Contents
- The weak point first
- Core banking migration: every record is reconciled and signed off
- Onboarding legal entities: complete, or a targeted request
- Tax and compliance engagements: the question comes a year later
- COO of a consulting boutique: the method tried at home first
- Five principles that came out of it
- Education, briefly and on record
- What this means for your office
- Frequently asked questions
Key points
- Every station in banking and consulting ended with the same question: who approved this, on what basis, and can it be shown a year later.
- From the core banking migration: every migrated record is reconciled and signed off. An agent that carries data from a document into an entry works by the same rule.
- From onboarding legal entities: a file is either complete, or what is missing is requested in one targeted request.
- From tax and compliance engagements: the question often comes a year later. That is why a replay log records which document, which rule, and who approved it when.
- Five principles came out of this: one process first, measurement in your own hours, approval by a named person, a source reference for every suggestion, and nothing corrected silently.
The weak point first
I studied law and economics at the University of St. Gallen (HSG), with a focus on financial market regulation. AI came later, through further training and through my own bookkeeping. I do not train models. Grow Amplified's review lanes are built and run with a permanent development team, and I lead every engagement personally.
What I bring is the other side: 15 years in banking and consulting, in transformation and finance operations, in places where work only counts once someone has signed it off. Most guides to AI agents start with building. Get it running, then see whether anyone asks. I learned in the opposite order. My full background is on my profile.
Core banking migration: every record is reconciled and signed off
At a Swiss bank, I led the organisational transformation together with the migration to a new core banking platform, with 25 people on the project and a seven-figure budget. It also covered the change of back-office provider and the vendor selection, from the long list to the final choice.
In a migration like that, nobody says “the data is across”. Every migrated record is reconciled against the old system: accounts, balances, master data. Every difference is explained or cleared, and at the end someone signs the acceptance. Until then, nothing is switched over.
An agent does the same thing on a small scale every day. It carries data from a document into an entry, a list or a draft. So the same rule applies: reconcile against the source, make every difference visible, have a named person sign off. In a bank migration that is taken for granted. With AI tools in everyday office work it is often skipped, because each single transfer looks so small.
Onboarding legal entities: complete, or a targeted request
At a major bank, I simplified the onboarding of legal entities. Taking on a company as a client means checking documents against a checklist: commercial register extract, articles of association, signing authorities, details of the beneficial owners.
What I learned there is that a file has only two good states. Either it is complete, or a request is out that states exactly what is missing. The state in between is the expensive one. A request that names three items when five are missing costs a second round with the client and a second round inside the bank.
For agents, that means an agent checks documents against your checklist and delivers one of two results. Everything is there, or it drafts the request listing every missing item, and someone approves it. That applies to tax documents in a fiduciary office just as much as to client files in a financial institution.
Tax and compliance engagements: the question comes a year later
On consulting engagements for global Swiss banks, I worked on tax, compliance and US regulatory matters, including a US Tax Program. There I learned when the real review takes place. Rarely on the day the work is finished.
The question often comes a year later, from the auditors, from an authority or from the client: who decided this, on what basis, show me. If you can only answer from memory, you cannot answer. People have moved on, the emails are archived, the spreadsheet exists in four versions.
That is why every agent at Grow Amplified has a replay log: which document, which rule, who approved it and when. It is built for the day someone asks. Which retention rules apply is covered in Replay log and GeBüV.
COO of a consulting boutique: the method tried at home first
As COO of a consulting boutique, I applied the method in-house first: bookkeeping in Bexio, a fixed Monday check, reporting and forecast with AI and approval. That won back 9 to 13 hours a week.
The weak point of that figure: it comes from a single firm. It is not an average and not a promise.
I learned two things there. First, approval needs a fixed slot in the week. A check that happens whenever there is time stops happening at some point; the Monday check is that slot. Second, only what is measured in your own hours counts. The free Inbox Hour Counter shows how to do that in a single working day: you count the emails that come in and the replies that go out the same day.
Five principles that came out of it
- Start with one process
- A single process you know well can be measured, signed off and explained. Five half-built processes can only be described.
- Measure in your own hours
- Count first, then build, then count again. A saving only counts once it shows in your hours, at your internal rate.
- Approval by a named person
- A person with a name, not a team and not a shared inbox. They approve or reject, and their name is in the log. What the four-eyes principle looks like with agents is covered in Four-eyes principle with AI agents.
- A source reference for every suggestion
- Every suggestion shows where it comes from: document, page, line, rule. Without one, the person approving can only check whether the suggestion sounds plausible.
- Nothing corrected silently
- An agent flags; it does not fix things on its own. A person makes the correction, visibly and in the log. That way it stays clear what the source said and what became of it.
Education, briefly and on record
I studied at the University of St. Gallen (HSG): a Bachelor in Law and Economics, with a focus on financial market regulation. On AI, further training followed: “Artificial Intelligence: Implications for Business Strategy” at MIT Sloan Executive Education, the Singularity Executive Program and the Transformational Leader Program at IMD. Every credential is linked on my profile.
Courses do not replace practice. They help in asking a model and a vendor the right questions. The answers came from our own bookkeeping.
What this means for your office
None of these lessons needs a bank. A fiduciary firm, a property manager or an owner-led SME faces the same question, usually without a supervisor behind it: who approved what, on what basis. If you can answer that today without searching, you can use AI agents safely. If you cannot, start with the approval step before you start with the agent.
I started from the inspection and worked backwards. That is why the approval step and the replay log are in the build, not in the pitch.
Timon Wick
How AI works in a fiduciary office in principle is explained in the first article on this blog, AI in the fiduciary office, explained simply. How the steps follow one another at Grow Amplified is described under Approach.
Frequently asked questions
Do you need a banking background to use AI agents safely?
No. You need the habit of letting work count only once someone has signed it off. Every good fiduciary office already has that habit; it is called the four-eyes principle.
Do banking rules apply to an SME as well?
The supervision is different, the review question is the same. In its Guidance 08/2024, FINMA names documentation, explainability and independent review, among other points, for institutions that use AI. It does not apply to an SME, but it works as a yardstick.
Where do the 9 to 13 hours a week come from?
From my time as COO of a consulting boutique: bookkeeping in Bexio, a Monday check, reporting and forecast with AI and approval. It is one firm, not an average and not a promise. What is possible in your office only your own hours will show.
Do you lead the engagements yourself?
Yes. I lead every engagement personally, and your first call is with me. The review lanes are built and run with a permanent development team, and everyone involved is bound to confidentiality by contract before they see a document.
What does “nothing corrected silently” mean in practice?
When an agent finds an exception, such as an amount that does not match the receipt, it flags it with its source reference. A person makes the correction, and the log shows both: what the source said and what was changed.
15 years in banking and consulting, in transformation and finance operations. Today he builds review lanes with Grow Amplified: agents read and reconcile, a person approves, and a replay log records who decided what.