Contents
- AI in one sentence
- Five terms are enough
- What AI does well, and what it cannot do
- Five concerns, and what addresses them
- How an agent works, in four steps
- Six tasks where AI takes manual work off your hands
- What an agent does not do
- An example from practice, anonymised
- The first step: one process, measured in your hours
- Frequently asked questions
Key points
- AI reads, compares and drafts. Deciding and signing stay with you, and so does the responsibility.
- An agent is AI with one fixed task that follows your rules, for example checking incoming tax documents against your checklist.
- Every proposal goes through an approval, and every step is logged. Your four-eyes principle stays; it only gets faster.
- Most manual work sits in reconciling: receipts, checklists, tax assessments, VAT and the inbox.
- The first step is one process, measured in your own hours.
AI in one sentence
Think of AI as a very fast trainee. It has read a great deal, it does not yet know your office or your clients, and it works in seconds rather than hours. As with any trainee, its work is checked before it leaves the office.
That picture explains almost everything that matters in practice: what you can entrust to it, what you need to explain to it, and why someone on your team always looks at the result. It also explains why the fear of AI often sits in the wrong place.
Five terms are enough
- AI
- A program that understands ordinary language, reads texts and writes drafts. You write to it the way you would write to a colleague. Nobody has to program.
- Agent
- AI with one fixed task that it carries out the same way every time, following your rules. For example: checking incoming tax documents against your checklist.
- Intake
- The agent takes in receipts, emails and PDFs the way they reach you today.
- Approval
- A person reviews the proposal and decides: release it as it is or correct it. That is your four-eyes principle.
- Log
- Every step is recorded: who, when and on what basis. Traceable for you, your team and the auditors. At Grow Amplified this is called the replay log.
You do not need more vocabulary to get started. With these five words you can place most AI offers, and you know which questions to ask a provider.
What AI does well, and what it cannot do
What it does well
- Read receipts, emails and PDFs and pull out what matters
- Compare two documents and flag every difference
- Draft letters and emails in your office's tone
- Explain long texts, such as new rules, in three sentences
What it cannot do
- Know what you agreed with a client, as long as it is not written down anywhere
- Take responsibility: that stays with you
- Calculate without errors unchecked: a person always checks the figures
- Send anything that nobody has approved
The right-hand column is not a flaw that disappears with the next version. It describes the division of labour. AI takes over the reading, sorting and reconciling; the judgement stays with the person who signs for it. Every agent is built around that division, which is why each one has an approval step.
Five concerns, and what addresses them
“AI makes mistakes.”
True, as with any new employee. That is why a well-built agent does not calculate freely from memory. It reconciles: receipt against booking, return against assessment, incoming documents against the checklist. Anything that does not match is flagged and never silently corrected. An error caught before approval is the reason approval exists.
“In the end, I am liable.”
Correct, and an agent does not change that. It signs nothing and files nothing. It prepares, a named person in your office decides, and the log shows even a year later which rule was applied to which document. Responsibility stays where it is today. What is new is that it can be proven.
“Client data does not belong on the internet.”
Also correct. Before an agent opens the first file, it must be agreed in writing where the data is processed, who has access and for what purpose. At Grow Amplified that means processing in Switzerland, a data processing agreement before the first access, only the data the engagement needs, and every access logged. Confidentiality obligations are passed on contractually to everyone involved. Ask every provider for these points, in writing.
“Then they will not need us for long.”
An agent takes over manual work: typing, searching, reconciling. It does not take over advice, the conversation with the client or the judgement on a borderline case. In a fiduciary office, expertise is rarely what is scarce. Time to use it is.
“This is too technical for us.”
You work with AI in ordinary language. The rules an agent checks against are written in plain words, the way you would explain them to a new colleague. New software is usually not needed: the agent works with the exports and documents you already have.
How an agent works, in four steps
- 01IntakeDocuments arrive as they do today: receipts, emails, PDFs and exports.
- 02ProposalThe agent prepares the work by your rules and flags whatever it cannot assign clearly.
- 03ApprovalA named person decides: approve or correct. Without approval, nothing leaves the office.
- 04LogEvery step stays traceable: which document, which rule, who approved it and when.
From office to office, the documents and the rules change. The approval and the log never do.
Six tasks where AI takes manual work off your hands
The examples below come from the typical day of a fiduciary office. They are starting points, not promises. Which of them costs your office the most time only becomes clear from your own hours.
| Task | Today, by hand | With an agent, you approve |
|---|---|---|
| Checking tax documents | Matching what came in against the checklist, requesting missing items by hand | Matches against your checklist and drafts the request for what is missing |
| Recording receipts | Typing up and coding receipts from paper, email and photos | Reads receipts and proposes the booking |
| Checking tax assessments | Reading the assessment and the tax bill line by line against the return | Compares both and flags every difference |
| Payroll administration | Registrations, deregistrations, accident and sickness reports by hand | Pre-fills the forms from the payroll data |
| VAT return | Reconciling revenue by hand against the books and the return | Reconciles and reports differences before filing |
| Inbox requests | Every email and every call lands with the owner | Sorts each request and drafts the reply |
All six tasks have one thing in common: reading, comparing and transferring. That is where AI is strong, and that is where the time goes that is missing for clients. How much recording receipts costs your office, you can work out with the free Shoebox Calculator on one month of your own receipts.
What an agent does not do
- It signs nothing and files nothing. That stays with you.
- It sends nothing to clients or authorities that nobody has approved.
- It does not invent figures. A saving only counts once it has been measured.
- It does not decide borderline cases. Whatever the rules cannot assign clearly comes to you as an open question.
An example from practice, anonymised
A fiduciary and property management firm bills much of its work as flat fees. The open question was how much recorded work disappears into those flat fees, and how much of it could be billed. Two mandate contracts against six time-tracking exports: estimated at four weeks by hand, done in one day with agents and approval.
The example shows where AI finds money first in fiduciary work: not by replacing professionals, but by reading documents that nobody would otherwise read again. With the free Flat-Fee Leak Check you can check the same for one mandate in your own office.
The first step: one process, measured in your hours
Nobody has to rebuild their office in one go. The safest start is a single process that costs a lot of manual work today and that you know well, for example checking incoming tax documents. First, count how many hours it takes today. Then an agent checks a limited number of files by your rules, with approval and a log. Then count again.
Whether it pays off is then decided by your own figures, not by a presentation. Ideally, start before tax season, while the office still has some room. How the steps follow one another at Grow Amplified is described under Approach.
Frequently asked questions
Will AI replace my staff?
No. An agent takes over reading, typing and reconciling. Advice, borderline cases and approval stay with your people. What you gain is time for the work clients pay you for.
What happens to my clients' data?
Before the first access, it is agreed in writing which data the engagement needs, where it is processed and who has access. At Grow Amplified, client data is processed in Switzerland, a data processing agreement is signed before the first file is opened, and every access is logged.
Who is responsible if the agent gets something wrong?
The same person as today. That is why approval exists: an error caught before approval is rejected, and the log shows which rule was applied to which document.
Do I need to program or buy new software?
No. The rules are written in plain language, and the agent usually works with the exports and documents you already have.
Where is the best place to start?
With a process that costs a lot of manual work and has clear rules, for example checking incoming tax documents against your checklist. First measure, then build, then measure again.
What does a first step cost?
It depends on the process. That is why it is first calculated in your own hours. You then get a fixed price with a clear start and end before you commit to anything.
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.