rate each one against two criteria: How structured is the task? and How expensive is a mistake?
Structured and error-tolerant (proposal drafts, meeting minutes, first replies in support) → pilot immediately. Structured but error-critical (invoices, contracts) → AI as an assistant with a four-eyes principle. Unstructured and error-critical → a human matter for now.
This is what the matrix looks like filled in, with tasks as they typically come up in our projects with mid-sized companies:
| Mistakes cost little | Mistakes cost a lot | |
|---|---|---|
| Structured task | Proposal drafts, meeting minutes, first replies in support → pilot immediately | Invoice checking, contract drafts → AI as an assistant, approval under a four-eyes principle |
| Unstructured task | Research summaries, first ideas for marketing copy → pilot, keep expectations low | Personnel decisions, price negotiations, crisis communication → a human matter for now |
The written form matters. As long as the list exists only in people's heads, the loudest tool wins, not the most rewarding task. One hour with the department leads is enough for a first version; the matrix does not have to be perfect, only honest. It also works as a corrective against wishful thinking: whoever sorts their pricing calculations into the top right cell has thereby also decided that no full automation belongs there.
Step 2: Two processes, one owner, 90 days
Do not pilot ten ideas; pilot two, and do those properly: with a named owner who knows the process (not the head of IT by virtue of office), with a baseline measurement (hours per week, turnaround time, error rate) and with a fixed end date. After 90 days there are exactly three permitted outcomes: scale, adjust and extend, or stop. The fourth, usual outcome is "it just keeps running on the side". It is forbidden, because it ties up attention and delivers nothing.
Three examples show what a baseline measurement means in practice. The numbers are typical observations from our projects and should be read as orders of magnitude: a proposal draft in technical sales often costs two to four hours, measured from the incoming enquiry to the document ready to send. A qualified first reply in support frequently takes 15 to 30 minutes per case. Meeting minutes tie up 30 to 60 minutes of follow-up work per meeting. Only with starting values like these can you decide after 90 days instead of going by feel.
The owner has three mandatory appointments in those 90 days. In week 1, they set the metrics and starting values and define what the tool is explicitly not allowed to do. In week 4, they check the first results with the team and adjust prompts, templates and responsibilities. In week 12, they present the numbers to management and recommend one of the three decisions. If nobody can be found for this role, that is itself a finding: then what is missing is not staff, but priority.
Step 3: Make your data foundation and website AI-ready
Two infrastructure topics decide whether pilots become systems:
Your company knowledge must be usable for AI: product data, price lists, documentation in searchable, structured form. That is usually the real project behind the AI project (the keyword is RAG).
Whether your data foundation is ready can be settled in a short meeting with IT. Three questions you can ask tomorrow:
1. Is our product and process knowledge available in searchable form? That means maintained databases, wikis or structured repositories. A drive full of scanned PDFs is not a data foundation but an upstream digitisation project. 2. Which source leads? If prices live in the ERP, in the web shop and in an Excel list, an AI will answer with three different truths. Before the AI project comes the decision which system counts. 3. Can we reach the data via an interface? Whatever is only reachable by manual export cannot be built into running workflows. An honest inventory of your interfaces takes a day and saves months later.
Your website is becoming the interface between your company and your customers' AI systems. ChatGPT, Perplexity and Google AI Overviews must be able to represent your offering correctly: structured data, clear statements, answered questions. A quick self-test: ask ChatGPT or Perplexity about your company and your core services. Whatever appears wrong or not at all there is already costing you enquiries today, because the answer then goes to a competitor whose pages are AI-readable. More on this in GEO: staying visible when AI gives the answers.
Step 4: Measure, decide, repeat
AI strategy is not a project with an end date but a rhythm: quarter by quarter, review two new tasks, re-measure what is running, document what was learned. A company that has established this rhythm no longer needs an "AI transformation". It has an organisation that routinely evaluates and adopts new tools.
In practice, the rhythm is a fixed quarterly meeting of about 90 minutes with four points on the agenda:
- Re-measure: Check running AI workflows against their starting values from step 2. What delivers, what has gone dormant?
- Decide: For each pilot, choose one of the three options: scale, adjust and extend, or stop. Postponing is not an option.
- Resupply: Pick two new tasks from the matrix in step 1 and name an owner for each.
- Record: Document what worked and what did not, including the training needs in the team.
A realistic order of magnitude for getting started: a two-digit number of person-days for the analysis and the first pilot, not a seven-figure programme. Mid-sized companies have a structural advantage over large corporations here: short paths, fast decisions, processes that one person can still oversee completely. Use it, and start with the task list from step 1.
Do we need an AI officer?
In most mid-sized companies, not as a new position. What matters more is that every single AI process has an owner from the business side who knows it and re-measures it. A lean coordination role, often held by the person responsible for digitalisation or processes, is enough to avoid duplicated work and tool sprawl. An AI officer without closeness to the processes, by contrast, quickly becomes a bottleneck that every idea has to pass through and little comes out of.
Our own model or an API?
For almost all mid-sized companies the answer is: rent models via API and keep them interchangeable. Training or operating your own model only pays off with very specific data and very high request volumes, and both are rarer than vendor brochures suggest. The economically more important question is who owns the integration layer between the model and your systems. How to make that decision cleanly is the subject of our article Make or buy for AI solutions.
What about data protection?
Data protection is solvable, but not retroactively. Clarify three points before the pilot: which categories of data the tool processes, where the provider processes them (EU region or data processing agreements with clear contracts) and whether personal data is needed for the task at all. Many pilot tasks work with anonymised or purely internal data. Involve your data protection officer in week 1, not after the first complaint.
How do we bring the workforce along?
With honesty and participation. Say clearly that the goal is relief from routine work, and prove it by letting the first pilots take over tedious tasks rather than prestigious ones. Let the people who do the process today help select, test and evaluate the tool; they spot weaknesses faster than any project group. In our observation, distrust rarely comes from the AI itself, but usually from a rollout over people's heads.
Go deeper in our knowledge base
Want to know what these topics mean for your company? The Future Check shows you the biggest levers within 2–4 weeks.
