AI Consulting: a close look at how your company works, then automation built to fit it
The six agents cover the processes I meet most often. If yours is different or you are not sure which one applies, the sensible start is a short piece of consulting: I go through your processes with you, find the steps where automation pays off and design a solution that fits them exactly — usually a custom agent or an integration rather than anything off the shelf.
When consulting is the right first step
The six agents cover HR questions, outreach, contracts, documents and CVs. If what eats your team’s time is quoting, scheduling, supplier paperwork or a support inbox, it needs its own design before anything gets built.
Four or five things look automatable and you can fund one this year. Consulting puts a number on each of them, so the choice rests on value rather than on whoever complains loudest.
A tool was bought, a pilot ran, everyone lost interest. Almost always one of two things went wrong: the data was never prepared or the wrong process was picked. Both are easy to find out before you spend again.
Which processes, in what order, what each costs to build and what it returns. Some owners want that map on the table before committing to anything, which is exactly what consulting produces.
What comes out of it
Which specific steps a machine can take over now, which it would get wrong and which are cheaper left with a person. Established by talking to the people who do the work, because the org chart never shows where the time actually goes.
Cost per unit of work today against cost once it runs automatically, then the same comparison at double and triple your volume. A person hits a daily ceiling and you hire another one; the automation keeps charging per unit. This is where the case is usually decided.
For each step: an agent, an automation or an integration, which of your systems it plugs into and what data has to be prepared first. Designed around your process, so you know exactly what gets built before anything is built.
What each choice means for personal data and GDPR, what a local model costs you in hardware and where the quality gap actually shows. Benchmarked on your own cases when the call is close.
The order matters more than the list. Which process to start with, what has to be ready before it starts and where to put the pilot so a failure costs two weeks rather than a year.
This list comes out every time. Steps where the arithmetic does not work and steps where the real problem is the process itself, so automating it would only make a mess run faster.
Three ways to run it
Conversations with the people doing the work, then the candidate steps, the numbers behind each one and the shape of a solution for the top ones. Ends with a written plan and a call to walk through it.
The audit plus the version you can take to a board or a bank: what it costs to build, what it returns per unit of work, what changes as volume grows and in what order to do it.
For companies with their own developers or an automation already running. Architecture reviews, model choices, reading what gets shipped and saying where it will break before your users find out.
What we'll need from you
A short list. Consulting runs on how things actually work, so the conversations matter more than the documents.
- a. Two or three conversations of about an hour with the people who run the process, not only with management.
- b. Whatever numbers you already have: volumes, how many people touch the process, how long one unit of work takes.
- c. A look at the systems involved. Read-only access is enough at this stage.
- d. One person who can say what the company will and will not change.
How an agent gets built (the HR Assistant as the example)
Six stages. The software is the easy part — the sequence below is what makes an agent answer correctly. The specifics here are from the HR Assistant; every other agent goes through the same six stages, only the data and the metric change.
A call, a walk through your process, an audit of the data you have. You get a written verdict: what to automate and what it will yield.
If the agent won't pay off — I say so and we stop here. Free of charge.
Your policies, records and templates get systematised, cleaned and structured — together with the person on your side who knows the actual rules.
This is 80% of the result and exactly what the €100 solutions skip.
A working agent on your prepared data. You and your team get access and try to break it with real questions.
A fixed set of test cases runs against local and cloud models. You get a table — quality, cost per dialog, privacy — and make an informed managed-vs-self-hosted call. For the HR assistant that took 16 rounds across Gemma, Qwen and Claude models before the choice was obvious.
A benchmark instead of an opinion.
A limited group uses the agent for real. We measure the share of conversations closed without a human and the response quality, tune the data and prompts and harden against prompt injection.
Deployment on my infrastructure or yours, team training, SLA support. Every month: a metrics report. Keeping the knowledge base current is part of support, not an extra.