AI & Automation for the Mittelstand

AI is everywhere — on every conference stage, on every sales slide, in every other LinkedIn post. And yet inside your own company the picture looks different: apart from a handful of employees using ChatGPT for their emails, little has changed in daily work. We build AI and automation solutions for mid-market companies (the German Mittelstand) that genuinely take load off day-to-day operations — and we say openly where AI is not the right answer and a 50-line workflow does the job better.

Does this sound familiar?

Your management asks: “Why aren’t we using Copilot? Others already do.” — and nobody quite knows what an honest answer would look like.
A few months ago you bought Copilot licenses because they were offered with the Microsoft renewal. Three months later nobody uses it regularly, and the reporting shows the money just keeps flowing out.
The IT service desk gets 40 tickets a day, and it feels like 70 percent are the same frustrations: password resets, printer drivers, access requests. Your first line grinds through them and never gets to anything else.
Your employees ask subject-matter questions in WhatsApp groups, in emails to colleagues, or worse: in ChatGPT — because the company knowledge sitting in SharePoint, the DMS and network drives can’t be found.
Marketing is experimenting with ChatGPT, sales with Claude, HR with some CV-screening AI — and nobody in the company knows which company data ends up where.

Why this happens

In the last two years, AI has gone from a research question to a sales push. Every Microsoft partner brings up Copilot, every consultant has an AI roadmap in their portfolio, and at the same time the expectations of management, boards and family shareholders keep rising: “We have to do something here.” What rarely happens is the sober question of where AI actually takes work off people’s hands in your specific operation — and where it remains an expensive toy that ends up in a drawer after three months.

On top of that, most Mittelstand companies don’t have their data in the state modern AI tools need. Copilot can only be as good as the SharePoint permissions it sits on top of — and in many companies those permissions have grown organically over years and nobody wants to touch them. If Copilot then gets “successfully” rolled out, it suddenly makes personnel files, payroll data or board minutes findable — documents that far too many people have long had access to through overly broad permissions. That’s not an AI problem, it’s a data-foundations problem — but it only surfaces because of the AI.

And finally: AI is not the same as automation. A lot of what is sold today as an “AI project” is at its core a workflow problem — a recurring procedure that was never cleanly mapped, and for which a large language model is now supposed to be harnessed because that sounds more impressive than “Power Automate”. We deliberately keep the two worlds apart: AI where language and ambiguity are involved, classic automation where the process is actually clear and simply never got implemented consistently.

What this covers in practice

Microsoft 365 Copilot — when it genuinely makes sense

Before we book a single Copilot license, we check two things with you: what your SharePoint permissions look like, and which data in your tenant is classified at all. If permissions have grown wild, Copilot makes documents findable that were never meant for everyone — and the question “How big was management’s last bonus?” suddenly becomes answerable for far too many people in the company. First get the data foundation in order, then Copilot. And even then not for everyone, but for the roles where text work, research and summarization are part of the daily routine — sales, marketing, management, back office with heavy mail volume.

Internal knowledge search

Employees ask a subject-matter question in Teams, and the answer comes from the company’s own knowledge — from SharePoint, the DMS, the wiki, with cited sources. Technically this builds on Azure AI Search and a classic RAG pattern (retrieval-augmented generation): the AI doesn’t generate an answer from thin air, but first finds the matching documents in your repository and formulates the answer along those sources. How you can tell it’s needed: when new employees take three weeks to learn where documents live — and the old hands carry the knowledge in their heads.

Workflow automation (Power Automate / n8n)

The less glamorous but usually more rewarding part. Recurring procedures that today run on mail distribution lists, Excel sheets and word of mouth become defined workflows: quote dispatch with automatic filing in the DMS, order confirmation with feedback to sales, onboarding a new employee with license assignment, group memberships and device preparation. We use what fits the situation — Power Automate if you’re in the Microsoft world anyway, n8n for more open scenarios or when you want to stay independent.

Ticket triage & classification

Incoming service-desk tickets are pre-classified (category, urgency, likely resolution path), briefly summarized and routed to the right place. For recurring standard questions — password, VPN, printer — the system suggests a resolution path that the responsible person only needs to confirm. The human stays in the loop. How you can tell: when 70 percent of your first-line workload is the same five topics and nobody has time left for the genuinely interesting tickets.

Governance — what the AI gets to see, and what it doesn’t

The invisible but decisive part. Data classification (what is public, internal, confidential, strictly confidential), sensitivity labels in M365, prompt filters and an audit trail for AI usage. Plus clear ground rules for employees: what may go into ChatGPT, what may not, and which internal tools are available. This is the answer to the question your data protection officer is going to ask you in the coming months anyway.

Where AI doesn’t help today — the honest answer

AI is not a cure-all, and we’ll say it openly in the initial conversation: “This is better solved today with a 50-line Power Automate flow than with an LLM.” For example:

Structured data extraction from identical forms — if the form looks the same every day, a Power Automate flow with AI Builder (or Azure AI Document Intelligence) is more stable, cheaper and more predictable than any LLM.
Closing tickets without a human in the loop — risky. An AI that grants access or resets passwords on its own is a security hole you won’t be able to explain in an audit. The human should make the call.
Complex legal or regulatory assessments — AI as an assistant for research and structuring, yes; AI as the source of record on liability-relevant questions, no.

A sentence that comes up a lot here: if your sales team writes 50 quotes a week and 90 percent of them differ only in quantities and prices, you don’t need AI — you need a clean template and a Power Automate flow. AI makes the difference where language genuinely varies, where documents look different every time, where questions are ambiguous. That’s where it’s a lever. With fixed patterns, it’s the expensive detour.

What you should look out for — even if you don’t go with us

Ask for the concrete use case before anyone talks licenses. Anyone who opens with “Copilot costs … per month” instead of “Which role in your company would this relieve every day?” is selling a license, not a solution.
Ask to see the data classification before you roll out a knowledge AI. If nobody can say which documents are confidential and which aren’t, the AI ends up seeing everything. That’s not the AI’s fault, but it is your problem.
Ask about the rollback plan. If an automation flow sends order confirmations and hits an infinite loop at three in the morning, your sales team will want to know the next day how you stop it. No answer means no plan.
Be suspicious of ROI promises with concrete numbers. Anyone promising “300 % productivity gains from Copilot” has no figure they can substantiate — the market is simply too young for that, and the measurements that do exist are mostly vendor-funded. The serious version is: “We define up front how we’ll measure, and after three months we take an honest look.”
Clarify early where the data ends up. Microsoft Copilot stays in your tenant. ChatGPT in its consumer variants (Free and Plus alike) does not. Private Claude use with company documents is a data protection incident. Anyone who doesn’t make these distinctions is muddling terms.
Start small before rolling out broadly. One department, six to eight weeks, clear success criteria — then decide. Buying Copilot for the whole company before anyone has genuinely used it is burning money.

When it’s time to act

Your management is actively asking about Copilot or AI, and you want a well-founded answer rather than a gut reaction.
Growth is stalling at a plateau that looks like a headcount problem but is really a productivity problem — recurring tasks are eating the capacity your actual business needs.
Hiring for first line or back office has gone nowhere for months — the labour market simply isn’t producing candidates, and you need a different kind of leverage.
A concrete data protection concern has surfaced: employees are using ChatGPT with customer data, or generated text suddenly appears in marketing and nobody knows where it came from.
Your Microsoft license renewal is coming up, Copilot is on the table, and you want an honest basis for the decision instead of sales pressure.
NIS-2 preparation is under way, and the question “How do you handle AI tools?” will show up in the questionnaire.

How we work

Phase 1 — Initial conversation & use-case inventory

A 30-minute initial conversation, then a structured look at the recurring procedures in your company: what happens daily, what happens weekly, where frustration piles up, where time is lost. Deliverable: a use-case list sorted into “worth doing with AI”, “worth doing with classic automation” and “not worth doing at all, because the process needs sorting out first”.

Phase 2 — Trial in one department

Together we pick a use case that is manageable, measurable and visible if it succeeds. Six to eight weeks of trialling in one department, with clear success criteria defined up front. Deliverable: a running use case, an honest evaluation (“what worked, what didn’t”) and a solid basis for deciding whether to roll out further or take a different approach.

Phase 3 — Roll out, or back to square one

If the trial holds up, we roll out step by step — department by department, use case by use case, with user training and inclusion in the governance. If it doesn’t, we say so openly — and we either look for a better use case or tell you straight that AI isn’t the right lever for you right now.

Phase 4 — Operations & ongoing adaptation

AI models change, licenses change, workflows change. Optionally, we support ongoing operations in a quarterly rhythm: what’s new at Microsoft, which new use cases have emerged, what’s no longer running as planned. Deliverable: an AI and automation estate that grows with you instead of rusting away.

What you can expect from us — and what you can’t

What you get:

Direct access to the founder as your permanent point of contact — no ticket carousel, no rotating account managers.
An honest use-case assessment before anyone buys a license or builds a flow.
Success criteria defined for every trial, instead of “let’s see how it goes”.
Documentation a successor can understand — no spaghetti workflows that only we can maintain.
Recommendations that sometimes cut against our own revenue — if the right lever is a simple Power Automate flow, that’s exactly what we build.

What we deliberately don’t do:

ROI promises with concrete percentages. The market doesn’t support them, and we won’t commit to gut-feel numbers that come back to bite you in front of the supervisory board.
AI as an end in itself. If the use case runs cheaper and more stably on classic automation, that’s what we do.
Full automation of critical decisions without a human in the loop. Closing tickets, approving contracts, triggering payments — the human stays.

Where we’ll also say no:

If you want to introduce Copilot “because everyone has it” and the data foundation doesn’t support it — then we clean up SharePoint and permissions first, and talk again after that.
If the honest answer is: “This isn’t an AI use case, it’s a process that was never properly defined.” Then we talk about the process, not the model.
If what you actually need is training your employees in the safe use of the AI tools you already have, rather than a custom build. That, too, is a valid answer.

How to get started

A 30-minute initial conversation — free, no strings attached, by video or phone.
What we clarify: where noticeable, recurring effort arises in your company today, and which tools you already have in-house.
Useful to have beforehand, but not required: your current Microsoft license packages, the workflow/automation tools in use, and a rough idea of which department could best carry the trial.
Engagement models range from a one-off trial project to ongoing support in a quarterly rhythm, or a hybrid — we’ll work out what suits you in the conversation.

Book an initial conversation

Frequently asked questions

Do we really need Microsoft 365 Copilot? That depends on two things: your roles — who works with text, research and summaries every day — and your data classification. If SharePoint permissions are clean and there are roles with a lot of text work, Copilot can be a real lever. If the data is a mess, you’re buying yourself a security risk along with the licenses. We assess that beforehand.

What does an AI project cost? That depends on three drivers: how many use cases go into trial, how clean the data foundation already is (or whether it needs tidying up first), and how many employees need training at the end. We give you an honest range in the initial conversation — a flat figure without a look into your tenant wouldn’t be credible.

How do we prevent the AI from sending company data to OpenAI? Through the choice of tool and through clear ground rules. Microsoft Copilot stays in your own tenant; so does Azure OpenAI Service. ChatGPT in its consumer variants does not — there, inputs can be used for training by default; Team and Enterprise agreements handle this differently. We set up the tools so that company data stays where it belongs, and we define with you what may go into which tool.

Can we use AI without Microsoft? Yes. Azure OpenAI is one option, Anthropic Claude via Amazon Bedrock another, local models (Llama, Mistral) on your own hardware a third. We’re not ideologically wedded to Microsoft — whatever fits the situation is what we use. For many mid-market companies Microsoft is the pragmatic route because M365 is already in the house; but it’s not mandatory.

Who is liable if the AI says something wrong? When in doubt, the company that acts on the answer. That’s why our architectures keep the human in the loop — the AI proposes, the human decides. That’s not a brake, it’s risk management. In knowledge searches with source citations, the AI is a research tool, not the final authority.

What do we do if our employees use ChatGPT privately for company work? First, don’t moralize — they do it because the internal tool is missing or too cumbersome. Second, set clear ground rules and provide a sanctioned internal tool, so the reflex no longer points to ChatGPT. Third, train people on what may go into which tool. Bans without an alternative don’t work.

Need a clean M365 tenant as the foundation first? Services overview