AI honestly placed — when we recommend it and when we do not.

There is a reflex in the consulting trade that annoys me: the moment a client asks about “AI”, a four-phase roadmap lands on the table. Phase one: data strategy. Phase two: use-case analysis. Phase three: pilot. Phase four: scaling. All of it over twelve months. The client nods, because it feels as though this is how it has to sound when people are being professional. The consultancy is pleased, because twelve months means a healthy fee. And at the end someone works out that a simple webhook and an if-condition would have solved eighty per cent of the problem.

That is my most honest observation from two years of AI consulting in mid-sized business. I am writing this article because I think that, as someone running a company, you have a right to a better answer than the usual roadmap. There are situations where AI makes complete sense. There are more situations where it does not. And there is a growing grey area where the right answer depends on details that a thirty-minute first conversation cannot supply.

Three layers that get confused constantly

Before I explain when we recommend AI, I need to clear something up, because three terms get mixed up in practice all the time.

Classic process automation is everything that runs to fixed rules. When A happens, do B. An order comes in, the confirmation is produced. A lead matches a category and is routed to the right person. A document lands in the mailbox and is moved to the right folder. This kind of automation is old, transparent, low-maintenance, and the right answer in ninety per cent of use cases at smaller companies. Tools: n8n, Make, Power Automate, conventional workflow systems.

AI-assisted automation is process automation plus an intelligent component that can cope with ambiguity. An email comes in, a model decides whether it is a complaint, a new order or a general enquiry, and routes it accordingly. A PDF invoice is scanned and a model extracts the relevant fields, even when the layout is new. This is where pre-trained models come in, often as API calls to OpenAI, Anthropic, Mistral or locally hosted open-source alternatives. Tools: everything in the first category plus a connected language or image processing API.

AI agents are the newest and most hyped layer. They combine several capabilities into a system that plans and carries out tasks on its own. An agent is given a goal — “answer all customer enquiries this morning” — and decides for itself which data sources to query, how to phrase the answer, whether it needs to ask something back. Frameworks: LangChain, AutoGPT, Claude Code for our own delivery chain. This layer is powerful, but it is also considerably more expensive, harder to control, and in many mid-sized contexts still too early.

Anyone saying “AI” usually means one of these three layers — but it is rarely clear which. The right recommendation depends on answering that question first.

When we do not recommend AI

Let me start with the uncomfortable part. There are three situations in which we deliberately do not recommend AI — even when the client originally asked for it.

When the problem can be solved with rules. A haulage firm wanted an AI a few months ago that would automatically classify incoming emails from drivers and match them to jobs. Sounds like a classic AI use case. During the assessment it turned out that the drivers send their emails in a structured way — job number in the subject line, fixed format. A regular expression in n8n solves that with five minutes of configuration and has run for months without maintenance. An AI model would have worked, but it would have cost two hundred euros a month, would have hallucinated occasionally, and would have tripped over a change to the letterhead format. The rule does none of those things.

When the data quality will not carry it. AI models are only as good as the data they work with. If your master data is faulty, if fields are maintained inconsistently, if spreadsheets are the real source of truth — then an AI solution produces systematically wrong recommendations. We had a case where a client wanted a recommendation engine to suggest similar products to customers. Analysing the data, we found that product categories were classified differently in two separate systems — and nobody knew which classification was more current. The right recommendation was not “let us train a model” but “let us get your master data in order first”. That took a year and nobody enjoyed it. But the recommendation engine works now, and it really only needs conventional statistics, no AI at all.

When the question of responsibility is unsettled. An AI that prepares decisions with legal consequences needs clearly defined responsibilities. Who is liable if the model decides wrongly? How is the decision made traceable? What happens when an audit asks? In many smaller companies these questions are unanswered, because the organisational structure was never built for them. In such cases we recommend using AI only in a suggesting role — the model proposes, a person decides. That halves the efficiency gain, yes, but it protects against risks a smaller company cannot carry.

When we do recommend AI

There are four situations in which we recommend AI without hesitation.

Classifying large volumes of incoming material. When a company receives hundreds of emails, documents, enquiries or calls a day and these have to be sorted into categories — and when the incoming data is not structured enough for a regular expression — then AI-assisted classification is a very good answer. We use it regularly for incoming mailboxes (complaint versus order versus general enquiry), for documents (invoice versus delivery note versus contract), for call transcripts (escalation versus routine).

Extracting structured data from unstructured sources. Incoming PDFs, scanned documents, free-text emails with order data buried in them — anywhere the information is present but not in the expected format. Modern language models are considerably better here than conventional OCR plus a rule set. A caveat: accuracy is never a hundred per cent. We always build such solutions with a validation step — the model extracts, a person or a second rule checks. With documents that feed into DATEV, that is mandatory.

Personalisation in marketing and sales. Copy for different audiences, briefings for social media campaigns, translations, summaries of longer documents. AI gains dramatically in value here, because the alternative — someone in marketing writing it by hand — is considerably more expensive and does not scale. Our own tool trendscout came out of exactly this need: website analysis plus trend data plus language-model-based briefing generation, tuned to the specific company.

Software development itself. This is where we use AI most deeply. Claude Code has been our central tool for months on every client implementation. That is not “let the AI write whatever it likes” but a very disciplined agentic way of working, with clear contracts, architecture sketches and version control. The result is that a small team reaches a delivery depth that would previously have required three or four times the staff. We pass that efficiency on to clients — through lower project prices, or through deeper involvement for the same budget.

The grey area: when the answer depends on details

Between “clearly recommend” and “clearly do not” lies a wide area where the right decision depends on factors a thirty-minute conversation does not settle. Three examples.

Chatbots in customer support. If you have many recurring questions that are clearly answered somewhere in your knowledge base, and if you manage to build a clean escalation path to a human — then an AI chatbot can work. If not, it will frustrate your customers and cost you time. Most implementations at smaller companies that I have seen over the past two years fall into the second category. But there are exceptions.

AI-assisted data analysis for management decisions. The problem here is not the AI but whether management is prepared to accept data-driven recommendations. If the answer is “when in doubt we go with our gut”, the investment is wasted. If the answer is “we want to lead on evidence but lack the time for the analysis”, AI can help enormously. That question is cultural, not technical.

Generated content for the website or the CRM. It depends on how much your brand voice matters to you. If your copy is interchangeable and you simply need to produce it, AI is an accelerator. If your brand voice is specific and does not survive without care — then AI produces well-meant text that all sounds the same after three weeks, because it comes from the same model. We use AI intensively when writing, but always as a first draft, never as the finished product.

What we actually do on client projects

So that this does not stay abstract, three concrete examples from recent months — all anonymised, because we take discretion seriously.

Example 1: an automated incoming mailbox. A client received around two hundred emails a day that had to be sorted into different categories — orders, complaints, general enquiries, supplier correspondence. Before: two employees each spent two hours a day sorting. We built an n8n workflow that sends the emails to a language model, receives a category suggestion and moves them into the right area of Microsoft 365. Correction rate: three per cent. The employees correct what is misfiled and now spend ten minutes a day on the mailbox. They deal with the actual cases instead.

Example 2: pre-checking documents in accounting. A client received hundreds of incoming invoices a month, which had to be checked manually for completeness, matched against the purchase order and passed to DATEV. We built a workflow that scans the invoices, extracts the fields with a language model, checks them against an SQL database of open orders, and puts only the unclear cases in front of accounting. Accounting now sees fifteen per cent of the invoices — the other eighty-five per cent are posted automatically. The validation step: every automatically posted invoice must have an open order at the correct amount. Otherwise it is diverted into the manual queue.

Example 3: a marketing briefing generator. Our own tool trendscout came out of this need. A client needed social media content regularly but had no marketing employee and no time for an agency. trendscout analyses the client’s website, compares it against current trend data and social signals, and suggests concrete posts — including image suggestions. The client checks what fits and posts it. That is AI as an accelerator, not as a substitute for brand work.

What we deliberately do not do

We do not sell AI strategies. We do not run multi-workshop programmes on “AI in mid-sized business”. We do not deliver twelve-month roadmaps. Not because we understand none of it — but because the format rarely suits the task. When a smaller company asks today what it should do with AI, the right answer is almost always: identify concrete problems, build solutions pragmatically, start small, deliver early. A three-day workshop on “AI strategy for your company” usually produces a document nobody can find six months later.

What we do: we take in your concrete pain points, look at the data and the processes, identify where AI genuinely brings efficiency, propose the simplest implementation that solves the problem, and build it. If that is a language model integration, it is a language model integration. If it is a regular expression, it is a regular expression. Both are legitimate, both are professional. What is not legitimate is selling AI because the word happens to sell well.

An honest closing note

I believe AI will work its way more deeply into every smaller company over the next five years. Not as a buzzword, but as a matter-of-course component — embedded in standard software, in office applications, in CRM and ERP. For most people running such companies it will not be a conscious decision at all. It will simply be there.

Until then there is a transitional phase in which AI is introduced as an additional layer, with all the questions that brings. During that phase you do not need an AI strategy. You need a sparring partner who answers honestly, for each concrete question, whether AI helps or not — and if so, how.

That is the role we see for ourselves. Not as AI consultants, but as consultants who have understood AI as one tool among many and use it where it holds up. That sounds less spectacular than “we will make you AI-ready”. But it is the more honest answer. And it is the one that will still hold in five years.

This is part two of a three-part series on digiFORMER’s position. Part three: digital sovereignty for smaller companies — more important than you think today.