The call it usually starts with
A managing director saw a demo at a trade fair. An assistant reads emails, answers them, logs everything in the CRM. Impressive. Back at the office the question comes up: what are we actually doing with AI? Surely something would be possible in sales. And in support anyway.
That is the point where most people call. The question is honest, but it is too big to answer. “What do we do with AI” leads to a workshop, a slide deck and then to nothing. The question that gets you somewhere sounds different: which task eats hours every week, follows recognisable rules and annoys everyone who has to do it?
That rephrasing is the actual core of AI consulting. Everything after it, the choice of model, the interfaces, the operation, is craft. The answer to the first question decides whether something runs in the end or whether a budget drains into a pilot nobody uses.
Four phases until the first agent works
The first two phases are consulting in the narrow sense. They cost little and decide a lot. The last two are implementation.
1. Pick the use case
We go through the processes where someone currently reads, sorts and triggers a next step. Pulling quotes together, sorting enquiries, maintaining contacts, typing up receipts. From that list the first candidate almost falls out by itself: the one that comes up often, can be described clearly and does no damage if it goes wrong. Which tasks those are in practice, I walked through in six cases under AI agent examples.
What helps most in this phase is listening to the people who do the work. Not to the management, who know the process from above, but to the person who answers the same email three times a day.
2. Check feasibility
This is where it shows whether the candidate turns into a project. I look at three things: whether the data is good enough, whether the systems hang on interfaces you can address, and whether it can be pinned down cleanly what the agent is allowed to decide by itself. The last question is the underrated one. Where the line between automation and human approval runs is a decision for the company, not a technical one.
This phase also decides whether it has to be an agent at all. If the process runs the same way every time, a fixed workflow automation is cheaper and more reliable. The difference between the two is covered in detail under what AI agents are.
3. Build the pilot
One process, properly connected, running dry. The agent reads along and suggests, but does not write yet. That way you see on real cases where it gets things wrong before anyone notices it in the system. This phase is unspectacular and the best protection against an expensive disappointment.
4. Rollout and expansion
If the pilot holds up, the agent gets write access for the uncritical parts, and the rest keeps running through approvals. After that the next use case follows. The first agent costs the most time because the connection and the rules are built once. Every further one feeds off that groundwork.
How you do the maths in phase 2
Whether a use case pays off is not decided by a gut feeling but by simple arithmetic: what does the process cost in time today? Everything the build costs has to stay well below that.
The numbers above are assumed, the logic is not. A process with 30 hours a month carries a pilot. A process with two hours a month does not, no matter how good the demo looked. The items on the other side, model costs, hosting and maintenance, I broke down separately under what an AI agent costs.
When I advise against AI
A good share of my consulting calls end with me advising against an AI project. That sounds like bad business, but it saves both sides money. Four cases come up again and again.
The process comes up too rarely. A task that happens three times a month does not justify a build, however annoying it is. The calculation above then comes out clearly.
The process only exists in one person’s head. If nobody can explain the rules by which decisions are made, there is nothing to automate. Then the description comes first, and that is organisational work, not AI.
The data is messy. Duplicates, outdated contacts, half-filled fields. An agent on top of that only writes errors faster. In that case a clean-up and enrichment of the CRM data comes first, and we talk about agents afterwards.
No AI is needed at all. The most common case. The process is rigid, the same every time, and an automation in n8n handles it more reliably and more cheaply than any language model. AI is then the more expensive answer to a solved question.
Starting with five use cases at once because every department wants to contribute something. In the end none of them is finished, nobody feels responsible and the technology gets the blame. One process, seen through properly, beats five half-finished ones.
You have a process in mind but are not sure whether AI is the right tool for it? Describe it to me in two sentences. You get an honest assessment, even if it turns out a simple automation is enough.
How you know you are ready
The more of this is in place, the shorter the consulting and the faster the pilot. The list does not have to be complete.
Before the intro call
- A concrete process is named, not just a department
- Someone in-house can explain the rules decisions are made by
- You roughly know how often the case comes up per month
- The systems involved (CRM, inbox, ERP) are known
- There is a person who grants approvals and decides exceptions
- It is clear how you would measure success: time, errors or response time
Frequently asked questions about AI consulting
I taught myself to code at 18 and have been building AI agents and automations with n8n and HubSpot ever since. Consulting for me means looking at a real process and saying whether an agent, a plain automation or nothing at all is worth it. Building comes after that, not before.
Auf LinkedIn vernetzen →Which process at your company carries an AI agent?
30 minutes, no strings attached. We go through one concrete process and afterwards you know whether it suits AI, what the build roughly means and where to start.
Pick a slot →